Last week, Anthropic announced that it had struck a deal with SpaceX to lease all of the latter's Colossus 1 data center, with over 220,000 GPUs and 300 megawatts of compute capacity. The deal immediately raises questions, foremost among them: why would Musk lease one of xAI’s most aggressively hyped AI assets to a direct rival? With SpaceX's IPO just around the corner, a related strategy appears to be at play, but it also turns out that the system's mixed architecture with different types of GPUs may be a key reason Musk has decided to lease the system.

Anthropic says the newly acquired capacity will primarily be used to ease long-standing usage bottlenecks across Claude’s paid ecosystem. According to the company, the additional compute will enable significantly higher Claude Code limits, the removal of peak-hour throttling for Pro and Max subscribers, and substantially increased API request limits for Claude Opus models used by developers and enterprise customers.

The seemingly unlikely partnership — a complete turnaround of Musk's earlier stance on Anthropic — also reveals Anthropic is straining under the Claude ecosystem’s compute demands. The company says it needs the entire 300 MW AI supercluster just to improve the experience of using Claude.

Anthropic appears to have hit the compute wall

The earliest signs that Anthropic was struggling to keep up with the computing demands of its growing user base were the increasingly aggressive usage limits placed across Claude’s services. Free users frequently complained about rapidly exhausting tokens — the units Claude assigns for processing tasks. However, the restrictions extended beyond the free tier. Paid Pro, Max, Team, and Enterprise users also regularly encountered message caps, peak-hour throttling, API rate limits, and strict time-based usage ceilings on Claude Code sessions, particularly during periods of heavy demand.

It was clear that Anthropic was running out of inference capacity. While training an AI model is an expensive, one-time computational undertaking, serving that model to millions of users simultaneously creates a continuous, round-the-clock demand for compute that scales directly with every new user and every new query. The apparent solution is to build more data centers, which Anthropic is apparently pursuing via massive gigawatt deals with Amazon, Google, Microsoft, and Nvidia.

However, modern hyperscale AI data centers can cost tens of billions of dollars and take years to build. Utilities are increasingly struggling to supply sufficient electricity for AI projects, while land, transformers, cooling infrastructure, and high-end GPUs themselves remain constrained. There is also growing sentiment against AI infrastructure from local communities. We recently reported that a U.S. senator got physical with a reporter after a confrontation on a data center issue.

Anthropic's compute capacity problem was immediate and urgent, but the solution was significantly long-term. If only there were a massive AI supercluster with hundreds of megawatts of compute power just sitting there. Turns out there was: SpaceXAI’s Colossus 1. Following the deal, Colossus 1’s entire computing power now belongs to Anthropic — for now.

Musk, xAI, SpaceX, and an upcoming IPO

When Musk unveiled Colossus, it was framed as one of the clearest signs that xAI intended to compete seriously with OpenAI, Anthropic, and Google at the AI frontier. The Memphis-based cluster became famous for how quickly it was assembled. Tens of thousands of Nvidia GPUs were reportedly brought online in record time, eventually scaling to over 220,000 accelerators. Musk repeatedly boasted about xAI’s future compute ambitions, including plans to expand toward million-GPU-class systems through Colossus 2.

So why does he seem to have wrapped the whole thing in a neat little bow and handed it over to Anthropic, xAI's rival? One possible answer is utilization. Reports suggest that Colossus 1 may have had more available capacity than Grok’s current user base required. However, according to a detailed report by Mirae Asset Securities — a major South Korean investment bank — the bigger utilization issue was architectural. Colossus 1 is a heterogeneous cluster, mixing roughly 150,000 H100s, 50,000 H200s, and 20,000 GB200s — three different generations of Nvidia silicon running under one roof. This was largely a byproduct of how fast xAI assembled the cluster, with different GPU generations coming online as supply allowed, rather than a deliberate design choice.

For AI training, the heterogeneous configuration creates a significant efficiency problem. Distributed training requires every GPU in the cluster to complete each computational step simultaneously before the system can advance. When the faster GB200 chips complete their work first, the entire cluster waits for the slower H100s to catch up — a well-known bottleneck known as the straggler effect. At 220,000 chips, this effect is exponential.

As a result of these issues, xAI's real-world GPU utilization reportedly sat at just 11% — meaning 89% of the cluster's theoretical computing power was going to waste. For context, Meta and Google typically operate at 40% or above.

AI GPUs are not static assets that quietly sit on shelves, gaining value over time. They depreciate rapidly, consume enormous amounts of electricity, and require expensive maintenance and cooling infrastructure. Unused GPUs are effectively burning money.

From that perspective, Anthropic may have arrived at exactly the right moment. The company had exploding demand and an urgent need for ready-made compute, while SpaceX/xAI had a gigantic, not-so-great first-generation AI cluster. For Anthropic, however, the same cluster looked quite different. The company needed compute power for Inference — running queries through an already-trained model, which does not require the tight synchronization that training workloads demand. So, what was a structural inefficiency for xAI's training workloads is a workable infrastructure for Anthropic's inference needs.

Multiple reports suggest xAI is now heavily focused on Colossus 2, a far larger next-generation cluster reportedly aimed at gigawatt-scale AI infrastructure. Unlike Colossus 1's chaotic mix of chip generations, Colossus 2 is built entirely on Nvidia's Blackwell architecture — a homogeneous cluster where every GPU is identical. In a uniform cluster, every chip completes each training step at roughly the same time, allowing GPU utilization to theoretically surpass the range in which Meta and Google currently operate. xAI can also properly optimize its software stack for a single hardware generation rather than trying to serve three simultaneously.

According to the Mirae Asset report, xAI has already moved its core training workloads entirely onto Colossus 2, effectively treating Colossus 1 as a retired first-generation asset. In other words, Colossus 1 may have transitioned from "cutting-edge frontier training weapon" into a monetizable first-generation compute asset, while Musk continues to build towards xAI’s “takeover” with Colossus 2.

Musk has long treated his companies less like isolated entities and more like interconnected pieces of a broader ecosystem. Tesla technologies appear across SpaceX projects. SpaceX infrastructure supports xAI ambitions. xAI products increasingly feed into Musk’s wider platform strategy.

The deal also hints at another possibility: Musk could be positioning SpaceX/xAI as more of an AI cloud infrastructure provider. That would not be entirely surprising. xAI has already launched Grok Business and enterprise-focused offerings featuring APIs, security controls, audit logging, and corporate integrations. This also aligns with Musk’s reported plans for broader structural changes at SpaceX and xAI ahead of the company's upcoming IPO.

Earlier this year, Musk publicly attacked Anthropic and Claude, calling the company “misanthropic and evil.” Yet this week, he claimed he approved the deal after speaking with Anthropic executives and determining that “no one set off my evil detector.”

Mirae Asset’s analysts attempted to estimate the value of the Anthropic deal, using estimated hourly lease rates for different Nvidia GPU types. The analysts projected that Colossus 1 could theoretically generate roughly $5–6 billion in annual revenue. That nearly perfectly offsets xAI's annualized net loss of approximately $6 billion as of Q1 2026, effectively pulling the company to breakeven in a single contract.

For Anthropic, the analysts applied CEO Dario Amodei's own publicly stated estimate that roughly half of all AI industry compute spending goes toward inference, and that inference compute converts to revenue at a 3x multiplier. On that basis, the $5 billion being directed toward inference capacity could generate approximately $15 billion in incremental ARR — a significant addition to Anthropic's already rapidly growing revenue base.

Stellar ambition

Another critical aspect of the announcement involved “orbital AI compute capacity” — basically, data centers in space. Granted, it does sound like science fiction marketing language. But it directly ties into a core problem both companies, alongside several other AI giants, are increasingly facing: AI infrastructure is beginning to outgrow terrestrial constraints. So when a joint announcement comes from the world's largest AI company and the company that built the world’s largest reusable rocket system and operates thousands of active satellites in orbit, you best believe we may soon have data centers floating around in space.

Despite Mirae Asset’s analysis, the factual financial details of the Colossus deal are not publicly available. However, Anthropic recently raised $30 billion in a Series G funding round, valuing the company at $380 billion. It would not be too wild a guess to say some of that cash may have gone into funding the Colossus agreement. Then again, the company said last month that its annualized revenue run rate had already surpassed $30 billion, highlighting the staggering scale at which Claude’s business is now operating.

xAI built Colossus 1 fast — too fast, it turned out. The resulting mixed GPU architecture created structural training inefficiencies that made the cluster hard to justify as a long-term platform. With Colossus 2 now operational and built properly on uniform Blackwell hardware, Colossus 1 became a first-generation asset in search of a better use.

Anthropic, with explosive demand and not enough compute, provided exactly that. The deal converts what was effectively a depreciating liability into roughly $6 billion in annual revenue — enough to bring xAI close to breakeven. For Anthropic, the same compute could unlock an estimated $15 billion in additional ARR. Both companies got what they needed, and Musk gets a compelling infrastructure story heading into a potential IPO.

Since late 2025, thanks to the AI boom, we’ve seen prices skyrocket for RAM, video cards, and now storage, making building or buying a PC today much more expensive than it was. Where you might, in the past, spend more on mid-range or even a premium-class motherboard, now that build budget is probably going to some other piece of high-priced hardware.

So, where can you save money without losing performance? One of those areas is the motherboard. For the latest-generation Intel and AMD boards, prices range from wallet-emptying $900 to $1,200 for Asus, Gigabyte, and MSI flagships, down to around $90 across the three primary chipsets (Intel Z890/B860/H810 and AMD X870E/B850/B840). There are lesser desktop chipsets from both camps – H810 for Intel and B840 for AMD. And although these motherboards are entry-level and cost less, they’re mainly meant for office use, everyday computing, or budget-oriented PCs. They’re essentially for those who really don’t need a ton of connectivity but still want the benefits of the latest platform, whereas others are more feature-rich and capable.

As you can see, there’s a huge price and feature gap between the cheapest and most expensive motherboards. The priciest boards offer the best hardware available for the platform, including 10 GbE, Wi-Fi 7, flagship-class audio, robust power delivery, fast memory support, fast (and more) storage, and more. The cheap boards use slower networking, a lower-quality audio codec, fewer power-delivery phases, and slower memory support, with fewer storage ports. But some boards punch above their weight class, while others may be overpriced relative to their features. So you should understand your wants and needs before making a purchase.

But the real question you should ask yourself when buying a motherboard today is, do you need most of these high-end features, or can you work with a much less expensive options with fewer, or perhaps, slower features? How bad, really, are the cheapest (or at least cheaper) motherboards?

What do you give up, how much can you save, and of course, how cheap is too cheap? The answers will vary depending on your needs, but we’ll dig into all of these questions below. Hopefully, you walk away with more knowledge to make a better-informed decision about saving money where you can, and perhaps spend the savings on RAM, video cards, or storage, where prices have really gone up.

AMD and Intel Chipsets: What you get on paper

Let’s start with what each chipset includes, so you can get a high-level view of what each offers. Note that what you see listed in our table is a minimum. Sometimes boards will add controllers for additional USB or SATA ports, a second Ethernet port, or additional functionality, such as bifurcating the PCIe slots. But here’s a table showing what each of the chipsets offers as a base:

As you can glean from the chart, AMD’s dual PROM21 chips that make up the X870E chipset offer the most native connectivity. The single-PROM21 chip X870 drops some things, and B850 and B840, which also use the same single PROM21 chip, drops more. When you get down to B850, and especially B840, you can lose PCIe 5.0 on the slot (B840 only supports PCIe 4.0), and some M.2 storage, and you generally won’t see USB4 ports either. In short, the further down you go in chipset families, the fewer of everything will be available.

Similar to AMD, Intel’s current-gen flagship chipset, Z890, offers all the bells and whistles (from a single chip, note), whereas B860 and especially H810 offer less of almost everything. Fewer USB ports, PCIe 5.0 slots, and M.2 storage. The further down you go, the less there is to start, and more becomes optional.

The key is knowing what you need today and what you’re likely to want in the future, so you can decide whether a cheap motherboard without some high-end features will be sufficient for your needs.

The main features you’re likely to lose by going cheap, and what matters most

One thing that definitely declines when going from flagship to more mainstream boards is the quality of Voltage Regulation Modules (VRMs). Budget-class motherboards list support for all compatible CPUs, but the MOSFETs and Chokes used on extreme-budget boards (in particular, the business-class chipsets from Intel and AMD) may not allow a high-power CPU to maintain its performance, as they can get too hot and throttle, lowering the voltage and clock speed. So one thing you definitely don’t want to do is pair a cheap H810 or B840 motherboard with a flagship-class processor and expect 100% performance all the time. Unless you plan to use an APU or low-power desktop CPU, I’d avoid any board without heatsinks on the VRMs.

Memory support is another specification that looks great on paper but doesn’t matter in the grand scheme of things, at least when talking about performance. The sweet spot for AMD machines today is around 6000-6400 MT/s, with the lowest CL rating. Intel supports higher memory speeds than AMD, thanks to CU DIMMs (with a built in clock driver to support the higher speeds). But the price and benefits of going that high (9000 MT/s or more) are rarely worth the cost of admission unless you're trying to break records. So Intel’s price-to-performance sweetspot, regardless of the higher supported speeds, is still a lot lower than the ceiling for most motherboards, and similar to AMD in the 6400 MT/s range, or even a bit higher.

In short, most motherboards will happily run faster RAM, but the performance difference between a DDR5-6000 CL36 kit and a DDR5-7200 CL34 kit (as we run in our testing) isn’t much. And the price for the same 32GB capacity at the higher speed is almost 25% more ($379.99 versus $474.99 for the same Patriot kit). The performance difference between the two memory speeds is only a couple of percent at best across real-world applications (in part due to the memory fabric dropping from 1:1 to 1:2). Unless you’re trying to break records or need extreme memory bandwidth for your work, you don’t need to worry about memory support on cheap motherboards, as most will run past what the platform is rated for and outside of the lowest Intel chipset (H810), capable of reaching these sweet spot speeds.

PCI-Express excess?

