The AI industry is about to hit an inflection point. According to Epoch AI, a nonprofit that tracks artificial intelligence trends, the compute power dedicated to running AI models will grow faster than the compute power used to build them by 2030.
The numbers behind the shift
Epoch AI’s projections paint a picture of an industry where the economics of deployment will increasingly dominate the economics of development. The organization estimates that nearly half of all inference compute will migrate to ASICs, or Application-Specific Integrated Circuits, by the end of the decade. These are chips designed to do one thing extremely well, as opposed to the general-purpose GPUs that currently power most AI workloads.
Meanwhile, the share of training compute in total AI operations is projected to hold steady at roughly 5%. Training compute for frontier AI models is currently growing at an annual rate of 4 to 5 times. The total installed AI compute base is expanding at a similar pace.
Historically, inference has already represented 60% to 80% of compute in actual deployments.
By 2030, frontier training runs are anticipated to reach approximately 2e29 FLOP. That progression is comparable to the leap from GPT-2 to GPT-4, supported by resources costing over $100 billion. Each of those frontier training runs may require between 4 and 16 gigawatts of compute power.
Epoch AI projects total AI power capacity in the US could surpass 50 GW by 2030, with global capacity exceeding 100 GW.
Why ASICs are eating inference
Google figured this out years ago with its Tensor Processing Units. Amazon has its Inferentia chips. The trend Epoch AI is projecting suggests this isn’t a niche strategy but the direction roughly half the inference market will take by decade’s end.
The constraints on this growth trajectory aren’t trivial. Power demand, chip production capacity, and data transfer limitations all pose real challenges. Epoch AI’s assessment is that these bottlenecks are manageable under current growth assumptions.
What this means for investors
If inference compute is where the growth is heading, the investment thesis for the semiconductor sector shifts meaningfully. Training still demands GPUs, and training budgets are still growing at 4 to 5x annually. But the higher-volume, recurring revenue opportunity increasingly lives in inference.
When you’re talking about 50 GW of AI power capacity in the US alone, that’s a massive buildout of data centers, power generation, and cooling systems.
The risk to watch is whether the 4 to 5x annual growth rate in compute is sustainable. Epoch AI’s projections assume current trajectories hold, but energy constraints, and geopolitical chip supply dynamics could all introduce friction.