Rapid growth of OpenAI as Sam Altman predicts ownership of over a million GPUs by year's end - ChatGPT architect expands rapidly
In the world of artificial intelligence (AI), the race to scale up compute resources is intensifying rapidly. At the forefront of this competition is OpenAI, the non-profit AI research company, with its ambitious plans to significantly increase its AI training infrastructure.
OpenAI's primary cloud backbone, Microsoft's Azure, currently supports the company's operations. With a vision of reaching approximately 1 million GPUs, OpenAI is set to become the single largest consumer of AI compute on the planet, surpassing even Elon Musk's xAI's Grok 4 model by five times. This unprecedented scale reflects the intense demand to train increasingly large and complex AI models.
However, OpenAI's strategy doesn't stop at 1 million GPUs. Sam Altman, the company's CEO, has hinted at visions of a 100x increase, aiming for potentially 100 million GPUs. Such a scale would pose extraordinary challenges for data center design, energy use, supply chain, and chip production, but it underscores the expected exponential growth in compute needs for training state-of-the-art models.
GPUs remain the dominant compute resource for AI training, with Nvidia leading 93% of server GPU revenue in 2024. The AI data center market and associated semiconductor investments are rapidly growing, with total data center semiconductor spending hitting $209 billion in 2024 and expected to nearly double by 2030, driven by AI and high-performance computing (HPC) workloads.
In response to Nvidia's dominance, other players like AMD, Intel, Google, Amazon, and Microsoft are pushing domain-specific AI chips to reduce reliance on Nvidia-centric GPUs and optimize costs. This shift towards tailored AI ASICs is adding complexity to the landscape, as the industry moves away from a single-vendor dependency.
OpenAI's strategy reflects an ongoing "arms race" in scale, with the company rumoured to be exploring Google's TPU accelerators to diversify its compute stack. The company has also partnered with Oracle to build its own data centers, including the world's largest single facility in Texas, currently consuming around 300 MW of power and expected to hit 1 gigawatt by mid-2026.
Despite the large scale of GPU deployment, challenges remain. In February 2025, OpenAI had to slow the rollout of GPT-4.5 due to a lack of GPUs. The energy demands of OpenAI's Texas data center are also drawing scrutiny from Texas grid operators. At current market prices, 100 million GPUs would cost around $3 trillion, a significant investment for any company.
The production and powering of 100 million GPUs would require infrastructure beyond current capabilities. High-bandwidth memory (HBM) is being bet on by various companies to keep monster models fed. However, the industry is moving quickly, with 1 million GPUs now feeling like a stepping stone toward something much bigger.
In conclusion, the AI compute competition is intensifying radically, fuelled by Nvidia’s GPU dominance but increasingly contested by tailored AI ASICs. As global hyperscalers and companies like OpenAI seek to multiply GPU deployments from hundreds of thousands towards millions, and conceptually beyond, the landscape is evolving rapidly. The 1 million GPUs coming online by the end of 2025 mark a new baseline for AI infrastructure, one that seems to be diversifying by the day.
The race to increase AI training infrastructure is escalating, with OpenAI aiming to become the single largest consumer of AI compute by reaching 1 million GPUs. However, Sam Altman, OpenAI's CEO, has hinted at ambitions to further scale up to potentially 100 million GPUs, a feat that would present extraordinary challenges for data center design, energy use, and chip production.
The rapid growth in AI compute needs is driving the AI data center market and semiconductor investments, with total data center semiconductor spending expected to nearly double by 2030. In response to Nvidia's dominance, other players are pushing domain-specific AI chips to optimize costs and reduce reliance on Nvidia-centric GPUs, adding complexity to the landscape.