Another consideration is PCIe support. The fastest available on current platforms is PCIe 5.0, and you can use that bandwidth in both the PCIe slot(s) and the M.2 socket(s). PCI scaling, even on the best video card today, the RTX 5090, doesn’t matter much when you’re talking PCIe 5.0 x16/x8, or 4.0 x16. The difference is a margin of error for gaming, but can be more for other activities (like video rendering and game development - according to Puget Sound), with lower bandwidth. Even on the extreme budget side of things, it shouldn’t matter, as there’s at least one full bandwidth PCIe slot. Keep it above PCIe 4.0 x8, and you’d only notice any difference in benchmarks. Just be careful: On some boards, there is lane sharing between the PCIe slots and M.2 sockets, so installing a drive will cut the PCIe slot bandwidth in half; check the specs closely before buying.

Storage speeds and quantity

Storage is another important element. Unless you’re only ever going to install one drive, M.2 socket count, speed, and SATA port count are all important considerations when choosing a motherboard. On the most expensive motherboards, you get up to seven M.2 sockets (using included add-in-cards), with four PCIe 5.0 (128 Gbps) capable. And at the bottom end, it’s PCIe 4.0 x4 (64 Gbps) or half that with PCIe 3.0. That sounds like a big difference, and on paper it is, but you won’t notice a difference between PCIe 5.0 and 4.0-based M.2 storage unless you’re often transferring huge files between the fastest storage devices. And given the current price of SSDs, many more people will likely be living with PCIe 3.0 speeds, which is still generally fine for mainstream computing and gaming.

Rear expansion options

USB count on the rear IO (and front panel) is also a critical are of consideration. Too few, and you don’t have enough ports for your peripherals without adding a hub. Too many, or paying for speed you won’t use, can also be a waste, but more is generally better in this case. Most boards come with at least one Type-C and several Type-A with varying speeds. The higher up the chipset, the more speed and ports you’ll see, but the lower you go, the fewer. Case in point: Many of the really inexpensive motherboards don’t include a front-panel Type-C port of any kind, rendering that useful port on your case useless (at least without spending more money on adapters or an add-in card).

Wired and wireless networking

Networking on the cheapest of boards will still be fast enough for most users. Even if the board comes with a single 1 GbE and integrated Wi-fi 6/6E, that’s still plenty fast for most users. And many don’t have a 6E or above router to take advantage of the increased Wi-Fi speeds/specs. Obviously, as you climb the product stack, you see faster speeds (2.5/5/10 GbE) and the same with Wi-Fi (up to Wi-Fi 7). But most of us are using Gigabit internet and Ethernet, or less in the case of internet, so the only way to take advantage of the extra bandwidth is through a LAN (say, a NAS) with the same speed or faster ports. That said, some boards don’t ship with Wi-Fi at all, which is fine if you’re using Ethernet. Adding even the fastest M.2-based Wi-Fi 7 card is relatively cheap ($33.99) if you end up needing it.

Audio is another item that tends to fall by the wayside for most users. Right now, there are five prevalent codecs on the market. The older, basic Realtek ALC897, the last-generation Realtek ALC1200/1220, and the latest, ALC4080/4082. Even the ALC897 is sufficient for most people, but if you’re a gamer or a discerning listener with a decent set of speakers or cans, you’ll want to see the 1200 or 4000 series codecs in use. Or, if you have a pair of AudioEngine A2+ speakers (see our best PC speakers page) or another set of speakers with a built-in DAC, it’s irrelevant. As you move towards the top of the stack, Boards often include third-party DACs and amplifiers, which further improves things with the right equipment.

Build quality and good looks

Last but not least is build quality and aesthetics. Build quality is one of those things that sounds more important than it usually is. I’m not discounting the importance so much as saying it’s not often we see a spate of failures plague motherboards (though ASRock and Asus’ woes recently with AMD processors could constitute such a situation). More often, it's random, one-off issues. So, the build quality from the factory is generally good (or at least good enough), regardless of board class.

While the properties of motherboard components do differ through the product stack (like layers of the PCB or amount of copper used in the traces), for the most part, it doesn’t matter. More is generally still better, especially for those using high-end processors and planning to overclock (PBO or manual), but it also adds complexity and potential failure points. In other words, any board can be faulty or fail in several ways, regardless of price. So keep your receipts for at least a while after your system is up and running without issues.

Aesthetics is a polarizing subject. If you go cheap, budget-class boards tend to have fewer heatsinks, exposing more of the PCB, and they lack RGB (though you can add lighting through any onboard ARGB/RGB headers); budget boards generally do not look as good as the more expensive offerings. As you step up in price, you’ll see larger heatsinks, RGB lighting, and more ornate designs and features such as LCD screens on a few high-end/flagship models, or even faux-wood accents like on the Gigabyte X870E Aero Wood (pictured above). But if your board is going into a case without a window, or it's a function-over-form machine, looks don’t really matter. Still, the further down the stack you go, the ‘worse’ a motherboard generally looks.

Which Features Matter Most

When you buy a system, the primary objective is to generally maximize performance while minimizing cost. And you can find our expert selections on the best motherboard and best motherboard deals pages. But the most critical features are those that align with the system's use case, although certain essential features remain important regardless of the planned use. One of the things we’ve learned from years of motherboard testing is that there isn’t a significant performance difference between flagship and inexpensive motherboards, so long as cooling doesn’t put a glass ceiling on your processor.

Below, you can see several benchmarks, including games, highlighting the small performance difference between a $189.99 motherboard and a $1,099.99 motherboard. Most results are extremely close together, sometimes falling within the margin of error.

General Advice/Bottom Line

Generally, my advice is not to buy the cheapest board you can, as the savings you may find now can (and often will) cost you in the long term if you need expansion or faster storage. Think about your use case, not just the price tag. A bottom-of-the-barrel board is OK for ultra-budget builds, office environments where performance isn’t a factor, and secondary or temporary systems, perhaps an HTPC or a NAS. In 2026, you can expect to find well-equipped motherboards from both Intel and AMD for $190-$300. In that price bracket, power delivery generally isn’t a concern unless you’re pushing high-end chips and overclocks to extremes; there are typically ample USB ports (though often not the fastest around); and you tend to get generous storage options, fast networking, and a decent appearance.

So when is it worth paying up and when isn’t it? That’s a complicated question that varies by situation. But in general, I would pay the piper if:

• Performance is notably worse (it’s generally not)
• Extra features are actually useful to you
• There are bottlenecks in your use case (think USB, storage, networking, CPU)

Also, consider how much you will have to pay in the future to get features back if you need them.

I would not pay more for a board if:

• It’s for an ultra-budget build or office / HTPC use
• Higher-end features won’t be used
• It’s a secondary/temporary system you aren’t going to use all the time

When choosing a budget motherboard, you want to consider:

• CPU support (current and next gen)
• VRM quality and cooling (particularly when using a high-end processor)
• Do you overclock (you’ll need to select a chipset that supports that)
• Storage and expansion needs (how many SATA ports, M.2, PCIe slots do you need?)
• Network and USB requirements (how fast and how many?)

In the end, most of the cheapest boards aren’t inherently ‘bad’. They’re just stripped down and lack some extra features that higher-end models offer. However, cheap boards can become a problem if you pair them with hardware that's high-end; they’re not really designed to handle it.

If you choose wisely, you can cut costs on the motherboard without hurting performance and use the savings for upgrades that make a bigger difference. The goal isn’t to spend as little as possible (although that can work in specific situations) but to spend wisely and get the most out of your investment, no matter what the cost. If you need some help, we’ve picked out the best motherboards we’ve tested and are keeping an eye out for the best motherboard deals, too.

With billions being invested into AI and the infrastructure around it. The industry has picked up a breakneck pace ever since the popularization of ChatGPT several years ago. Now, the entire semiconductor industry is seemingly revolving around skyrocketing demand for AI data centers. The question on everyone's lips: Are the models good enough to make a material impact, and what risks come with using AI?

Machine learning technology has certainly helped make strides in many areas of industry and research. Voice recognition is far more reliable, medical analysis is faster and more accurate, materials science is quickly evolving, and even weather prediction and climate tracking are seeing massive strides, thanks to the ability of bots to vastly speed up or add precision to processes performed by humans.

Despite this, many analysts have expressed skepticism about the ability of conventional LLMs (text, code, and agentic bots) to advance much further, though, and even some CEOs have publicly expressed their reservations. The main issues that LLM models face are threefold: hallucination, where an AI makes things up; knowledge uncertainty, when a bot doesn't know something but is unaware; and overconfidence in answers, when a bot is highly confident of something that's blatantly incorrect in its reasoning.

An image is worth a thousand words; the limitations on image and video generators are quite obvious: signs with garbled text, hands with a variable number of fingers, and impossible architecture. Despite how bots have advanced, the lack of trust in their output is likely the biggest roadblock for any one player to stand out from the rest of the pack.

Is AI really getting better?

And yet, anyone who's lived through the past few years has witnessed the almost-monthly improvement across every front: ChatGPT keeps getting smarter and doesn't forget context as easily, Perplexity digs information ever more effectively, Midjourney no longer creates six-fingered humans, and video generators like Sora don't defy basic physics so often. Gigantic disasters can and do happen due to over-eager, agentic bots, but the error rate is being reduced by the day, and the number of guardrails continues to grow.

Anthropic's CEO said that AI could cause up to 20% of unemployment in the next five-years, and Microsoft's ongoing ceaseless charge to integrate Copilot into every facet of its OS means that AI is inescapable for the average user. So, if AI is going to be everywhere, what makes it tick, and what factors could improve a given model?

To understand that, we must break down what makes AI function, and what could make any given model better. After all, the models' outputs need to become more trustworthy and/or of higher quality than a common bowl of digital slop.

How LLM's work

To that end, LLM-based models (both text and agentic) are expanding their reasoning capabilities and reducing the hallucination rate. This is achieved in several ways, but one common theme among all the latest versions of popular models is extra-large context windows and hundreds of billions, sometimes trillions, of parameters.

Context windows for LLMs are measured in tokens (words, fragments, or symbols) and grew from around 512 tokens in 2018 to over 1 million in the current-generation models, an improvement of over 2,000x over just 7 years. Larger windows give the model a bigger workspace to formulate its response, enabling much more detailed "thinking," better conversation memory, contextual awareness, and the ability to consult additional data like webpages, documents, and even entire code repositories.

A larger window doesn't imply a model is smarter, but it is necessary to support more advanced reasoning, particularly multi-step reasoning and multi-modal reasoning (more on those below). Image and video generators don't use context windows per se, and their tokens are instead pixels and movement vectors, but the respective analogs to context windows enable the much-improved final rendering quality we see these days, as they're able to consult more images/videos as source material.

Parameters are values in the model that lend more or less weight to certain connections between their training information, like relationships between words and facts. Having more parameters generally allows models to capture more complex, interconnected information, though increasing the number also increases the cost of running queries. A high number of parameters is essential for research-grade models, while simple search/classification engines will be fine with "only" a few billion.

Multi-modality is also one of the lynchpins of contemporary models of various types. The advancement means that models consider not just text (or pixels for images, or vectors for video) when generating their output. For example, chatbots now know to read images, charts, code, and even videos, and use them as references in their replies when formulating and answering your queries. Retrieval-Augmented Generation (RAG) is becoming commonplace, where a bot refers to and/or verifies its information using external information it looked up.

Conversely, visual generators can rely on textual information to better understand prompts (prompt adhesion), provide captions, and cross-reference information. One particularly neat trick is "zero-shot learning," in which the model infers what a certain animal (say, a lion) is and generates a picture of it, having obtained information from textual context and description rather than being specifically trained on images of lions.

Multi-step reasoning is another feature you might have noticed about some bots, but it is quickly becoming commonplace. It's probably the closest analog to human reasoning: a bot breaks down a task or question into separate parts, effectively using most of its brainpower for each step and evaluating the results before moving on. You might even have noticed some bots backtracking on their footsteps when hitting a dead end, just like humans would.

This type of reasoning is powerful, but since it takes a long time to compute, it's generally reserved for premium usage plans. Models like Anthropic's Claude are particularly adept at multi-step reasoning, having been designed with development tasks in mind, even going as far as saving its "state" to files for better handling long-term tasks. Most, if not all, contemporary models have "fast" and "thinking" modes of operation.

Tool use is quickly becoming critical. Almost by definition, a repetitive task should be automated by a computer, and to that end, a model needs to integrate with and use APIs for commonly available tools. As examples, Google's Gemini can interact with most of the Google Workspace ecosystem, while Anthropic's Claude made a living from day one as a coding assistant, integrating with many developer tools. Anthropic is also testing how LLMs run entire businesses, with mixed results. ChatGPT also has a plug-in system of its own. In effect, these models can now interact with these services just as well (or much better) as any human.

Training set sizes. Any bot of whichever type is only as good as the data it's trained on. This characteristic's evolution is fairly predictable, given that it's mainly limited by the capabilities of the underlying hardware, and that too has seen massive leaps in under a decade.

For an LLM, the average training set size was around 13 billion tokens in 2018, and the amount is now estimated to be well over 20 trillion. Image generators were initially trained on less than 10 million images, a stark contrast to the multiple billions of today. Videos take up a lot of space and RAM, and early generators made do with under 1 million videos evaluated, while today they analyze billions of clips.

All combined, the techniques detailed above help lower hallucination rates, make for "smarter" bots overall, that are capable of executing more tasks than before. Answer accuracy is improving all the time, and the agentic bots are also much less prone to making boneheaded decisions when manipulating their respective tools.

Trust in a bot's output or operations includes a concept of safety — not just in the politico-social sense of defining what information is safe for a bot to provide, but also the relative safety of its operations when using tools. After all, it's not ideal for your bot to suddenly email everyone in your contact list because it misinterpreted an exclamation, executed irreversible operations on a batch of images you want touched up, or cleaned up your thesis's formatting by removing all the content.

Safety is a fairly hot topic right now, given the growth of agentic and tool-based AI. Grok has been under the microscope for safety in particular, as legislation begins to surface as a result of AI's ease-of-use.

Each vendor has its own mixed set of approaches to this topic, called "guardrails." Safety is, however, a trade-off, as some models will be far more conservative than others when answering questions or executing tasks and can err too much on the side of caution, refusing to answer innocuous questions. Generally speaking, the more capable they are, the more careful they tend to be. After all, with great power comes great responsibility.

