OpenAI has run out of GPUs, says Sam Altman — GPT-4.5 rollout delayed due to lack of processing power
Tens of thousands of GPUs will arrive next week.

OpenAI has just released its latest model, GPT-4.5. However, it’s currently limited to Pro subscribers who pay $200 a month. Its CEO, Sam Altman, said on X (formerly Twitter) that it had to stagger the model’s release because “…we’ve been growing a lot and are out of GPUs.” He then added, “We will add tens of thousands of GPUs next week and roll it out to the Plus tier then.” So, even if you’re only paying $20 a month to OpenAI, you won’t have to wait long to get access to its most advanced model.
GPT-4.5 is ready!good news: it is the first model that feels like talking to a thoughtful person to me. i have had several moments where i've sat back in my chair and been astonished at getting actually good advice from an AI.bad news: it is a giant, expensive model. we…February 27, 2025
Altman added in his post that hundreds of thousands more GPUs are coming soon. Shortages like these are what’s pushing OpenAI to develop its own AI silicon in partnership with Broadcom. But because it will take the company years before it can release its own chips, for now it must rely on Nvidia and other providers for its needs for now.
This shows how Nvidia is in a good position, with the chipmaker saying that its latest Blackwell GPUs are sold out until October this year. And with institutions and individuals planning massive data center expansions, Team Green will likely be on a roll for the next couple of years. For example, OpenAI and Microsoft are working on a massive AI supercomputer that would cost $100 billion, while Elon Musk wants to scale his Colossus supercomputer in Memphis, Tennessee, to over a million GPUs.
Other investors are also getting on the data center game, with one 3-GW facility getting the go-ahead from the South Korean government, and another team experimenting with storing data on the moon. However, all this expansion of AI infrastructure has got Microsoft CEO Satya Nadella saying that there will be an overbuilding of AI systems, even as AI models become more advanced and require more computing power.
OpenAI’s GPT-4.5 is a good example of this happening. Sam Altman says that “it is a giant, expensive model,” with GPT-4.5 costing $75 per million input tokens and $150 per million output tokens. By comparison, GPT-4o only costs $2.50 per million input tokens and $10 per million output tokens. Despite its outrageous pricing, Altman says that it “isn’t a reasoning model and won’t crush benchmarks.” Still, he adds that “it’s a different kind of intelligence and there’s a magic to it I haven’t felt before.”
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Jowi Morales is a tech enthusiast with years of experience working in the industry. He’s been writing with several tech publications since 2021, where he’s been interested in tech hardware and consumer electronics.
-
DS426 Does AMD have shortages on MI300X and MI325X? Anyone whining about AI GPU shortages needs to realize not everyone can go to a single vendor and expect there's plenty to go around. Partnering with Broadcom will eventually help, but there's only so much leading-edge node capacity available, so some compromises have to be made.Reply
And boo hoo for Sam Altman. It'll arrive when it arrives. -
sseemaku But Intel did inventory write off of gaudi and AMD isn't selling as many MI300s as expected. Seems like there is no room for more than one company in the AI accelerators space.Reply -
bit_user
For quite a while, the limiting factor has been HBM & advanced packaging capacity for it, not fab capacity for the compute dies. All HBM production for 2025 has already been sold.DS426 said:Does AMD have shortages on MI300X and MI325X? Anyone whining about AI GPU shortages needs to realize not everyone can go to a single vendor and expect there's plenty to go around. Partnering with Broadcom will eventually help, but there's only so much leading-edge node capacity available, so some compromises have to be made.
Also, I might have this wrong, but I think I read that Micron diverted some of its GDDR7 (?) production to make HBM, instead. If so, perhaps it could have something to do with the RTX 5000 scarcity problems, at launch. -
bit_user
OpenAI is still losing money. So, any delays in placing their orders might've come from the need to get more financing.phead128 said:All those Nvidia chips got smuggled to China, so no more for Mr. Altman. -
JRStern I'm boggled by these numbers.Reply
I think very highly of 4o, when properly used, yet here he is bottlenecked on a version 4.5 that he admits is virtually the same just 10x-15x more expensive.
This must be the $80b that Nadella says Microsoft is still spending this year, and it seems a waste.
Within ten years they'll have algorithms that can run this stuff 100x more efficiently on even today's hardware.
Also, do we know if Altman is using only new B200s or only old H100s or what, for inference? -
bit_user
Naw, more efficiency is certainly possible, but you're going to need different hardware. Spiking neural nets are a good example. They're a lot more efficient, but don't run well on GPUs.JRStern said:Within ten years they'll have algorithms that can run this stuff 100x more efficiently on even today's hardware.