OpenAI, a leader in artificial intelligence research and development, has openly acknowledged significant challenges related to the supply of Graphics Processing Units (GPUs). While the company has indeed faced and continues to grapple with GPU shortages, impacting the rollout and accessibility of its advanced AI models, it is simultaneously implementing aggressive, multi-faceted strategies to secure and expand its computational infrastructure. This complex situation highlights the intense demand for specialized AI hardware and OpenAI's ambitious long-term vision.
Key Information
As of early 2025, OpenAI CEO Sam Altman explicitly stated that the company had "run out of GPUs," directly attributing this shortage to the staggered and slower-than-anticipated rollout of their highly anticipated GPT-4.5 model. The new model's computational demands were described as "giant" and "expensive," with its input and output token pricing being 30 and 15 times higher, respectively, than its predecessor, GPT-4o, underscoring its resource intensiveness.
The surge in demand for features like GPT-4o's image generation in June 2025 further strained OpenAI's infrastructure, leading to temporary rate limits and even the unusual step of borrowing computing power from its internal research division to alleviate overloaded GPUs. Reports from March 2025 indicated that OpenAI's GPU supply was "completely saturated," a unprecedented situation where the company could not readily sell access to its GPUs. In response to these immediate shortages, Altman pledged to integrate "tens of thousands" of additional GPUs in the weeks following the GPT-4.5 announcement to expand access for ChatGPT Plus subscribers.
Context and Background
OpenAI's journey with GPU reliance dates back to a foundational partnership with Nvidia (NASDAQ: NVDA) established in 2016. Nvidia has been a critical technology partner, notably in OpenAI's ambitious "The Stargate Project." Despite this deep and long-standing collaboration, OpenAI has reportedly expressed dissatisfaction with the "slow" flow of high-performance GPUs like the H100 and H200, indicating that even for a major customer, demand continues to outpace the supply from manufacturers.
The general landscape of AI development is characterized by an insatiable demand for high-end GPUs, which are essential for training and running large language models (LLMs). This demand has created a bottleneck in the industry, making access to sufficient computational power a strategic imperative for AI companies. OpenAI's predicament is a microcosm of this broader industry trend, amplified by the groundbreaking complexity and scale of its own models.
To mitigate immediate shortfalls and diversify its compute sources, OpenAI has actively pursued new agreements. In March 2025, a significant $11.9 billion agreement was struck with CoreWeave, an Nvidia-backed cloud service provider, granting OpenAI access to CoreWeave's AI infrastructure, which includes over a quarter million Nvidia GPUs. Further diversification efforts were seen in June 2025 when OpenAI began renting Google Cloud's (NASDAQ: GOOGL) Tensor Processing Units (TPUs) to support ChatGPT, marking its first substantial use of non-Nvidia AI chips and a strategic move beyond exclusive reliance on Microsoft Azure (NASDAQ: MSFT) and Nvidia GPUs. This expansion to include Oracle (NYSE: ORCL) and TPUs underscores a growing urgency to secure compute resources beyond existing limitations.
Implications
Short-Term Implications
In the short term, OpenAI's GPU constraints have led to practical limitations for users, such as staggered model rollouts, temporary service rate limits, and occasional performance issues during peak demand. The strategic partnerships with CoreWeave and the utilization of Google Cloud TPUs are critical stop-gap measures aimed at addressing immediate capacity needs and maintaining service availability. These moves demonstrate an agile response to an immediate crisis but also highlight the persistent challenges of scaling cutting-edge AI.
Long-Term Outlook
OpenAI's long-term strategy is exceptionally ambitious and multifaceted, aiming for significant compute independence and unprecedented scale. A cornerstone of this strategy is the development of custom AI chips to reduce dependency on external manufacturers like Nvidia. OpenAI began collaborating with Broadcom (NASDAQ: AVGO) in 2024 to design a custom AI chip for both training and inference, targeting mass production by TSMC (NYSE: TSM) in 2026 using a 3nm node. A dedicated team of approximately 20 engineers, including experts from Google's TPU project, is working on finalizing the design for their first custom chip by early 2025.
Beyond chip development, OpenAI is investing in massive infrastructure projects. "The Stargate Project," announced in January 2025 in partnership with Oracle, SoftBank, and MGX (with key technology partners Microsoft, Nvidia, and Arm), is a colossal joint venture aimed at building an AI infrastructure system in the United States, estimated to cost $500 billion over four years, with an initial $100 billion deployment. There have been reports of delays in this project, signaling the complexity of such undertakings. Additionally, "Stargate Norway," a European extension, is projected to house 100,000 Nvidia GPUs by the end of 2026.
OpenAI CEO Sam Altman has articulated an audacious long-term vision, expressing a willingness to invest "trillions of dollars" into AI infrastructure in the "not very distant future," exploring novel financing instruments to fund this monumental spending. He envisions a future where OpenAI could operate 100 million GPUs, a staggering 100 times their target of "well over 1 million GPUs" by the end of 2025.
Factors That Could Change the Situation
Several factors could significantly alter OpenAI's GPU supply situation. The successful and timely development and mass production of their custom AI chips will be crucial for long-term autonomy. The execution speed and scale of "The Stargate Project" and its European counterpart will determine their ability to build out proprietary data centers. Continued strong partnerships with existing and new cloud providers, as well as maintaining a robust relationship with Nvidia, will be vital. Finally, the broader global supply chain for high-end semiconductors and the pace of innovation in GPU technology itself will inevitably influence OpenAI's access to the necessary computational power.
Summary
While OpenAI has certainly experienced acute GPU shortages, impacting its ability to roll out models and services without constraint, the narrative is not simply one of "running out." Instead, it's a dynamic situation where intense demand for cutting-edge AI hardware meets aggressive, multi-billion-dollar strategic investments. OpenAI is simultaneously diversifying its cloud providers, entering massive infrastructure deals, and undertaking the formidable task of developing its own custom AI chips. These efforts underscore the company's commitment to overcoming current limitations and securing a long-term, scalable, and ultimately self-sufficient computational foundation for its ambitious AI development goals. The journey to trillions of dollars in AI infrastructure and millions of GPUs is a testament to the scale of OpenAI's vision and the challenges inherent in pushing the boundaries of artificial intelligence.