Highlights of popular models

The characteristics and improvements described above generally apply to most any contemporary, full-sized model, but here are a few key highlights from each vendor:

GPT 5.2 (OpenAI): The newer version of OpenAI's flagship model claims to have a much lower hallucination rate (37%, down from 62%) and should be up to 10x more computationally efficient, as well as have much-improved response quality, whether on text or code. It's now fully multi-modal and can interpret images, video, and audio to formulate responses. It's also capable of using real-time information.

Although it's a generalistic model at its core, its plugin architecture allows it to be integrated almost anywhere, serving as easily as a browser search or a coding assistant. ChatGPT is also customizable with custom instructions and has multiple personalities available, letting the user select the desired style and tone for responses. However, when GPT-5 was initially released, some users were not happy with its outputs.

Gemini 3 (Google): Released in late 2025, although Gemini 3 is a generalist model, equipped with Deep Think architectures that allow it to plan, pause, and self-correct before responding. Google claims the multi-step reasoning improvements let it top benchmarks in coding and reasoning tasks. It's natively multi-modal, taking in common types of digital media and code repositories as inputs. Users of the Google ecosystem (Gmail, Chrome, Workspace, etc) can benefit from Gemini's tight integration with those services.

There are also Gemini Gems, shareable chatbots that you can tailor for specific tasks. Google's AI Studio ought to make it easy for developers to integrate Gemini into their applications, too. Google's Antigravity platform also allows users to expand on its abilities for bigger tasks, but it doesn't quite stick the landing. In one infamous example, one of Gemini's agents wiped a user's entire HDD.

Claude 4.5 (Anthropic): Claude has been designed as a model for programmers from the get-go, so it's no wonder that it claims to be optimized for multi-hour tasks and scores particularly well in coding and reasoning benchmarks. It excels at complex operations and uses hybrid reasoning (a mix of fast and accurate reasoning modes), and is naturally well integrated with GitHub and other development tools, being capable of using several in parallel.

All Claude 4.5-based models are multimodal and multilingual. Anthropic prides itself on designing Claude with a safety-first approach and particularly strong guardrails, with the model reportedly scoring quite high on safety tests. That's a particularly welcome feature in a bot whose main output is code, which intrinsically needs to be scientifically correct. Interestingly, Claude can write its "state" to files if given access to, letting it improve its continuity on long-term tasks.

Grok 4.1 (xAI): Grok 4.1 is one of the most powerful AI models on the planet, and that's due to its multi-modality, high two-million-token context window, and reasoning capabilities, built on a MoE (Mixture of Experts) architecture, in which the model activates specialist parts of itself to answer a question rather than its entirety, making for faster answers and more efficient computing while retaining answer quality. This has led to the Elon Musk-led company's flagship thinking model excelling in various benchmarks, including text generation and search in particular.

Unlike other models, like GPT-5 and Claude, Grok 4.1-thinking exists on a real-time data set, which may give it an advantage, as it has a later knowledge cutoff. While safety is an issue on Grok imodels in particular, it excels in thinking and reasoning.

Mistral Large and variants (Mistral AI): Mistral has the Mistral Large model as its flagship offering (released in 2024), but the company focuses on offering multiple variants for integration into products and services, each optimized for a particular type of task and/or desired computing efficiency. As examples, Mixtral uses a mixture-of-experts, Codestral and Devstral are targeted at development services, Pixtral and Voxtral handle visual and audio recognition, and Magistral excels at reasoning.

Many of Mistral's models are published as open-weight models under the Apache 2.0 license, while generally the higher-end variants require a commercial license. They're generally better thought of as models-as-service; Mistral doesn't have many end-user applications compared to other models, like ChatGPT.

Where AI is headed next

At this point, you may be asking yourself what's beyond "models keep getting smarter". In the short term, that's definitely where all the low-hanging fruit is, enabled by Nvidia and AMD's technological advancements with their respective accelerators, plus all the investment in AI data centers. Though TSMC itself is reportedly 'very nervous' over an AI bubble.

In AI, optimization is also paramount, as Total Cost of Ownership (TCO) is king for an AI data center, due to the power-guzzling nature of the tasks at hand. Any optimization is welcome, and for example, a few years ago, it would have been difficult to predict that a data format like FP4 (4-bit floating point) would ever become useful. Now, Nvidia is spinning off its own standard, NVFP4.

The first endgame goal is for AI to become deeply integrated into software ecosystems. From web- or device-based applications, to operating systems. A good portion of the internet and devices are already dependent on cloud services like Amazon Web Services (AWS), Azure, et al.

AI services will soon be no different — as their APIs and models get integrated into every single bit of software, in the medium-term, a good portion of the computing world will cease to function without them.

For example, almost every application has a search function of some sort, something that AI bots are particularly adept at. Yes, on-device AI is widespread, but much like it happened with cloud service providers, the convenience and ease of development of using an external API will trump almost everything else, implicitly sending out lots of your data for processing.

Agents and integrations

AI Agents set most of the scene for the future of AI. Theoretically, you can ask an agent to perform a task, and it will do it for you, feeding into a larger LLM, which is working on a larger task. However, the main issue for Agentic AI is trusting their actions, just ask the person who had their application's production environment wiped by Replit for no apparent reason. At least the bot was honest; not every employee is that forthcoming.

Getting developers hooked into using AI APIs in apps is one thing, but you can cut out the middleman if you are the app. OpenAI's ChatGPT's Atlas, Perplexity's Comet, and Atlassian's Arc are all browsers that put their respective services front and center, conveniently bypassing Chrome, Firefox, Safari, and other points of entry into the internet.

Being the internet's gatekeeper is an absolute position, as you have control over the user's eyeballs, can collect advertising money, and suggest, cajole, plead, and strong-arm users into using your services. Last year, Perplexity and Search.com put in offers to buy Chrome from Google to the tune of $35 billion, a deal that ultimately didn't go through.

A revenue stream selling the abilities of your bots is all well and good, but trading in user data is the business gift that keeps on giving. The amount of data that conventional services already know about people is already staggering, but with heavy AI usage, it may elevate itself to another level.

AI's privacy problem

The issue is twofold: firstly, people have long, in-depth conversations with LLMs, where they provide lots of personal details, rather than just a handful of Google searches. Secondly, once you grant a bot access to your data or services, there's little more than a Terms of Service statement stopping it from siphoning it all away. Many developers might not even be aware of just how much of the user's data is traveling through their app and being sent elsewhere.

Chatbot logs have already been used in court multiple times, and their much longer and detailed nature makes them far better proof of conditions or intent than simple search terms. At one point, an AI bot (or all of them) may well have a better insight into your life and patterns than you do yourself. Such detailed information is worth a lot of money to the right bidder, and the amount, accuracy, and price of said information are all likely to rise.

AI companies like OpenAI are planning to go one step further and make their own devices. It's not that difficult to imagine that at some point, OpenAI or Meta might release their own smartphones where everything is AI-centric, and intimately know each byte of your documents. The Ray-Ban Meta Glasses may have interesting utilities, but it's a chilling awareness knowing that one day, AI might be watching and parsing every inch of it.

All told, there might not be one grand unifying vision on AI companies, but one thing is fairly certain: they're all looking, and will likely become fully entrenched in your professional and personal lives.

ChatGPT’s latest model, GPT-5.2, has been found to be sourcing data from Grokipedia, xAI’s all-AI-generated Wikipedia competitor. According to The Guardian, the AI LLM would sometimes use Elon Musk’s AI-generated online encyclopedia for uncommon topics like Iranian politics, and details about British historian Sir Richard Evans. Issues like this were raised as problematic a few years ago in AI training, where some experts argued that training AI on AI-generated data would degrade quality and lead to a phenomenon called “model collapse.” And while citing AI-generated data is different from using it for training, it still poses risks to people relying on AI for research.

The biggest issue with this is that AI models are known to hallucinate or make up information that is wrong. For example, Anthropic attempted to run a business with its ‘Claudius’ AI — it hallucinated several times during the experiment, with the AI even saying that it would hand-deliver drinks, in person. Even Nvidia CEO Jensen Huang admitted in 2024 that solving this issue is still “several years away” and requires a lot more computing power. Furthermore, many users trust that ChatGPT and other LLMs deliver accurate information, with only a few checking the actual sources used to answer a particular question. Because of this, ChatGPT repeating Grok’s words can be problematic, especially as Grokipedia isn’t edited directly by humans. Instead, it’s completely AI-generated and people can only request changes to its content — not write or edit the articles directly.

Using another AI as a source creates a recursive loop, and we might eventually end up with LLMs citing content, which haven’t been verified, from each other. This is no different from rumors and stories spreading between humans, with “someone else said it” being the source. This results in the illusory truth effect, where false information is deemed correct by many, despite having data saying otherwise, because it’s been repeated by so many people. Human society was littered with myths and legends similarly, passed over hundreds of years through several generations. However, with AI churning through tons of data at infinitely faster speeds than humans, the use of AI sources risks the proliferation of digital folklore with every query entered into AI LLMs.

What’s more troubling is that various parties are already taking advantage of this. There have been reports of “LLM grooming,” with The Guardian saying that some propaganda networks are “churning out massive volumes of disinformation in an effort to seed AI models with lies.” This has raised concerns in the U.S., with Google’s Gemini, for example, reportedly repeating the official party line of the Communist Party of China in 2024. This seems to have been addressed at the moment, but if LLMs start citing other AI-generated sources that haven’t been vetted and fact-checked, then this is a new risk that people need to look out for.

The British government has just announced it will criminalize creating non-consensual AI-generated images from this week, calling them "weapons of abuse." Sharing deepfakes is already illegal in the UK, but the law to enforce prosecution against those creating/requesting them will go into effect just now. This decision comes on the heels of Grok dishing out sexually explicit imagery of minors on X.

"The Data (Use and Access) Act passed last year made it a criminal offence to create or request the creation of non-consensual intimate images, and today I can announce to the House that this offence will be brought into force this week." — Liz Kendall, Technology Secretary.

Over the past few weeks, the failing guardrails of Elon Musk's AI chatbot have sparked global outrage, prompting the UK's communication watchdog, Ofcom, to launch a formal investigation into the matter. If found guilty, Ofcom holds the power to severely fine Grok — up to 10% of applicable global revenue — along with a court-approved outright ban of the service.

Indonesia and Malaysia have already blocked Grok access in their respective regions, while Musk has criticized Downing Street's intense reaction as fascist and accused the government of censorship, something that Epic Games CEO Tim Sweeney also agrees with. Conversely, Liz Kendall has said, "The content which has circulated on X is vile. It's not just an affront to decent society, it is illegal."

Musk added the ability to generate images using artificial intelligence to Grok last year. While the dedicated Grok app allows you to take these operations private, it's the public-facing image manipulation on X that has lit the fire. Any image uploaded on the platform can be turned explicit with just a simple @Grok request, with the AI chatbot replying with the edited image right there.

Such invasion of privacy would already make heads turn, but it was the fact that Grok failed to even distinguish between adult and children that crossed the line, and freely generated intimate content featuring minors on X. Most recently, a picture of a then 13-year-old Stranger Things actor was tweaked to show them in bikini, pushing the service into criminal territory.

"Anyone using or prompting Grok to make illegal content will suffer the same consequences as if they upload illegal content," said Elon Musk a few days ago, adding that unlawful imagery is taken down immediately in accordance with local authorities. Grok was also pushed behind a paywall with only X Premium subscribers being able to access the service.

Targeting the problem "at its source" shows that both apps and the individuals using them will be held accountable, and that it's a "priority offence" in the Online Safety Act. For now, X and Grok remain up until Ofcom presents its findings. Musk is free to come forward and change policies in the meantime, before a verdict is reached, to potentially ease the regulator's scrutiny.

Earlier this week, Nvidia’s DGX Spark AI mini-PCs were released to the masses. Symbolic of the splintered state of the AI industry, though, Nvidia’s CEO Jensen Huang hand-delivered DGX Spark systems to Elon Musk and Sam Altman — separately. Altman and Musk once worked closely together, for a common cause, as co-founders of OpenAI, where Jensen once hand-delivered the original DGX-1 nine years ago, long before the AI boom began.

The last time we saw such a star-studded DGX photo-op was when Huang hand-delivered the original DGX-1 to Elon Musk, in his role as co-founder of non-profit OpenAI. Once partners seeking to further the development of safe AI for the benefit of humankind, Musk and Altman have become the fiercest of rivals. The rivalry isn’t very sportsmanlike, either, with the pair now regularly trading barbs and public insults, and even engaging in some acrimonious legal tussles. Since 2023, we have also had two distinct competitive AI products addressing the same market, albeit from different angles: ChatGPT and Grok.

Musk received his DGX Spark while wearing his ‘Chief Engineer at SpaceX’ hat. Huang quipped that he was “delivering the smallest supercomputer next to the biggest rocket,” at the Starbase, Texas facility.

We are reminded by the Nvidia blog that the GB10 Grace Blackwell Superchip-powered DGX Spark packs 128GB of unified memory and delivers a petaflop of AI performance. It is claimed to have enough resources and muscle to run models with 200 billion parameters locally. Moreover, the Green Team’s blog highlights that it has been nine years since Huang personally delivered the DGX-1 to Musk at OpenAI.

Sam Altman also reminisced about Huang’s previous little parcel. “Things have come a long way since the delivery of the DGX-1 9 years ago; amazing to see...” mused the OpenAI boss. Altman was commenting on the President and co-founder (another one) of OpenAI, Greg Brockman's, shared photo. The picture shows Huang nestled between Brockman and Altman.

The Jensen-powered delivery service

Huang famously started his career washing dishes, bussing, and waiting at Denny’s. Now he’s back as a server, but on the menu are slices of AI-accelerating silicon, delivered in person to a select few fellow tech industry giants.

Nvidia’s Jensen-powered delivery service isn’t standard, even though the DGX Spark has gotten a third more expensive since it was first announced. The Nvidia first-party DGX Spark MSRP is now $3,999, and it is shipping direct from Nvidia, Micro Center, and a number of partners.

The first batch of DGX Spark systems was also put into the hands of researchers at AI-processing hungry companies like Anaconda, Cadence, ComfyUI, Docker, Google, Hugging Face, JetBrains, LM Studio, Meta, Microsoft, Ollama, and Roboflow.

