The GPU shortage explained

A graphics processing unit, or GPU, is the engine of modern artificial intelligence. Training and running AI models is mostly parallel math, the same operation repeated across enormous amounts of data, and GPUs are built to do exactly that kind of work at scale. As AI moved from a research curiosity to a mainstream tool used by hundreds of millions of people, demand for that specific kind of hardware climbed faster than almost anyone planned for.

The trouble is that supply cannot follow at the same speed. Advanced GPUs depend on a small number of chip fabrication plants, specialized high-bandwidth memory, and advanced packaging steps that take years and billions of dollars to expand. You cannot add a factory the way you add a shift at a warehouse. When the whole industry wants the same hardware at the same moment, a shortage is the natural result.

It helps to think of the GPU shortage not as a single problem but as a gap between two curves. One curve is demand for AI compute, which bends upward sharply. The other is the supply of finished, deployable hardware, which rises in slow steps. The space between those curves is the shortage, and for now it keeps widening rather than closing.

Three forces push demand higher at the same time. First, models keep getting larger. Epoch AI finds that the compute used to train frontier AI models has grown roughly 4 to 5 times per year since 2010, a pace that compounds with startling speed. A model trained two years from now may need many times the hardware of one trained today.

Second, far more people use AI every day. That means more inference, the compute needed to actually run a model after it is built. Every search summary, coding assistant, and chat reply is an inference call, and the number of those calls keeps multiplying as AI gets woven into everyday tools. Third, each new hardware generation, such as the NVIDIA Blackwell platform, is designed for ever larger workloads, so the most capable buyers race to secure it the moment it appears.

The International Energy Agency reports that data centre electricity demand could more than double to around 945 TWh by 2030, with electricity use by accelerated AI servers growing about 30 percent per year. When the appetite for compute grows that fast, the hardware that provides it stays in short supply even as factories run flat out.

Even when chips are available, they need somewhere to run. That means data center space, high-density power delivery, and serious cooling. In the United States, Lawrence Berkeley National Laboratory estimates that data center electricity use rose to 176 TWh in 2023, up from about 58 TWh in 2014, and could reach 325 to 580 TWh by 2028 depending on how fast the buildout continues.

Building that capacity takes years, not weeks. New facilities need land in the right place, a grid connection that can deliver large and steady power, water or other cooling capacity, and skilled teams to operate everything around the clock. Any one of those can become the limiting factor, and increasingly the slowest link is power. A site can hold approved hardware and still wait on the grid.

This is why access to professionally hosted GPU capacity has become almost as valuable as the chips themselves. Owning a chip that has nowhere to run is not the same as owning working compute.

When a resource is scarce, it tends to flow to buyers who can commit early, purchase at scale, and operate the hardware properly. That is part of why large technology companies have moved aggressively to lock up supply, sign multi-year agreements, and build their own facilities. Priority access is itself an advantage, separate from price.

For everyone else, the practical question becomes how to get a real position in AI compute without trying to win a global supply race alone. The gap that matters most is between owning hardware and actually operating it, because a chip only becomes useful compute once it is powered, cooled, monitored, and connected to demand.

There is also a timing dimension. Because allocation rewards buyers who reserve capacity ahead of need, the people who act while the shortage is still widening tend to secure a position that latecomers cannot buy at any price for a stretch. Waiting is itself a decision, and in a tight market it carries a cost that is easy to underestimate.