Sam Altman: OpenAI’s GPUs Are Depleted, GPT-4.5 Rollout Will Be Phased

OpenAI has recently launched its latest model, GPT-4.5. However, access is currently limited to Pro users paying $200 per month. On February 28, OpenAI CEO Sam Altman stated that the model’s release had to be delayed because “...(OpenAI) is growing so fast that we have run out of GPUs.” He then added, “We will be adding tens of thousands of GPUs next week and will roll it out to Plus users after that.” This means that soon, users will only need to pay $20 to access OpenAI’s most advanced model.

Rumors suggest that the ongoing GPU shortage may push OpenAI to collaborate with Broadcom in developing its own AI chips. However, since it would take years for the company to produce its own hardware, OpenAI must currently rely on NVIDIA and other suppliers to meet its growing demands.

This puts NVIDIA in a strong position. The chipmaker has stated that its latest Blackwell GPUs won’t be fully delivered until October this year. As institutions and individuals plan to expand data centers on a massive scale, NVIDIA is expected to remain dominant in the coming years. For instance, OpenAI and Microsoft are working on a $100 billion AI supercomputer, while Elon Musk aims to expand his Colossus supercomputer in Memphis, Tennessee, to over one million GPUs.

Other investors are also joining the data center race. A 3GW facility has been approved by the South Korean government, while another team is even exploring the idea of storing data on the moon. However, with AI infrastructure expanding at such an aggressive pace, Microsoft CEO Satya Nadella has warned that the industry may already be overbuilding AI systems—despite the increasing sophistication of AI models and their growing computational demands.

GPT-4.5 serves as a prime example. According to Sam Altman, “This is a huge and expensive model.” The cost for GPT-4.5 is $75 per million input tokens and $150 per million output tokens. In contrast, GPT-4o costs just $2.50 per million input tokens and $10 per million output tokens. Despite its steep price, Altman noted that it is “not a reasoning model, nor will it achieve overwhelming results in benchmarks.” However, he added, “It is a different kind of intelligence—it has a kind of magic I’ve never felt before.”