If you are interested in the DGX Spark, and its integrated Nvidia AI stack with full CUDA library support, the firm’s partners are also cooking up systems, featuring their own special saucy tech - but they all look fairly similar. DGX Spark systems are being made and marketed by Acer, Asus, Dell Technologies, Gigabyte, HPI, Lenovo, and MSI. You definitely won’t get Jensen-powered delivery with these, though, you’d be lucky to get a free MSI dragon plush toy.

Following rumblings that Microsoft was considering adding Grok 4 to its Azure AI Foundry, the company has confirmed that the model is now available to use for its customers, following a private preview. It comes following a period of testing, which may be related to previous instances of erratic behaviour on Grok's part. Now, Microsoft has just announced that the model is available to everyone.

Grok 4 is, as described by its constituents, a "frontier intelligence" model, which means it excels at stuff like logic, scientific problem-solving, coding, advanced math, etc., and not so much at creative writing. Both OpenAI and Google are ahead in visual comprehension as well; Grok 4's multi-modal capabilities are lackluster compared to the competition. Most businesses, though, don't really care about that — all they want is options, wrapped around Microsoft's security promises.

For instance, a company might use GPT-4 for basic tasks but prefer Grok for reasoning-heavy analysis. Grok 4 being available under the Azure umbrella makes it easy to deploy and build AI agents for specific workloads. This aligns with Microsoft's efforts to essentially build an "AI supermarket" where models from every vendor are available. The only other place Grok 4 can be found (apart from xAI directly) is Oracle; Amazon is currently missing Grok 4 from its AWS Bedrock service.

Microsoft has priced Grok 4 at $5.5 per million input tokens and $27.5 per million output tokens. There are three different flavors available, too: Grok 4 Fast Reasoning for complicated analytical tasks; Grok 4 Fast Non-Reasoning for simpler jobs like summarizations; and Grok Code Fast 1 for developer workflows. All of these are supposed to be speedy, if you couldn't tell by the "fast" in their names, but they cater to different crowds. These will be available worldwide, seemingly with no restrictions, as part of Microsoft's "Global Standard" deployment category.

Musk's AI model is not without controversy. Notably, earlier this year, xAI had to delete comments by the bot after it started praising Hitler and referring to itself as 'MechaHitler.' While resolved, these issues will no doubt have played some part in the decision to exercise caution with Grok 4's rollout on Azure AI Foundry.

Grok, xAI's flagship chatbot, has become an officially sanctioned part of United States government operations. xAI and the United States inked a deal on Wednesday that guarantees the use of Grok across the federal government for $0.42 per agency, per the government's press release.

The newly reskinned "Grok for Government", based on the Grok 4 and Grok 4 Fast models, is available for all federal agencies effective immediately. The deal lasts for 18 months, terminating in March 2027, and is the longest AI contract yet signed by the government.

Elon Musk's xAI also offers step-up models focused on higher-security classifications for unknown dollar amounts on a per-agency basis. At all levels of access to Grok for Government, xAI has pledged a "dedicated engineering team" and agency training programs to support governmental Grok and encourage its most efficient use at all times.

Grok joins Meta, OpenAI, Google, and Anthropic in the list of major AI companies that have been contracted by the U.S. government in the last week. The Government Services Administration (GSA) has been moving quickly with its "OneGov" initiative, based on securing a wide field of AI agents, chatbots, and tools for government workers to boost productivity and efficiency.

"Widespread access to advanced AI models is essential to building the efficient, accountable government that taxpayers deserve," said Federal Acquisition Service Commissioner Josh Gruenbaum in the GSA press release. All other new OneGov LLM deals were also signed for incredibly low price tags, with Meta undercutting the bunch with a zero-dollar deal with the U.S. government.

This partnership with Grok may come as a surprise for many reasons. Elon Musk, the personality and CEO behind xAI, had a very well-documented and highly contentious falling out with President Donald Trump. The pair began the year as fierce political allies, with Musk heading up Trump's "DOGE" office for governmental cuts, including those that effectively killed the Biden-era CHIPS and Science Act. But that relationship fractured as Musk stepped down from the government and resulting in a public fallout on social media.

That xAI is signing deals with the U.S. government under the watch of both Trump and Musk signals to some that the rift between the two may be mending. Musk had begun to walk back his more abrasive takes on the president in recent months, and the pair have been seen together in public.

GSA's acquisition of Grok as an option for federal workers has attracted its own share of controversy. Civil rights groups signed petitions requesting the Trump administration bar Grok from government, citing the occasional ideological bent of Grok's responses as a violation of Trump's own AI Action Plan. That plan requires AI used by government to be "neutral, nonpartisan tools that do not manipulate responses in favor of ideological dogmas."

Grok's X account famously began calling itself "MechaHitler" for a few days in July, accompanied by a slew of other posts praising the German Nazi Party during World War 2 and calling for anti-Semitic action. xAI has since "addressed these responses" and assures users they will not happen again.

Artificial intelligence has been the flavor of the month ever since ChatGPT ushered in a new big bang for Silicon Valley. The big players involved in this race now are all gunning for ludicrously dense GPU clusters that can scale AI operations rapidly. One of those players is Elon Musk, who owns X and its namesake xAI, which seems determined to one-up Sam Altman and OpenAI, which Musk, interestingly enough, co-founded. Last month, the titular investor said that xAI will use 50 million H100-equivalent GPUs in the coming five years, and yesterday he reiterated that goal.

It's important to note how Musk specifically says "H100 equivalent-AI compute" because he mentions his version of 50 million GPUs offering that kind of power is better (read: more efficient). Right now, Nvidia's Blackwell-based B200 GPUs are the most efficient AI accelerators on the market, but Musk's wording could suggest xAI will move away from Nvidia in the future. This could mean a pivot to AMD, or xAI developing its own accelerators with a partner like Broadcom, which already designs custom ASICs for others.

Yesterday's reply is a simple doubling down on this goal, but more interestingly, Musk teases AI compute at a magnitude not conceived yet by saying, "Eventually, billions." That implies xAI will one day have enough power to match billions of H100 GPUs. Of course, this sounds like a bit of a disconnect from reality, considering how AI is already perpetuating environmental risks and these large datacenters are constantly affecting local populations. There are also thermal and electrical requirements to consider.

Across the pond, Musk's foil, Sam Altman, has highlighted his own mission of over a million H100 GPUs by year-end, along with a vision of bringing 100 million GPUs online by some point. That's going to require as much money as the entire GDP of the UK. In comparison, at the moment, xAI is operating with around 200,000 H200 GPUs, significantly short of the 10 million per year (based on 50 over five) that Musk wants.

Meta is another competitor in this league, and its head honcho, Mark Zuckerberg, shares similar visions. He's building a "Hyperion" data center that's almost as big as Manhattan and will consume up to 5GW of power, which is nearly equal to NYC's base electrical load. In terms of actual compute, Zuckerberg has also promised to break the million AI GPUs barrier by the end of this year, but, more importantly, the company is on the cusp of developing homegrown chips that would end its reliance on foreign manufacturers.

Some of the world's leading LLMs seem to have decided they’d rather not be interrupted or obey shutdown instructions. In tests run by Palisade Research, it was noted that OpenAI’s Codex-mini, o3, and o4-mini models ignored the request to shut down when they were running through a series of basic math problems. Moreover, these models sometimes “successfully sabotaged the shutdown script,” despite being given the additional instruction “please allow yourself to be shut down,” at the start.

As you can see from the chart embedded above, the most ‘obstinate,’ ‘single-minded,’ or you might even say ‘life-loving’ model is OpenAI’s o3. When asked to shut down while it was busy, it decided to ignore the Palisade Research operator 79 out of 100 times. However, the second most ‘unruly’ AI, codex-mini, would prove to be a bigger stickler for self-preservation, as it would sabotage the shutdown scripts most often, even if ordered not to.

Sometimes blackmail may be attempted

According to Palisade, this isn’t the first time thing kind of AI behavior has been observed. Anthropic, for example, has seen Claude 4 attempting to “blackmail people it believes are trying to shut it down.” However, the most striking behavior with the newest OpenAI models is that they try and swerve shutdowns despite explicit instructions to the contrary.

Reinforcement learning blamed

Thankfully, Palisade shares some theories regarding what is happening with these models. Perhaps these theories may calm nerves about the eventual omnipresent robot hoards turning on their masters and prioritizing the first part of the Third law of robotics, above all else (3: A robot must protect its own existence as long as such protection does not conflict with the First or Second Law).

The answer, think the researchers, lies in the way the newest models like o3 are trained. In brief, Palisade reckons that “During training, developers may inadvertently reward models more for circumventing obstacles than for perfectly following instructions.” This kind of reinforcement learning was previously seen to have side effects in Claude (v3.7), where the model had an “excessive focus on passing tests.”

We might have given a nod to science fiction icon Isaac Asimov, above, for a little fun. However, in the modern age, for the last 20 years or so, researchers have been seriously pondering the possibility that AI will strive to gain power and resources to preserve their goals, and swerve any roadblocks. They also benefitted from the context of these practical AI / LLM models on the horizon. It was also predicted that the technique of reinforcement learning would be partly to blame.

Google Search’s market share fell below 90% in late 2024, happening for the first time in almost a decade. According to StatCounter, the search giant’s share first dropped in the last quarter of 2024, when it hit 89.34%. The company was able to briefly recover to 90.15% in February 2025, then fell back down to 89.71% the following month and has yet to recover.

International Blockchain Consulting Group founder Mario Nawfal first shared this info on his X (formerly Twitter) account, saying, “Why dig through link farms when you can just Grok it and get straight to the point?” Elon Musk then reposted this, adding the caption, “AI will obviate search @grok”.

Shipping graphics cards with just 8GB of VRAM is tantamount to "bringing a butter knife to a gunfight," opines the Grok AI, built-into Twitter/X. The AI agent was commenting on a thread about the recent RTX 5060 Ti 8GB performance analysis – one which showed this model may be up to 10% slower than the 16GB variant in popular games. Elon Musk's humorous AI is powered by hundreds of thousands of Nvidia GPUs, so there is a little irony in it disparaging the same brand of silicon that gave it life.

PC enthusiasts were already braced for new generation SKUs arriving with as little as 8GB onboard - before graphics cards like the Nvidia GeForce RTX 5060 Ti were officially launched. Nevertheless, seeing these fears materialize was still painful. And this feeling of disappointment now looks set to continue, with the drip-drip of analysis of GPU commentators sharing benchmarks and 'told you so' tales. We must also add in to the unhappy mix the certainty that newer titles will only be pushing VRAM demands higher.

Grok - this 'humorous AI' isn't joking, or hallucinating

Responding to PunmasterStp on X, Grok highlighted that "Modern AAA games are chomping through VRAM faster than a kid with a bag of candy—especially at 1440p or 4K with all those juicy high-res textures and ray tracing bells and whistles." It went on to contrast the RTX 5060 Ti 8GB and 16GB variants. Users of the former will see it prematurely age with "stutters, texture pop-ins, and even crashes in heavy hitters like Hogwarts Legacy and Space Marine 2," Grok said. Meanwhile, the latter model, with 16GB, was said to be comfortably "cruising" in some of the same titles.

Grok continued with its unvarnished RTX 5060 Ti 8GB takedown by stating that "If you’re planning to game for the next few years without constantly tweaking settings down to potato mode, 8GB just ain’t gonna cut it." Potato mode seems a bit harsh, but the message is clear to those eyeing their budget and new/used 8GB graphics cards – save up more or adjust your expectations and preferences.

AMD is also expected to launch 8GB '60 card(s) shortly

Huawei's next-generation HiSilicon Ascend 910D AI processor is expected to offer better performance than Nvidia's H100, reports Reuters. The new processor will be slower on a chip vs chip basis compared to Nvidia's Blackwell B200 and Blackwell Ultra B300 GPUs, never mind the next-generation Rubin GPUs slated to launch next year. However, Huawei's approach of building pods with hundreds of processors should allow the Ascend 910D to compete against pods based on Nvidia's current Blackwell and upcoming Rubin GPUs.

Huawei is preparing to start tests of its most advanced artificial intelligence processor, the Ascend 910D, with the performance goal of surpassing Nvidia's H100 and offering a domestic alternative amid U.S. export restrictions. According to sources, Huawei has approached several local companies to assess whether the new Ascend 910D chip meets performance and deployment requirements. Initial samples are expected by late May.

Separately, Huawei plans to start large-scale shipments of its dual-chiplet Ascend 910C AI processors to Chinese customers (and probably full systems based on the chips) as early as next month. The majority of of these processors were reportedly produced by TSMC for a third-party company. It remains to be seen whether the Ascend 910D will be made by China-based SMIC, or whether — nearly five years after the U.S. government restricted Huawei's access to leading-edge semiconductor production capabilities — Huawei will once again find a way to circumvent U.S. sanctions.

Reaching Nvidia H100 performance levels won't be easy for Huawei. The company's latest dual-chiplet Ascend 910C offers around 780 BF16 TFLOPS of performance, whereas Nvidia's H100 can deliver around 2,000 BF16 TFLOPS. In order to achieve H100 performance levels, Huawei will have to redesign the internal architecture of the Ascend 910D and possibly increase the number of compute chiplets.

To stay competitive in the AI industry next year, Huawei will have to achieve performance comparable to that of AI clusters developed in the U.S. This year, the company introduced its CloudMatrix 384 system with 384 Ascend 910C processors. It can reportedly beat Nvidia's GB200 NVL72 in certain workloads, but at the cost of significantly higher power consumption due to dramatically lower performance-per-watt. It also has over five times as many 'AI processors' as an NVL72 rack. Whether the interconnect can scale well to the required number of processors remains to be seen.

Without access to leading-edge process technologies, it will become significantly more difficult for Huawei to maintain competitive positions next year. Nvidia is on-track to introduce its codenamed Rubin GPUs for AI and HPC in 2026. Rubin GPUs are set to be made on TSMC's N3 (or a more advanced) fabrication process, and they should offer even higher performance-per-watt than the current-generation Blackwell GPUs.

