Ask what limits AI progress and the intuitive answer is ideas. The real answer, increasingly, is hardware. The story of modern AI is inseparable from the story of the chips it runs on.
Why compute became king
Training a frontier model means performing an astronomical number of calculations. The chips best suited to that work — GPUs, originally built for graphics — do the parallel math that neural networks demand. As models grew, so did their appetite for these chips, and demand outran supply. Compute stopped being a background detail and became the scarce resource that gates everything else.
Compute as strategy
When capability tracks access to chips, chips become power. Companies with large GPU fleets can train models others cannot; the same logic plays out between nations weighing the strategic importance of semiconductor supply. Access to compute now shapes the competitive map of AI as much as talent or data.
The scramble for custom silicon
Depending on a single supplier is uncomfortable, and general-purpose GPUs are not always the most efficient option. So the largest players are designing their own chips, tuned to their workloads, to cut costs and reduce dependence. Expect this trend to intensify.
Efficiency is the new frontier
The other response to scarcity is doing more with less. Smaller models, better training methods, and hardware tuned for efficiency are shifting the goal from “biggest model” to “most capability per chip and per watt.” In a compute-constrained world, efficiency is not a nice-to-have — it is the edge.
What it means
The AI race is, under the hood, a compute race. Watch the chips, the fabs, and the efficiency gains as closely as the model announcements — they determine who actually gets to compete.