Background: The Unstoppable Rise of AI Compute Demand

AI’s demand for compute has surged exponentially over the past decade.The size and complexity of AI models have grown exponentially, driving an insatiable demand for compute. For instance, OpenAI’s GPT-2, launched in 2019, contained around 1.5 billion parameters, while GPT-4, the latest iteration, reached an astonishing 1 trillion parameters . OpenAI reported that the compute used for AI training increased by over 300,000 times, with compute doubling approximately every 3.4 months.

The growth of the AI chip market underscores the immense demand for compute. Nvidia’s data center revenue grew by 126% in 2024, driven by the rapid expansion of AI workloads . Analysts predict that the AI accelerator market will exceed $150 billion in annual sales by 2025, and Nvidia alone is on track to earn $60.9 billion in revenue in 2024, up from just $17.5 billion in 2021. With the AI hardware market expected to reach $500 billion by 2028, the demand lear, and the race to scale AI infrastructure is intensifying.

Despite this clear growth in demand, some have questioned whether the skyrocketing investment in compute is sustainable. A notable example is Grok 3, Elon Musk’s AI model, which used 100,000 GPUs but reportedly showed only modest performance improvements . This has led some to argue that scais no longer a viable path to progress. However, this view misses the broader picture. Scaling up remains essential for achieving new breakthroughs in AI. Even with diminishing returns per unit of compute, larger models still push the boundaries of what AI can do, unlocking new capabilities and applications that smaller models cannot. Grok 3's limited improvement should be seen not as a sign of compute’s diminishing returns, but rather as evidence of the increasing complexity and optimization required at the cutting edge of AI development.

Another common misconception arises from the DeepSeek model, which has shown that it is possible to achieve significant performance improvements at a fraction of the compute cost. Some believe that this will lead to a future where AI needs less compute. However, as AI becomes more efficient, it often leads to broader adoption and the development of larger models, both increasing the total demand for compute. Efficiency improvements usually result in more AI applications across industries, leading to a rise in the need for computation. For example, DeepSeek's efficiency makes it easier and cheaper for more businesses to deploy AI, but that also means more training, more inference, and ultimately more compute consumption. One piece of evidence is that within a month after the open-sourcing of the DeepSeek-R1 671B model, the rental price of H200 increased by 10%.

The surge in demand for compute is also evident in the infrastructure constraints facing the industry. The supply of high-performance GPUs is struggling to keep up with demand. Waiting times for the latest GPUs can stretch up to 11 months, and even then, access is limited to a few large players with the financial resources to secure this critical hardware . Simultaneously, the AI-ready data center market to meet the growing needs of companies training and running AI at scale. To keep up, the industry would need to build twice as much data center capacity in the next few years as has been built since 2000. Yet, even with all this effort, the gap between supply and dues to grow.

The future of AI hinges on compute. From healthcare to finance to autonomous driving, every new AI application is driving the need for more compute. Whether through larger models or more efficient algorithms, the total demand for compute will only continue to rise as AI transforms more industries. Exabits is uniquely positioned to address this growing need by providing optimized GPU compute services that are scalable and accessible, helping to bridge the gap between AI’s exponential growth and the compute infrastructure needed to support it.