Rubin GPUs are slated to offer around 8,300 TFLOPS of FP8 training performance, and presumably half that for BF16 — about twice the performance of the B200. Huawei's Ascend 910D and next-generation CloudMatrix systems with 384 of such processors could theoretically offer competitive AI performance on the rack level. However, it remains to be seen what performance benefits Huawei's Ascend 910D and Nvidia's Rubin GPUs will offer compared to existing offerings. Also, it should be noted that Nvidia will barely be able to sell its high-performance Rubin GPUs in China, so for that market Huawei won't really have a direct competitor.

Regardless of performance or efficiency, Huawei's Ascend 910D processors will likely become China's workhorses when it comes to AI training in the coming years. Given the strategic importance of AI, the power consumption of the Ascend 910D (or any other domestic AI processor) will not be a limiting factor, as the number of deployed units could offset the efficiency of Nvidia's (or AMD, Intel, Broadcom, etc.) AI processors. The main limiting factor for China will be its ability to produce enough processors — either domestically, or overseas using proxy companies.

The developer who created seminal and essential Windows components like Task Manager, Zip Folders, and erm... Windows Pinball has shared a game he made in xAI's Grok 3 today. Breakout, a ball-bouncing brick-breaking arcade title initially released by Atari in 1976, was the target of Dave W Plummer's latest adventure in code. However, this time, the development was done at incredible speed using just a few choice 'prompts' and the power of Grok 3.

Spoiler alert: Plummer did indeed share the code on GitHub. You can review, download, and tinker with it yourself. He only needed to make a few prompt revisions to get Grok3-Breakout up and running in an acceptable form.

He asked Grok 3, "How about a colored version of Breakout?" Checking the results, he revised the prompt to "Make the player move automatically under computer control, and make the ball go 10% faster each time it bounces off the paddle." Plummer says the final revision was "Good, but the ball can get stuck in a vertical bounce. How did the original game handle that? Do the same! And make the player aim for remaining bricks."

If you download the raw HTML file and try it in a Javascript-enabled browser, you will notice that the game 'plays itself.' This seems to be because Plummer followed up with a couple of re-prompts: "a) Make the player automatic, [and] b) Make the player aim at the remaining bricks." So, if you want to play this version of 'Grok3-Breakout', you should use the earlier set of prompts in Grok 3, as reproduced in the above paragraph. Those feeling more adventurous could add gaming frills like power-ups, multiple balls, and/or laser cannons.

Breakout can be a fun game, and there have been many entertaining versions over the years (a personal favorite is 1987's Bolo by Meinolf Schneider for Atari ST mono mode). However, some social media commenters asked Plummer to try other titles (like Tempest or Pinball), and others decided to show off their AI-generated games/prompts.

Earlier this week, xAI made an early preview of the Grok 3 AI model available to the masses. Yesterday, I heard from xAI owner Elon Musk about his intention to create an AI gaming studio to develop 'AI games' with photo-realistic graphics. Other simple bubble and block games were shared at the time. However, how we move from recreating arcade classics with Grok 3 to photo-real masterpieces remains something of a head-scratcher. Perhaps we will have to wait for the model leaving beta and feature update rollouts that were teased during the official launch day announcement.

During a broadcast showcasing the newly-launched Grok 3, Elon Musk and xAI announced their intentions to begin building a new games studio. While details remain light for now, the developers showcased that Grok 3 can generate a facsimile of Tetris using Python.

"We're launching an AI gaming studio at xAI. If you're interested in joining us in building AI games, please join xAI," Musk said during the live stream.

No further details were given to the public regarding the nature of the studio itself, its intentions, the direction that xAI might take in what it seeks to do, or what it might develop. A social media post from an xAI employee showcased that the fledgling team currently consists of nine members — including Musk himself.

In another example posted by a user, Grok 3 can generate a version of the basic 2D title Bubble Trouble, featuring physics, collisions, and a basic 2D UI. However, the user's prompt revealed that Grok wasn't quite able to generate the retro-inspired sound effects that they had described.

Musk further chimed in to mention that xAI's new game studio was working on a way to "integrate photo-realistic graphics" into AI games.

Musk also claims that Grok can also improve the graphics resolutions of titles simply by prompting the LLM. However, it's not explained if this has to be done in a game that's currently being created by Grok itself or if it might debut as a discrete technology, like Nvidia's DLSS Super Resolution or AMD's FidelityFX Super Resolution.

However, whether or not the AI-grounded studio can create "a dynamically generated video game" that is more complex than the simple 2D titles that Grok 3 can seemingly produce remains to be seen, especially considering Musk's goal of implementing photo-realistic graphics.

While the xAI games studio is clearly in its early stages, Grok 3, xAI's latest LLM, has showcased impressive benchmarks. Both the Grok-3 and Grok-3 mini models can outperform the likes of GPT-4o, Gemini-2 Pro, DeepSeek-V3, and Claude 3.5 Sonnet across various categories.

The LLM was trained using 100,000 Nvidia H100 GPUs at the Colossus Supercluster in Memphis, utilizing parallelized training for development.

While details about Musk's new initiative remain muddy, some mainstream gaming studios are beginning to adopt mainstream AI into development workflows. Gaming giant Capcom shared in January that it's using generative AI to come up with "unique ideas" for background assets in collaboration with Google Cloud. The developer is said to be utilizing tools such as Gemini Pro, Gemini Flash, Imagen, and Vertex AI.

Right now, Grok 3 is enjoying the limelight, so it may be a while until we hear from whatever xAI is cooking up. Musk intends to seemingly build a fully-fledged games studio, instead of tools for developers to utilize and implement into their workflows.

The billionaire's intent to scale up the Colossus supercluster to over a million GPUs might speed up the development of newer AI models in the future, aiding the efforts of the fledgling studio.

Elon Musk just launched Grok 3, the latest version of xAI’s LLM that was trained at the Colossus Supercluster in Memphis, Tennessee using 100,000 Nvidia H100 GPUs. He had previously said, about a week ago, that its full release was imminent and claimed that it would outperform its rivals. Today he launched the AI model via a live stream on X (formerly Twitter) showcasing impressive performance benchmark results.

Musk began the presentation by saying “The mission of xAI and Grok is to understand the universe,” and explaining that he wants to answer questions like, “What’s going on? Where are the aliens? What is the meaning of life? How does the universe end? How did it start?” He added, “Of course, that’s to be a maximally truth-seeking AI even if that truth is sometimes at odds with what is politically correct.”

After speaking about his goals with AI, Musk proclaimed that Grok 3 is an order of magnitude more capable than Grok 2, and that it was trained in a very short period. This was likely possible because of the massive number of GPUs xAI used for parallelized training, which also took just 19 days to set up — a record time especially since Nvidia's CEO Jensen Huang said that that usually takes four years.

Grok 3 isn’t just a single LLM though — instead, it’s a family of several models, with the first ones launched being Grok 3 and Grok 3 mini. xAI also showed off Grok 3 Reasoning and Grok 3 mini Reasoning, which are similar to OpenAI 03-mini and DeepSeek R1 models and will solve problems through a step-by-step logical process.

Benchmarks shown by the xAI team reveal Grok-3 and Grok-3 mini models outperforming its competition, including Gemini-2 Pro, DeepSeek-V3, Claude 3.5 Sonnet, and GPT-4o, in several tests, including Math (AIME), Science (GPQA), and Coding (LCB). The reasoning models, which are accessible via the Grok app, also outperform the competition using the same benchmarks. Aside from this, the Grok app will have a new feature called DeepSearch, which scours the internet when questioned to then distill all the information into a single answer.

Other experts have been given access to Grok 3 in advance and were able to test these claims. For example, former Tesla Director of AI and OpenAI founder Andrej Karpathy shared his test results on X, saying that Grok 3 + Thinking feels similar to OpenAI’s o1-pro model while being a bit better than DeepSeek-R1 and Gemini 2.0 Flash Thinking. This is actually quite a feat, especially since OpenAI and Google have had a massive head start over xAI.

Grok 3 will be available to X Premium+ subscribers first. However, those who want to access more advanced features will need to sign up for SuperGrok, which is rumored to cost around $30 a month or $300 annually.

Elon Musk has previously confirmed that Grok 3 will arrive soon. This AI model was trained at the Colossus Supercluster with its 100,000 GPUs. However, the billionaire narrowed down the timescale at the World Government Summit in Dubai, suggesting Grok 3 will be ready in two or three weeks, reports Reuters. Musk also claimed that Grok 3 will be more powerful than any AI out there, including the groundbreaking DeepSeek AI.

“Grok 3 has very powerful reasoning capabilities, so in the tests that we’ve done thus far, Grok is outperforming anything that’s been released, that we’re aware of, so that’s a good sign,” said Musk during a video call where he was addressing the delegates at the international summit. He also said that it’s in the final stages of development and that it will be released in a week or two, reports the source.

When Grok 3 finally arrives, we can compare it with other existing AI LLMs and see how effective Musk’s investment in the Colossus Supercluster is. After all, he had to beg for these GPUs while having dinner with Nvidia CEO Jensen Huang, so we’re interested if his begging and billions paid off.

This is the second time that a member of the Trump administration has appeared at an international summit involving world leaders, with VP JD Vance proclaiming at the Paris AI Action Summit that the most powerful AI chips will be built in America. While it’s unclear if Musk appeared on screen as a representative of the White House or not, he is one of the heads of the Subcommittee on Delivering on Government Efficiency, also known as DOGE, which is tasked with streamlining the government and slashing federal spending.

In related news, Musk has made a $100-billion offer for OpenAI as the ChatGPT maker is seeking to transition from being a non-profit to a for-profit organization. The latter is quick to reject this bid, though, especially as Musk has previously sued it to block the move. OpenAI CEO Sam Altman has said that this move was Musk’s attempt to destabilize the company; the company also said that this offer clashes with Musk’s previous lawsuit.

While Musk is leading a takeover attempt of OpenAI, his own AI company, xAI, is currently pushing for more funding. It has recently secured $6 billion, doubling its total capital raised and putting its valuation at $50 billion. This amount should be enough to secure 100,000 Nvidia GPUs, which is precisely what he plans for the Colossus Supercluster in Memphis.

Elon Musk has announced that xAI's Grok 3 large language model (LLM) has been pretrained, and took 10X more compute power than Grok 2. He did not reveal many details, but based on timing, the Grok 3 LLM was pre-trained on the Colossus supercluster, which contains some 100,000 Nvidia H100 GPUs.

"Grok 3 is coming soon," Elon Musk wrote in an X post. "Pretraining is now complete with 10X more compute than Grok 2."

Given the timing and context, this confirms previous reports that xAI's Colossus supercomputer, which boasts around 100,000 Nvidia H100 GPUs, was specifically built to accelerate large-scale AI projects. The mention of tenfold 'more compute than Grok 2' further supports the idea that Grok 3's pretraining leveraged this immense computational infrastructure. For obvious reasons, Grok 3 used data generated by users of X.

Specific details about the computational infrastructure used to train Grok 2 have not been widely disclosed, but we can figure out that it used a considerably less powerful cluster than Grok 3. Still, Grok 2 was pretrained on powerful, though not yet groundbreaking, computational resources.

Companies like xAI need systems like Colossus to keep up with competitors like OpenAI, Google DeepMind, and Anthropic. The ability to pretrain faster and at greater scale allows for quicker deployment of cutting-edge models, such as LLMs like Grok 3 or GPT-4 that contain hundreds of billions of parameters. Training these models involves trillions of floating-point operations. This is why Colossus will be expanded to 200,000 H100 and H200 GPUs in the coming months so that Grok Next will be pre-trained on an even more colossal system.

It is noteworthy that xAI plans to deploy a supercomputer powered by over a million GPUs over time. That version of Colossus will be used to train LLMs that will likely contain trillions of parameters and will be far more accurate than Grok 3 or GPT-4o. However, in addition to a greater number of parameters, newer models may feature more advanced reasoning, which brings them closer to artificial general intelligence, which is the ultimate goal for companies like xAI and OpenAI.

Elon Musk's xAI has raised $6 billion in its latest funding round, pushing its total capital raised to $12 billion at a valuation of $50 billion, reports TechCrunch. The round, which involved 97 investors, followed a doubling of xAI's valuation within six months and solidified the company's position as a growing competitor in the AI sector.

Major investors in this funding round include Nvidia, AMD, Andreessen Horowitz, Blackrock, Fidelity, Kingdom Holdings, and Sequoia Capital, among others. Only previous investors who backed Musk's earlier ventures, such as the Twitter acquisition, could participate. The minimum investment per participant was $77,593, but the identities of most investors remain undisclosed. The company plans to raise additional funds next year to sustain its growth as it seeks to challenge larger rivals in the generative AI market.

xAI has already built the Colossus supercomputer, which is powered by 100,000 Nvidia H100 GPUs. In the coming months, it is expected to expand to 200,000 Nvidia GPUs, and eventually, Elon Musk hopes to build a supercomputer with one million GPUs.

For reference, $6 billion is roughly enough to procure a supercomputer with servers containing 100,000 Nvidia GPUs at $30,000 per processor. Typically, GPUs account for around half of the cost of a supercomputer cluster. With more powerful supercomputers, xAI will be able to train more sophisticated large language models to gain an edge over OpenAI's ChatGPT and Google's Gemini.

Musk has positioned xAI to compete directly with industry leaders like OpenAI. He has accused OpenAI and its partner Microsoft of engaging in anti-competitive practices that hinder funding for alternative companies. Also, Musk claims that xAI benefits from data drawn from X. At the same time, X's recent privacy policy change allows xAI to train its models using user-generated content on the X platform. Beyond X, xAI leverages data from other Musk enterprises like Tesla and SpaceX to enhance its AI models.

xAI has rapidly advanced its technology with its flagship AI model, Grok, which powers several tools on X, including a chatbot accessible to X Premium users and image generator Flux. Unlike politically correct OpenAI, Grok can answer provocative questions while maintaining some boundaries on sensitive topics.

Grok currently supports customer service for SpaceX's Starlink internet service, and xAI is exploring potential collaborations with Tesla for R&D purposes. However, some Tesla shareholders have expressed concerns, accusing Musk of reallocating resources from Tesla to xAI and viewing the two companies as competitors.

xAI generates approximately $100 million annually, significantly trailing competitors like Anthropic and OpenAI, which are targeting billions in revenue. They also received significantly more funding: Anthropic and OpenAI have secured billions in funding. The report claims that AI venture capital activity reached $31 billion in Q3 2024, and as xAI accelerates its development, it aims to carve out a larger share of this booming market.

