The AI Compute Crisis: Why We Need Thousands of Times More Computing Power | Fractile Founder

Walter Goodwin argues that AI inference is heading into a massive compute crunch, not a slowdown. He says the industry will spend far more on running models in 10 years than today, because frontier intelligence is still not commoditizing and because test-time compute is becoming central to capability, especially for reasoning models and long-running agents.

The core technical thesis is that memory bandwidth, not raw compute, is now the main bottleneck. Goodwin says flop counts have scaled about a million-fold over 25 years, while bandwidth has scaled only about 40-fold, making today’s AI chips increasingly bandwidth-limited. Fractile’s answer is 3D memory stacking: putting memory on top of logic to combine SRAM-like speed with DRAM-like capacity, rather than forcing a trade-off between the two.

He also argues that the bottleneck is shifting from silicon to data centers, land, power, and grid connection. That makes better chips even more important, because the winner will be whoever can squeeze more intelligence per megawatt per rack. On competition, he thinks Nvidia remains extraordinary, but specialist chipmakers can win in specific workloads where performance and economics matter most.