OpenAI President teases our 10-billion GPU future — says always-working AI future calls for 'every person to have their own dedicated GPU'

(Image credit: Getty / Bloomberg)

The various recent partnerships involving OpenAI and Nvidia have been discussed at length by the entire internet as well as the stock market. Besides enhancing shareholder value, though, it's hard to say what the endgame of these large corporations is. The answer is a bit clearer after a CNBC interview, where OpenAI's President Greg Brockman talked up a future with agentic AI that works in your sleep — pointing out we'll need 10 billion GPUs or the equivalent thereof for the future he envisions.

The interview in question includes OpenAI CEO Sam Altman, Nvidia CEO Jensen Huang, and Brockman. The discussion goes back and forth, but Altman offers some perspective on pointing out that the scale of its partnership with Nvidia is larger than even the Apollo program that put mankind on the moon. He thinks of the future AI proverbial collective as a "superbrain", embedded in our daily lives.

The most interesting remarks, though, come from Brockman, whose vision for the future is that of "an agent that's going to do work proactively for you while you're asleep", with every person on Earth having their dedicated GPU on the job. "You really want every person to have their own dedicated GPU," he told CNBC. That's definitely a hot take, one that might elicit some eye-rolls given how many folks are weary of AI features being shoehorned into every product.

Nvidia will invest $100 billion in OpenAI Jensen Huang Sam Altman Greg Brockman ~ 2025-SEP-22 Monday - YouTube

Watch On

However, keep in mind that in the early nineties, Microsoft's ex-CEO Bill Gates envisioned not just one computer in every home, but on every desk, a notion that seemed even more far-fetched and met with excitement and derision in equal measure. And yet here we are today, with one in every pocket.

Brockman continues on to state that the industry is still "three orders of magnitude off" where it needs to be in terms of AI computing power, and posits that to fulfill the always-on, always-working AI vision, about 10 billion GPUs would be necessary — a figure above the 8.2 billion people currently inhabiting the planet. He minces no words, saying that the world is heading to a state where "economy is powered by compute."

Reinforcing the notion, he went on to say he world is heading towards "compute scarcity," and in a direction where "the economy is powered by compute," implying that AI datacenter service could become de facto currency. While some could call that remark dystopian in nature, it's also hard to deny at this point in time, given that Nvidia's GPUs have become flashpoints in Sino-American diplomatic relationships — and that situation is bound to repeat itself in the future quite frequently.

Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.

TOPICS

Bruno Ferreira is a contributing writer for Tom's Hardware. He has decades of experience with PC hardware and assorted sundries, alongside a career as a developer. He's obsessed with detail and has a tendency to ramble on the topics he loves. When not doing that, he's usually playing games, or at live music shows and festivals.

6 Comments Comment from the forums

JRStern

And then, finally, at last, we will all be able to keep up with the Kardashians.
Reply
Stomx

I remember during Internet Bubble we've also heard that the heaven will come to Earth when we will be buying and selling while we sleep. When we woke up, all that ended with the crash and monopolization by Amazon

Sounds like a hysteria is unfolding. To feed these 10B GPUs needed total electric power of 5 countries like USA
Reply
Stomx

OpenAI, can you give us your promised heaven today or at least to show how it looks? Here at Tom's many already have that one GPU.

Citing: "The H100 PCIe 80GB has a higher peak half-precision (FP16) performance of 204.9 TFLOPS compared to the RTX 5090's 104.9 TFLOPS. In large language model (LLM) inference, a dual RTX 5090 setup has been shown to outperform a single H100 (50 million of which wants Elon Musk) in sustained output token generation, achieving over 80 tokens per second compared to the H100's 78 tokens per second, despite the H100's higher price

This performance advantage is attributed to the effective use of tensor parallelism in multi-GPU configurations, which is particularly beneficial for reasoning models requiring high sustained speeds.

When VRAM is not a constraint, the RTX 5090 performs nearly as fast as the H100, with only a few minutes difference in inference time for complex video generation tasks.

For example, generating a 960x960, 81-frame video with the Wan 14B model took 1 hour on the RTX 5090 and 1 hour on the H100, with the H100 being slightly faster. However, the RTX 5090 is significantly more cost-effective, with rental costs on platforms like Runpod being less than one-third of the H100's"

Also heard that 8bit FP8 in RTX5090 also there,it is just locked and can be unlocked by upgrading its BIOS
Reply
jlake3

"or the equivalent thereof" is doing some SERIOUSLY heavy lifting, because otherwise this feels like an admission that the vision they're targeting won't work in the near or medium term.

When talking about AI datacenters "A GPU" isn't an RTX 5050, "A GPU" is an H100 that sells for somewhere in the range of $25,000-$30,000. If you project that every person is going to need to have their own GPU and intend to make a profit, each customer needs to be paying you MORE than that over the useful service life of the equipment, which tax depreciation tables define as 5 years (if I remember correct). They're probably getting a bulk discount from Nvidia on hardware, but they also have to pay for the server it goes into, power, building construction, cooling, etc., so if we say that brings their overhead per customer back up to list price to keep the math simple, they'd need to charge $415-500 a month. And that's JUST to cover running all these buildings full of servers, that doesn't pay developer salaries or return a profit to shareholders.

Statistically most of the US doesn't have an extra $500/mo to throw at having their own dedicated AI agent, and even in other "high income" countries that's a pretty significant ask. Anything close to 100% adoption is going to require AI to basically be a free pack-in... and while they could subsidize the service with ads, people already don't like AI without ads in it, and getting $500/user/month worth of ads doesn't feel realistic (If people don't have that much disposable income, expensive ads will struggle to make a return).

If hardware improves by two orders of magnitude, then the numbers all start to get more reasonable... but with diminishing returns on process nodes, that's a big ask.
Reply
fiyz

And then deepseek appeared using clever design, showing OpenAI's approach of brute forcing raw numbers is a fools errand
Reply
trica

Go build.ypur AI data centers and the nuke plants to power them on the moon. All the free cooling they could want, and they can stop building them in my back yard and driving my electric bill up 15% annually like clockwork - even though the electric companies are still bringing in hundreds of millions in profit.
Reply

Show more comments