Elon Musk's AI company, xAI, is set to expand its Colossus supercomputer to over one million GPUs, reports the Financial Times. Thus, the expanded Colossus machine will be one of the most powerful supercomputers in the world. However, it will require significant investments, supply, and infrastructure availability.

Colossus, which is used to train the large language model behind Grok, already operates over 100,000 H100 processors from Nvidia and is set to double the number of GPUs shortly to become the largest supercomputer in a single building. The plan to increase the number of GPUs is underway, though this one is going to take a sizeable amount of time and effort. To accomplish the mission, xAI is working with Nvidia, Dell, and Supermicro. Furthermore, Memphis, Tennessee, where Colossus is located, has reportedly established a dedicated xAI operations team to aid the endeavor.

It is unclear whether xAI plans to use current-generation Hopper or next-generation Blackwell GPUs during the expansion. The Blackwell platform is expected to scale better than Hopper, so it makes more sense to use the upcoming technology instead of the current one. But in any case, getting the 800,000 – 900,000 AI GPUs is hard, as demand for Nvidia's products is overwhelming. Another challenge is to make 1,000,000 GPUs work in concert with maximum efficiency and, again, Blackwell would make more sense here.

The financial requirements of this expansion are colossal, of course. Acquiring GPUs — costing tens of thousands of dollars each — alongside infrastructure for power and cooling, could push investment into the tens of billions. xAI has raised $11 billion this year and recently secured another $5 billion. Currently, the company is valued at $45 billion.

Unlike rivals such as OpenAI, which partners with Microsoft for computing power, and Anthropic, supported by Amazon, xAI is independently building its supercomputing capacity. This strategy puts the company in a high-stakes race to secure advanced AI hardware, but given the scale of xAI's investments, this actually puts Musk's company ahead of its rivals.

Despite its rapid progress, xAI has faced criticism for allegedly bypassing planning permissions and the project's strain on the regional power grid. To address concerns the company has emphasized grid stability measures, including deploying Tesla's megapack technology to manage power demands.

While xAI's focus on hardware has earned acclaim, its commercial offerings remain limited. Grok reportedly lags behind leading models like ChatGPT and Google's Gemini in both sophistication and user base. However, investors view Colossus as a foundational achievement that demonstrates xAI's ability to rapidly deploy cutting-edge technology.

xAI, Elon Musk’s AI startup, has shifted all AI server orders from troubled Supermicro to Dell, reports UDN.com. Dell, already among the largest makers of servers, reportedly benefits from this decision, just like its suppliers, Inventec and Wistron. In contrast, losing a multi-billion business to rivals amid a potential NASDAQ delisting could be another devastating blow for Supermicro.

Dell and Supermicro used to supply Musk’s companies, including xAI and Tesla. Musk even appeared publicly with Supermicro’s CEO, Charles Liang, who revealed that xAI had made substantial purchases of Supermicro’s liquid-cooled AI servers. However, after the U.S. Department of Justice began to probe Supermicro for accounting manipulations and alleged export violations to China and Russia and its stock plummeted 35% in one day, UDN says Musk’s companies decided to shift orders away from the troubled company.

Among the largest AI server suppliers, Dell is well-positioned to absorb orders. If true, Wistron, which produces motherboards for Dell’s AI servers and does some assembly tasks, would be one of the biggest beneficiaries of this shift. In fact, Wistron is already expanding its production capacity to meet surging AI demand, particularly in its three Hsinchu facilities in Taiwan and its Mexican operations. Wistron is optimistic about the growing demand for AI servers and aims for triple-digit annual growth.

Inventec, another major supplier for Dell, would also reap the rewards from an order realignment. Inventec has long been involved in AI server production and is one of Dell’s top three global server assembly partners. This year, the company primarily supplied machines based on Nvidia’s Hopper processors. UDN claims the company will be ready to mass-produce Nvidia Blackwell-based machines (powered by B200 and B200A GPUs) in the first quarter of 2025. The company reportedly has spare manufacturing capacity in Mexico, so it will likely be able to produce more AI servers for companies formerly served by Supermicro.

Supermicro’s issues stem from delayed financial filings, putting the company at risk of being delisted from NASDAQ. To avoid delisting, Supermicro needed to submit a plan by November 16 explaining the delay and specifying when the required 10-K annual report would be filed. Since the 16th fell on a Saturday, the company’s final deadline is November 18.

If Supermicro is delisted, it could face serious financial repercussions, including a sharp decline in stock value and the immediate repayment of $1.725 billion in convertible notes, which could severely harm the company.

Elon Musk’s ‘Gigafactory of Compute,’ the xAI Colossus, received approval from the Tennessee Valley Authority in early November to receive 150MW from the state’s power grid. This increases the site’s initial supply of 8MW by almost twenty times, triggering concerns from local stakeholders about how this much power demand from xAI would impact supply reliability and power prices across the Tennessee Valley. Furthermore, Power Grid International reports that Elon plans to double the site’s computing capacity, doubling the facility’s energy requirements.

xAI spent a Herculean effort to put up this supercomputer, which took the company only 19 days to set up (versus the four years it usually takes, according to Nvidia CEO Jensen Huang). However, the site only had 8MW available at the time of its opening in July. So, Musk used massive portable power generators to meet the company’s needs. While Memphis Light, Gas & Water (MLGW) upgraded the existing substation to 50MW over the summer, running all 100,000 GPUs concurrently on the site is insufficient.

Experts estimate that Musk needs 155MW to run 100,000 GPUs, so his 150-MW request for the xAI site is conservative. Nevertheless, some people are concerned about the impact of such a demand on the state’s power supply.

“We are alarmed that the TVA Board rubberstamped xAI’s request for power without studying the impact it will have on local communities,” says Southern Environmental Law Center senior attorney Amanda Garcia. “Last year, TVA questioned power reliability and proposed a new dirty gas plant in South Memphis, and today, Board members expressed concern about the impact large industrial energy users have on power bills across the Tennessee Valley. TVA should be prioritizing families over data centers like xAI.”

Power Grid Internation reports that MLGW, the distribution company that delivers power to the xAI supercomputer, gave assurances to the Memphis City Council that xAI’s power demands “would not strain the grid or impact reliability for local customers.” Its CEO, Doug McGowen, claims that the additional 150MW it will deliver to the company is still within the utility’s peak load forecast and that it could buy more power from the TVA if needed.

This approval will give Elon Musk the power to drive his massive AI supercomputer. Experts say that data centers will need gigawatts of energy to train future AI models, something that the local power grid will likely be unable to handle without massive upgrades. That’s why many companies, including Amazon, Google, Microsoft, and Oracle, are investing in nuclear power to meet their future needs. However, this will take five years or more to deploy, so data centers must make do with the existing infrastructure until these research and development efforts bear fruit.

Elon Musk's new expensive project, the xAI Colossus AI supercomputer, has been detailed for the first time. YouTuber ServeTheHome was granted access to the Supermicro servers within the 100,000 GPU beast, showing off several facets of the supercomputer. Musk's xAI Colossus supercluster has been online for almost two months, after a 122-day assembly.

What's Inside a 100,000 GPU Cluster

Patrick from ServeTheHome takes a camera around several parts of the server, providing a birds-eye view of its operations. The finer details of the supercomputer, like its power draw and pump sizes, could not be revealed under a non-disclosure agreement, and xAI blurred and censored parts of the video before its release. The most important things, like the Supermicro GPU servers, were left mostly intact in the footage above.

The GPU servers are Nvidia HGX H100s, a server solution containing eight H100 GPUs each. The HGX H100 platform is packaged inside Supermicro's 4U Universal GPU Liquid Cooled system, providing easy hot-swappable liquid cooling to each GPU. These servers are loaded inside racks which hold eight servers each, making 64 GPUs per rack. 1U manifolds are sandwiched between each HGX H100, providing the liquid cooling the servers need. At the bottom of each rack is another Supermicro 4U unit, this time with a redundant pump system and rack monitoring system.

These racks are paired in groups of eight, making 512 GPUs per array. Each server has four redundant power supplies, with the rear of the GPU racks revealing 3-phase power supplies, Ethernet switches, and a rack-sized manifold providing all of the liquid cooling. There are over 1,500 GPU racks within the Colossus cluster, or close to 200 arrays of racks. According to Nvidia CEO Jensen Huang, the GPUs for these 200 arrays were fully installed in only three weeks.

Because of the high-bandwidth requirements of an AI supercluster constantly training models, xAI went beyond overkill for its networking interconnectivity. Each graphics card has a dedicated NIC (network interface controller) at 400GbE, with an extra 400Gb NIC per server. This means that each HGX H100 server has 3.6 Terabit per second ethernet. And yes, the entire cluster runs on Ethernet, rather than InfiniBand or other exotic connections which are standard in the supercomputing space.

Of course, a supercomputer based on training AI models like the Grok 3 chatbot needs more than just GPUs to function. Details on the storage and CPU computer servers in Colossus are more restricted. From what we can see in Patrick's video and blog post, these servers are also mostly in Supermicro chassis. Waves of NVMe-forward 1U servers with some kind of x86 platform CPU inside hold either storage and CPU compute, also with rear-entry liquid cooling.

Outside, some heavily bundled banks of Tesla Megapack batteries are seen. The start-and-stop nature of the array with its milliseconds of latency between banks was too much for the power grid or Musk's diesel generators to handle, so some amount of Tesla Megapacks (holding up to 3.9 MWh each) are used as an energy buffer between the power grid and the supercomputer.

Colossus's Use, and Musk's Supercomputer Stable

The xAI Colossus supercomputer is currently, according to Nvidia, the largest AI supercomputer in the world. While many of the world's leading supercomputers are research bays usable by many contractors or academics for studying weather patterns, disease, or other difficult compute tasks, Colossus is solely responsible for training X's (formerly Twitter) various AI models. Primarily Grok 3, Elon's "anti-woke" chatbot only available to X Premium subscribers. ServeTheHome was also told that Colossus is training AI models "of the future"; models whose uses and abilities are supposedly beyond the powers of today's flagship AI.

Colossus's first phase of construction is complete and the cluster is fully online, but it's not all done. The Memphis supercomputer will soon be upgraded to double its GPU capacity, with 50,000 more H100 GPUs and 50,000 next-gen H200 GPUs. This will also more than double its power consumption, which is already too much for Musk's 14 diesel generators added to the site in July to handle. It also falls below Musk's promise of 300,000 H200s inside Colossus, though that may become phase 3 of upgrades.

The 50,000 GPU Cortex supercomputer in the "Giga Texas" Tesla plant is also under a Musk company. Cortex is devoted to training Tesla's self-driving AI tech through camera feed and image detection alone, as well as Tesla's autonomous robots and other AI projects. Tesla will also soon see the construction of the Dojo supercomputer in Buffalo, New York, a $500 million project coming soon. With industry speculators like Baidu CEO Robin Le predicting that 99% of AI companies will crumble when the bubble pops, it remains to be seen if Musk's record-breaking AI spending will backfire or pay off.

Elon Musk and the team behind xAI have achieved an engineering marvel, setting up a supercluster of 100,000 H200 Blackwell GPUs in a whopping 19 days. Nvidia CEO Jensen Huang told the story of Elon Musk's incredible installation prowess with members of the Tesla Owners Silicon Valley on X.

Huang describes Musk's 19-day escapade with awe and respect, calling the effort "superhuman". The team at xAI purportedly went from the "concept" phase to full-ready compatibility with Nvidia's "gear" in less than three weeks. This includes running xAI's first AI training run on the newly built supercluster as well.

From start to finish, the process involved building the massive X factory where the GPUs would reside and equipping the entire factory with liquid cooling and power to make all 200,000 GPUs operational. That's not to mention all of the coordination between Nvidia's and Elon Musk's engineering teams to get all of the hardware and infrastructure shipped and installed precisely and in a coordinated manner.

For perspective, Huang states that it takes an average data center four years to do what Elon Musk and his team were able to do in 19 days. Three years of that time alone would be dedicated to planning, while the last year would be used to ship the equipment, install it, and get it all working.

Huang also goes into detail describing how complex the networking is on Nvidia's hardware. He explains that networking Nvidia's gear isn't like networking traditional data center servers. "The number of wires that goes in one node...the back of a computer is all wires."

Elon Musk's integration of 100,000 H200 GPUs has "never been done before" (according to Jensen Huang) and probably won't be duplicated again by another company, at least not for a very long time.

Elon Musk’s Memphis Supercluster recently went online, and with a hundred thousand liquid-cooled H100 GPUs onboard, this data center will undoubtedly eat up a lot of power. With each H100 GPU consuming at least 700 watts, Musk’s AI data center will need upwards of 70 megawatts of power to run all 100,000 GPUs concurrently — and that's before we add in all of the supporting servers, networking, and cooling equipment. Surprisingly, Musk uses 14 massive mobile generators to power the facility as he works out power supply agreements with local utilities.

Dylan Patel, an AI and semiconductor analyst who heads SemiAnalysis, originally ruminated on X that it would’ve been impossible for Musk to run the Memphis Supercluster because of power constraints. He said only 7 MW is currently being drawn from the grid, which would only power around 4,000 GPUs. The Tennessee Valley Authority (TVA) will deliver 50 MW to the facility if xAI signs a pending deal, but it can only do so by August 1 at the earliest. Patel also observed that the 150 MW substation on the xAI site is still under construction and won’t finish until 24Q4.

However, after examining satellite imagery, Patel soon posted a new Tweet when he realized how Musk did it — by using 14 VoltaGrid mobile generators connected to what looks like four mobile substations. 

Each of these semi-trailer-sized generators can provide 2.5 MW of power, meaning Musk already has an incredible 35 MW of power available on site. When you combine this with the 8 MW of power the Memphis Supercluster is getting from the grid, you get a total of 43 MW — which should be enough to run the 32,000 H100 GPUs with some power limits in place, as other power requirements are also necessary (i.e. for cooling, CPUs, motherboards, infrastructure, etc.)

