There has been much anxiety in the last couple of years over the future availability of affordable computing power, the hardware that underpins the modern economy’s digital infrastructure. Such worries are understandable for anyone tracking the rapid advancements in generative artificial intelligence (GenAI). The time it takes to double compute demand to train such models is now faster than Moore’s Law—and poised to accelerate further, as the tech giants continue to bet on scale as the driver of progress in artificial intelligence (AI). The joint Microsoft and OpenAI plan for a $100 billion supercomputer is one recent indication of this trend. Spending on computing power by businesses has climbed alongside the growing compute intensity of AI: 2023 was the year when Google for the first time spent more on computing than people.

While demand for AI workloads continues to rise, the supply for specialized hardware, especially Graphics Processing Units (GPUs) or Tensor Processing Units (TPUs)—manufactured by a very small number of companies—appears unable to keep pace. A rapid increase in supply is unlikely for two reasons. On the one hand, this sector is hard to disrupt by outsiders, as there are significant barriers to entry due to intellectual property rights on the complex designs and architecture of GPUs and TPUs. Catching up to the incumbents’ IP would require substantial upfront R&D investment. On the other hand, an expansion of existing manufacturing capacity by incumbents is no easy feat given the complexity of the supply chain and sheer capital intensity of high-end chip production. The cost of setting up a chip production facility is estimated to have increased threefold in the last six years, from $7 billion in 2017 to $20 billion in 2023, mainly driven by the sophistication of the required machinery.

As hyperscalers such as Google or Microsoft build ever-larger GenAI models requiring more and more specialized hardware, won’t we face a severe scarcity of computing power? As we argue in this piece, however, once one clearly understands the different computational and hardware demands of model training and model inference, this scenario starts to seem less likely than might appear at first. In fact, even under bullish assumptions about the growth and intensity of GenAI demand in the coming years, it is far from obvious that we’re headed toward a structural scarcity of computing power. To test this claim, we built a quantitative model that is moderate in its supply estimates but bullish in its demand assumptions and found that GenAI workloads would account for only ~34% of global data center-based AI computing supply by 2028. The rise of GenAI, and the computational requirements associated with it, is unlikely to “break” the decades-long regime of affordable, widely available computing power.

Of course, there may be other factors that could severely constrain computing power supply—notably the energy required to power the data centers where much of the world’s AI workloads are and will continue to be processed. But the central contention of our analysis remains: The rapid adoption of GenAI will not by itself outpace the world’s capacity to produce the required computing hardware.