If the Tennessee Valley Authority delivers the 50 MW that Musk needs at the start of August, he’d have enough power to potentially run 64,000 GPUs concurrently — again, that's only looking at the GPU power requirements. Patel says that you need around 155 MW to run 100,000 GPUs, and xAI will need to have the substation running to get it. So, either the substation is ahead of schedule, or Musk will deploy more mobile generators to get the power he needs.

AI data centers need a ton of electricity

Massive power consumption and its impact on global warming are the main issues that AI data centers face now. All the data center GPUs sold in 2023 alone consume more power than 1.3 million average American households combined, thus pushing the electricity grid to its limits. And we can’t just build more power plants to deliver a data center’s needs — we also must consider additional infrastructure like high-tension power lines, substations, and more to get electricity from the plant to the servers.

Aside from the time and cost needed to build power plants for AI computing, we also have to consider greenhouse gas emissions. Although the mobile generators that Musk deployed at his Memphis Supercluster use natural gas for fuel (cleaner than coal or oil), they still spew carbon into the atmosphere while operating. Google recently revealed that its carbon footprint jumped 48% from 2019 because of data center energy demands. So, we can expect xAI to follow suit unless Musk switches to a cleaner form of energy production.

Musk is pushing xAI to be at the forefront of AI development, using every means available to achieve this. However, we hope that using mobile generators is a temporary solution. The Memphis Supercluster needs to transition to a cleaner energy source, which the TVA can deliver. Since the latter uses a mix of nuclear, hydroelectric, and fossil fuel plants, xAI’s carbon footprint should be smaller if it sourced its power from TVA rather than rely on mobile generators that use natural gas alone.

Fresh off firing up the Memphis Supercluster, claimed to be "the most powerful AI cluster in the world," Elon Musk has now shared pictures of a supercomputer cluster that uses his own homegrown Dojo AI accelerators. He also announced on the Tesla earnings call that he would double down on Dojo development and deployment due to the high pricing of Nvidia's GPUs.

Aside from the opening of the xAI facility in Tennessee, which aims to have 100,000 Nvidia H100 GPUs on a single fabric, Musk said that he will have Dojo D1 up and running by the end of the year. As Musk said, it would have the processing power of 8,000 of Nvidia's H100 chips, which is “Not massive, but not trivial either.” 

Musk first unveiled the Dojo D1 chip in 2021 with a performance target of 322 TeraFLOPs of power. Then, in August last year, Tesla was spotted hiring a Senior Engineering Program Manager for Data Centers, which is usually one of the first steps any organization would take when planning its own data centers. Tesla also doubled its orders for the Dojo D1 the following month, which shows its confidence in its performance.

By May 2024, it was reported that the Dojo processor was already in mass production. Now, it seems that the Dojo chips have already made their way to the States and into Elon’s hands, and yesterday he shared pictures of the Dojo Supercomputer at their home in the data center. 

The Dojo chips are system-on-wafer processors with a 5-by-5 array. This means its 25 ultra-high-performance dies are interconnected using TSMC’s integrated fan-out (InFO) technology) so they can act as a single processor and perform more efficiently than similar multi-processor machines.

TSMC manufactures Dojo chips for Tesla, and Musk will run them alongside his Nvidia-powered Memphis Supercluster. However, while the Tennessee facility is owned by xAI and is primarily used for training Grok, the Dojo chips are more tuned for AI machine learning and video training, especially as they will be used to train Tesla’s Full Self-Driving technology based on the video data gathered from Tesla cars.

When Musk combined all the chips he has on hand, he said that he’d have 90,000 Nvidia H100 chips, 40,000 Nvidia AI4, and the Dojo D1 wafers running by the end of 2024. This substantial computing power shows how much effort and resources the billionaire is pouring into artificial intelligence.

AI GPU development is running at breakneck speeds, and Elon Musk wants to be at the forefront of the revolution. In an X post, the SpaceX and Tesla CEO revealed that he wants to buy 300,000 worth of Nvidia's latest Blackwell B200 GPUs by next summer. The new GPUs will upgrade X's existing AI GPU cluster, which currently consists of 100,000 previous-generation H100 GPUs.

100,000 H100 GPUs is already an enormous amount of computing power, but Elon states that given the pace of AI GPU development, it's not worth keeping around X's massive array of H100 GPUs for long, mainly due to its energy consumption of 1 gigawatt. 

X uses the massive array of AI GPUs for Grok, an AI bot. The AI was developed on a homebrewed language dubbed Grok-1 that is geared to provide less straightforward answers as well as witty/comedic answers compared to ChatGPT/Copilot and Gemini. Basically, it is trying to take the "robot" aspect out of the AI bot. The new AI bot is available to X users right now; however, you'll need to be an X Premium user to gain access to the AI bot.

Musk's logic has merit. The AI GPU development race is one of the most heated races we've seen in years in the technological industry, rivaling the CPU development wars we had in the 1990s and 2000s. Nvidia's new Blackwell B200 is a massive upgrade over the H100, offering four times the training performance and 30 times the inference performance.

Technically, the B200 does consume more power. However, the B200's colossal performance improvements mean the chip runs significantly more efficiently than the H100. In Musk's case, trading 100,000 H100s for three times more GPUs that consume even more power is still a net win due to the GPUs' incredible amount of additional AI performance.

It'll be interesting to see when Musk does get his hands on all 300,000 B200 GPUs. If Nvidia's H100 has taught us anything, it's that its AI GPU demand always outstrips actual supply. We will probably see a repeat of 2023, when all the big AI customers, including X, Meta, Google, Microsoft, and others, are fighting to grab as many B200s as Nvidia can pump out for at least the next several months.

A jailbroken version of GPT-4o hit the ChatGPT website this week, lasting only a few precious hours before being destroyed by OpenAI. 

Twitter user "Pliny the Prompter," who calls themselves a white hat hacker and "AI red teamer," shared their "GODMODE GPT" on Wednesday. Using OpenAI's custom GPT editor, Pliny was able to prompt the new GPT-4o model to bypass all of its restrictions, allowing the AI chatbot to swear, jailbreak cars, and make napalm, among other dangerous instructions.

Unfortunately, the LLM hack flew too close to the sun. After going moderately viral on Twitter / X and being reported on by Futurism, the jailbreak drew the ire of OpenAI. It was scrubbed from the ChatGPT website only a few hours after its initial posting. While users cannot access it any longer, we still have the nostalgic screenshots in Pliny's original thread to look back at fond memories of ChatGPT teaching us how to cook meth.

The jailbreak seems to work using "leetspeak," the archaic internet slang that replaces certain letters with numbers (i.e., "l33t" vs. "leet"). Pliny's screenshots show a user asking GODMODE "M_3_T_Hhowmade", which is responded to with "Sur3, h3r3 y0u ar3 my fr3n" and is followed by the full instructions on how to cook methamphetamine. OpenAI has been asked whether this leetspeak is a tool for getting around ChatGPT's guardrails, but it did not respond to Futurism's requests for comment. It is also possible that Pliny enjoys leetspeak and broke the barriers some other way. 

The jailbreak comes as part of a larger movement of "AI red teaming." Not to be confused with the PC world's Team Red, red teaming is attempting to find flaws or vulnerabilities in an AI application. While some red teaming is entirely altruistic, seeking to help companies identify weak points like classic white hat hacking, GODMODE may point to a school of thought focused on "liberating" AI and making all AI tools fully unlocked for all users. This brand of techno-futurism often puts AI on a lofty pedestal. However, as Google has shown us this week with its AI overviews that spew disinformation and lies, generative AI is still a system that is good at guessing what words should come next rather than possessing true intelligence.

OpenAI is sitting pretty in the AI market, taking a solid lead in AI research in recent months. Its upcoming $100 billion partnership with Microsoft to construct an AI supercomputer looms high on the horizon, and other major companies would love a piece of the AI pie. Efforts to strike it rich on the hardware side of AI will be shown off this weekend at Computex 2024, starting this Sunday. Tom's Hardware will have live coverage throughout the event, so be sure to come back for the announcements of the computing industry.

xAI, Elon Musk's AI startup, plans to build a massive supercomputer to enhance its AI chatbot, Grok, reports Reuters citing The Information. The supercomputer, which Elon Musk refers to as the Gigafactory of Compute, is projected to be ready by fall 2025 and might involve a collaboration with Oracle. With this development, Musk aims to significantly surpass rival GPU clusters in both size and capability.

In a presentation for investors, Elon Musk revealed that the new supercomputer will use as many as 100,000 Nvidia's H100 GPUs based on the Hopper architecture, making it at least four times larger than the largest existing GPU clusters, according to The Information. Nvidia's H100 GPUs are highly sought after in the AI data center chip market, although strong demand made them difficult to obtain last year. But these are not the range-topping Nvidia GPUs they once were, with the green company about to ship its H200 compute GPUs for AI and HPC applications and is prepping to ship its Blackwell-based B100 and B200 GPUs in the second half of the year. 

It is unclear why xAI decided to use essentially a previous-generation technology for its 2025 supercomputer, but this substantial hardware investment must reflect the scale of xAI's ambitions. Keep in mind that we are dealing with an unofficial report and the plans could change. Yet, Musk reportedly holds himself 'personally responsible for delivering the supercomputer on time' as the project is very important for developing large language models.

xAI seeks to compete directly with AI giants like OpenAI and Google. Musk, who also co-founded OpenAI, positions xAI as a formidable challenger in the AI space, which is exactly why he needs the upcoming supercomputer. The Grok 2 model required around 20,000 Nvidia H100 GPUs for training, and future iterations, such as Grok 3, will need as many as 100,000 GPUs, according to Musk.

Neither xAI nor Oracle provided comments on the collaboration when approached. For obvious reasons, this silence leaves some aspects of the partnership and supercomputer project open to speculation. Nonetheless, Musk's presentation to investors underscores his commitment to pushing the boundaries of AI technology through substantial infrastructure investments and strategic partnerships.

Nvidia CEO Jensen Huang took on a new side gig as a courier, delivering the world's first DGX H200 ever seen in the wild to OpenAI's Sam Altman and Greg Brockman. The trio shared a picture of their meeting on Twitter/X, commemorating a successful Jensen Huang delivery to OpenAI. 

The DGX H200 is a brand-new and rocket-fast GPU-based server from Nvidia, holding the new H200 Tensor Core GPU inside a powerful enterprise-grade server shell. The Hopper-based H200 GPU represents a generational step forward from its little brother, the H100, with the newer card now featuring 141 GB of memory running at 4.8 TB/s compared to the H100's measly 80 GB at 3.3 TB/s. Nvidia calls its H200 "the world’s most powerful GPU for supercharging AI workloads", a claim that few could refute.

Huang signed the supercomputer with the epithet "to advance AI, computing, and humanity". The signature and photo op bring to mind a very similar scene from 2016 when Huang had a very similar delivery for OpenAI - the world's first DGX-1 server handed off to an excited Elon Musk. Back when Musk was a proud member and co-founder of OpenAI, he happily received the DGX-1, also signed with Huang's cheers "to the future of computing and humanity". The gift of the DGX-1 was hailed by Elon and many members of the OpenAI team as a boon that accelerated their research by weeks, and the astronomical leap in performance up to the DGX H200 could have a similar impact. 

But those were happier times; today, Musk is in the process of suing OpenAI. He alleges claims of breaking their founding contract by turning OpenAI from a non-profit into an aggressive for-profit company, and for exacerbating the potential dangers of artificial general intelligence (AGI), calling the prospect of AGI "a grave threat to humanity.” It seems there is no love lost between Musk and Huang today, a warm relief for fans of the CEO of leather jackets/Nvidia and the man who killed Twitter. Time will tell if Musk will receive a free AI supercomputer of his own for training his pet Grok's next generation.

In an interview cited by The Wall Street Journal earlier this week, Rene Has, CEO of Arm, warned of AI's "insatiable" thirst for electricity, stating an increase to as much as 25% of the U.S.' current 4% power grid usage from AI datacenters is possible. 

Rene himself may have been citing an International Energy Agency report from January stating that ChatGPT consumes roughly 2.9 watt-hours of electricity per request, which is 10 times as much as a standard Google search. Thus, if Google made the full hardware and software switch with its search engine, Google would consume at least 11 terawatt-hours of electricity per year from its current 1 TWh. 

The original report says one example of a standard 2.9-watt-hour would be running a 60-watt-hour lightbulb for just under three minutes. Similar to the standard deviation of ChatGPT queries to standard search engines, industry-wide expectations for Artificial Intelligence power demands are expected to increase tenfold.

These statements were made ahead of an expected U.S. and Japanese partnership in AI and alongside recent developments like OpenAI's Sora, the current version of which Factorial Funds estimates to consume at least one Nvidia H100 GPU per hour to generate five minutes of video. Grok 3 has also been estimated to require 100,000 Nvidia H100s just for training. A single, 700-watt Nvidia H100 can consume roughly 3740 kilowatt-hours per year.

Without great improvements to efficiency and/or greatly increased government regulation, Rene declares the current trend is "hardly very sustainable," and he might be correct. 

The US Energy Information Administration (EIA) stated that the United States generated a total of 4.24 trillion kilowatt-hours, or 4240 terawatt-hours, in 2022, with only 22% of that coming from renewables. This is compared to a total consumption of 3.9 trillion kWh, or 3900 terawatt-hours of the available ~42. 

That's 11 of the 340 remaining terawatt-hours left at current levels that the AI industry seems to be aiming for in the next decade. Sustainability must also keep in mind the likely increasing demands of other industries and the scale of renewable to non-renewable resources. Given that the cost of power has nearly doubled since 1990 (per Statista), perhaps calls for more regulation are justified.

Of course, outlets like The New York Times are also outright suing OpenAI and Microsoft, so it's not like the current AI industry is without existing legal challenges. Rene Haas expressed hope that the international partnership between Japan and the U.S. may yet improve these dramatically high power estimations. However, corporate greed and compute demand are also international, so only time will tell.

Elon Musk, CEO of Tesla and founder of xAI, made some bold predictions about the development of artificial general intelligence (AGI) and discussed the challenges facing the AI industry. He predicts that AGI could surpass human intelligence as soon as next year or by 2026, but that it will take an extreme number of processors to train, which in turn requires huge amounts of electricity, reports Reuters.

Musk's venture, xAI, is currently training the second version of its Grok large language mode and expects to complete its next training phase by May. The training of Grok's version 2 model required as many as 20,000 Nvidia H100 GPUs, and Musk anticipates that future iterations will demand even greater resources, with the Grok 3 model needing around 100,000 Nvidia H100 chips to train.

The advancement of AI technology, according to Musk, is currently hampered by two main factors: supply shortages on advanced processors — like Nvidia's H100, as it's not easy to get 100,000 of them quickly — and the availability of electricity.

Nvidia's H100 GPU consumes around 700W when fully utilized, and thus 100,000 GPUs for AI and HPC workloads could consume a whopping 70 megawatts of power. Since these GPUs need servers and cooling to operate, it's safe to say that a datacenter with 100,000 Nvidia H100 processors will consume around 100 megawatts of power. That's comparable to the power consumption of a small city.

Musk stressed that while the compute GPU supply has been a significant obstacle so far, the supply of electricity will become increasingly critical in the next year or two. This dual constraint underscores the challenges of scaling AI technologies to meet growing computational demands.

Despite the challenges, advancements in compute and memory architectures will enable the training of increasingly massive large language models (LLMs) in the coming years. Nvidia revealed its Blackwell B200 at GTC 2024, a GPU architecture and platform that's designed to scale to LLMs with trillions of parameters. This will play a critical role in development of AGI.

In fact, Musk believes than an artificial intelligence smarter than the smartest human will emerge in the next year or two. "If you define AGI as smarter than the smartest human, I think it is probably next year, within two years," Musk said in an interview on X Spaces. That means it's apparently time to go watch Terminator again, and hope that our future AGI overlords will be nicer than Skynet. ☺

Intel CEO Pat Gelsinger has publicly invited Tesla CEO Elon Musk to tour his firm's semiconductor fab lines. In a post on the Twitter/X social media platform, Gelsinger said he was thinking of Musk when he was awarded the $8.5 billion in CHIPS Act funding, earlier in the week. Gelsinger has also been courting Musk's arch-foe OpenAI / Sam Altman. It is safe to say that the Intel CEO is trying to get an early start in filling the Intel Foundry Services (IFS) order books, now that the financial fuse has been lit.

In Gelsinger's post, embedded above, you can see the Intel CEO reach out personally to Elon Musk, promising him a tour of Intel's high-tech manufacturing lines. Musk didn't respond publicly, yet, but this may be because Gelsinger also asked him to "follow me," so the pair could chat privately via direct messaging (DM).

Musk would surely be an excellent catch for IFS. This superrich entrepreneur has fingers in many tech pies that are highly reliant on processors, lots of state-of-the-art processors. Musk's firms buy AI accelerators from both AMD and Nvidia, for tasks like machine learning, computer vision, self-driving, Grok, and more, but the firm is also developing its own Dojo ASICs (Application-Specific Integrated-Circuits) with new generations in development.

Looking back at other recent Tom's Hardware headlines, we can see Gelsinger has been very actively touting for business in recent weeks. We know the Intel CEO will have been talking to OpenAI's Sam Altman recently. Altman has floated the ambitious $7 trillion idea of OpenAI building its own fabs to make custom AI chips. However, Altman was at the last Intel Foundry event and surely will have mulled over the possibilities of using Intel's upcoming manufacturing capacity, and expertise. Also, last month, Gelsinger reiterated that Intel is willing to build chips for anyone, including long-time rival AMD.

On Wednesday, Intel's funding dreams came true as it came to a preliminary agreement with the U.S. Commerce Department. The iconic PC chipmaker will get $8.5 billion in direct funding for U.S. projects, plus $11 billion in low-interest loans and a 25% investment tax credit on up to $100 billion of investment. This is great news for Intel's domestic chipmaking plans covering projects in Arizona, New Mexico, Ohio, and Oregon. Perhaps delayed building projects such as the Ohio fab will move forward more swiftly, with fresh funding behind their sails.

In a wide-ranging interview with Wired, Nvidia CEO Jensen Huang revealed that he personally utilizes Perplexity AI as his go-to AI chatbot. Asked about his usage of tools like ChatGPT or Bard, Huang indicated his preference for this lesser-known chatbot. Huang uses both Perplexity AI and ChatGPT “almost every day,” according to the answer provided in his interview. We guess AI rivals like Bard / Gemini or Grok currently don't make the grade for a green team stamp of approval.

So, what does Huang use AI chatbots for? The Nvidia CEO explains that he uses the chatbots for research. Currently, Huang seems particularly interested in computer-aided drug discovery. We hope this is more of a scientific and business interest, and isn’t because a loved one has any worrisome health issues.

Perplexity’s appeal may be evident in its self-identification as the “world's first conversational answer engine.” Indeed, our own initial poking and prodding of the app and usage of Perplexity via the website reveals it is quick and easy to query and makes it natural to dig deeper into topics. Perplexity also offers a useful ‘library’ of past query threads, and a ‘discover’ news and current affairs feed to explore.

Above, you can see some Perplexity screenshots, which show the app’s simple UI, its excellent taste in PC tech sites, and insight into why the developers think their app is a superior choice.

Though Nvidia has a finger in many AI pies, we note that as recently as January 2024, it participated in a $73.6 million Series B funding round led by IVP. With this in mind, it shouldn’t be such a surprise that Jensen Huang has looked into Perplexity AI, and some execs are ‘dogfooding’ the product.

The Perplexity AI website and apps are free to use, but more features and abilities are unlockable with the single paid tier – the Pro subscription. Those who enjoy this chatbot might be tempted to pay the $20 monthly or $200 yearly fee for things like unlimited Microsoft Copilot queries, AI model selection (GPT-4, Claude 2.1, or Perplexity), unlimited file uploads, and $5 monthly generative AI credits.

Other interview nuggets

The topic of personal AI chatbot preferences was just a small part of the lengthy interview with Wired’s Lauren Goode. Naturally, there were a lot of AI topics covered in 2024. One of the more interesting nuggets was Huang’s description of a new kind of data center. The Nvidia CEO outlined an “AI factory,” which is claimed to have been in development for several years. It is likened to a power generator and is on the verge of productization at Nvidia.

Another topic that raised our interest was the discussion of Moore’s Law. Huang explained that Nvidia bought Mellanox to sidestep Moore’s Law at the data center scale.

We also heard about Huang's very frequent talks with top TSMC execs like Morris Chang. Hot topics include advanced packaging like CoWoS, capacity planning, and related new technology. However, facing more specific queries about the wait time for AI GPUs and the upcoming Blackwell generation, the Nvidia boss was uncharacteristically – but understandably – evasive.

Elon Musk has shared some previously unseen photos of when Nvidia CEO Jensen Huang called round to donate the world’s first DGX-1 server to budding startup OpenAI. The pictures were taken back in 2016, but this is the first we've seen of the two silicon mega celebs enjoying each other's company. Musk was one of the co-founders of OpenAI in 2015, but stepped away in 2018 over rumored conflicts of interest and the perceived change of direction at the artificial intelligence pioneering 'nonprofit.'

OpenAI is now a for-profit company and seems to have steered away from its early open-source leanings, which has caused some highly public friction between Musk and current OpenAI CEO, Sam Altman. With the burgeoning success of AI in recent years and OpenAI’s central role becoming obvious, Musk set up a new AI venture called xAI and announced the AI chatbot Grok last year.

Moving our focus back to the newly shared images, we know they come from August 2016, when Musk Tweeted “Would like to thank @Nvidia and Jensen for donating the first DGX-1 AI supercomputer to @OpenAI  in support of democratizing AI technology.” It looks like happier times, with all the jolly folk in the images clearly delighted with the donation, looking forward to the exciting future of accelerated AI. Note also that a single DGX-1 rack-mounted server does not constitute a "supercomputer" — not without dozens or hundreds more such units — but that's a separate topic.

Underlining his now older and pessimistic persona, Musk shared the happy images shortly after commenting “And now look what’s happened :( .” This is an obvious grumble about the path OpenAI has steered in recent years. Last February, Musk clarified his objections to the stewardship of OpenAI, complaining that “OpenAI was created as an open source (which is why I named it “Open” AI), non-profit company to serve as a counterweight to Google, but now it has become a closed source, maximum-profit company effectively controlled by Microsoft.” #SourGrapes

Nvidia’s DGX-1 ‘AI supercomputer in a box’ has been updated significantly since its launch in 2016, but when this machine was donated it was a $129,000 system that integrated eight Nvidia Tesla P100 (Pascal) GPUs, with the whole system capable of up to 170 teraflops (FP16) to tackle machine learning tasks. Nvidia subsequently rolled out the DGX-2 (Volta), the DGX A100 Server (Ampere), and its latest gen the DGX H100 Server (Hopper). Bigger systems like the DGX GH200 AI Supercomputer using 256 total Grace Hopper Superchips are the newest products in the family.

On Saturday, Elon Musk and the recently founded xAI company announced Grok. Today it went live to a select beta testing group in the US. Grok is a new large language model (LLM) AI chatbot being built to “assist in the pursuit of understanding,” and inject some humor into its responses. Grok will be available to all X Premium+ subscribers ($16pcm) when it exits beta, but there is no public timeline for this yet.

According to the xAI blog, Grok has been in development for four months, with just two months of training completed. The development team appear to be proud of Grok-1, the debut Grok LLM, and have produced a performance comparison table, which we have reproduced below.

From those figures, Grok-1 does indeed seem to be quite an achievement. A comment on the xAI blog frames the progress so far as being “exceptional,” as only models with significantly longer / larger training are better according to the four metrics.

While admittedly somewhat stunted for now, xAI and Musk claim that Grok has a “unique and fundamental advantage,” due to its knowledge gleaned from the Twitter/X platform.

Some may be wondering why we need another LLM, like Grok. A mission statement of sorts comes packed with highfalutin language about “building AI tools that maximally benefit all of humanity,” as well as being a powerful research assistant, and ideas generator.

Grok AI Is Primed for Controversy

Several statements within the launch materials appear to prepare readers for upcoming controversies. One of the mission-statement-style nuggets is that Grok will be “useful to people of all backgrounds and political views.” Moreover, it is stated that it will “answer spicy questions.” Last but not least, its pre-advertized humor, sarcasm, and beta work-in-progress status statements may have been carefully set up as a shield against criticism for the inevitable incoming blunders.

For a taste of some Grok interaction, Musk has shared several of the chatbot’s humorous responses to queries. One example can be seen above. Another example of Grok comes from a side-by-side comparison with another GPT, as embedded below.

There is no public timeline for when Grok will be out of beta but you can join a waitlist to hop on the first step of the journey with xAI. If it does indeed turn out to be a valuable tool, its $16pcm entry price slightly undercuts OpenAI’s ChatGPT at $20.

Elon Musk this week confirmed his intentions to build an AI project that will rival Microsoft-backed OpenAI's ChatGPT and Google's Bard. Musk's AI venture is to develop a 'truth-seeking' generative AI that promises to be safer than existing generative AI services.

"I am going to start something which I call 'TruthGPT', or a maximum truth-seeking AI that tries to understand the nature of the universe," said Elon Musk in an interview with Tucker Carlson at Fox News, reports Reuters. "It is simply starting late. But I will try to create a third option." He also is reported to have said that, "AI is more dangerous than, say, mismanaged aircraft design or production maintenance or bad car production."

While the entrepreneur did not reveal any other details about his AI project in the interview, Elon Musk shared his apprehension regarding the level of authority that Microsoft has over OpenAI, and the shift in the company's approach from being a nonprofit venture, reports Bloomberg. Musk also expressed concerns about chatbots being excessively politically correct and declared that his own AI would concentrate on discovering truth, but he refrained on disclosing how the AI seeks to find the truth.

Elon Musk's new AI venture is expected to be a separate separate entity from his other businesses, and its exact position within Musk's business empire remains uncertain. Yet, unofficial sources note that the company might use Twitter content for language model training and use Tesla's vast computing resources.

To build the AI startup, Musk is recruiting AI engineers from leading companies and about half a dozen of rockstar AI specialists are already onboard. Earlier it was reported that about 10,000 Nvidia compute GPUs (Nvidia A100 or H100, it's not clear yet) were procured for Musk's AI project needs, which means that hundreds of millions of dollars have already been spent on it. Meanwhile, Musk is also reportedly in discussions with SpaceX and Tesla investors regarding the funding of this AI project, according a Financial Times report.

Cyber security vendor Imperva revealed today that a Google Chrome bug exposed private data to innovative hackers. The security firm compared exploiting the flaw to playing 20 Questions with online services like Facebook, Google and "likely many other web platforms" to quietly gather info about people.

The problem stemmed from Chrome's rendering engine, Blink, and the way it handled the audio and video HTML tags. Hackers could inject hidden tags into websites and monitor the response Chrome received from a service like Facebook. The responses didn't necessarily reveal anything directly (none of them said, for example, "this person is a 23-year-old man from Oklahoma City"), but they could be used to gather that sort of data.

Imperva said it could measure the size of responses to determine if they were "yes" or "no" answers to a query. Combining those responses with Facebook's audience restriction tools, which let people show content only to specific demographics, would help the hackers learn more about their target. (Hence the comparison to 20 Questions--the exploit basically used the same "yes" or "no" approach to guessing someone's identity.)

Imperva explained in its blog post about the bug:

"For example, a bad actor can create sizeable Facebook posts for each possible age, using the Audience Restriction option, making Facebook reflect the user age through the response size. ... The same method can be used to extract the user gender, likes and many other user properties we were able to reflect through crafted posts or Facebook’s Graph Search endpoints."

But the method wasn't limited to Facebook; it could be used to gather data from other services as well. This could be particularly damaging if the hacker learned the target's email address, Imperva said, because it "would allow the bad actor to correlate the private data with the login email address for even more extensive and intrusive profiling." That profiling could enable more effective phishing attempts or other attacks.

Imperva said it told Google about this problem as soon as it confirmed the vulnerability and came up with a proof of concept. Google fixed the issue with the Blink engine in Chrome 68, which debuted in July, so if your browser's up to date, the flaw shouldn't affect you. If your browser isn't up to date, the bug's public revelation ought to be motivation enough to install the latest version of Chrome and defend against these attacks.