Meta Compute Just Rewrote the AI Capex Math
The biggest buyers of GPUs are becoming sellers of GPU time. That sound you hear is a thousand data center models being rebuilt from scratch.
On July 1, reports emerged that Meta plans to launch Meta Compute, a cloud unit selling surplus AI training and inference capacity plus access to its Llama models to enterprise customers. The market needed about a day to process what that means, then the chip complex started leaking.
Here is the simple version. The AI infrastructure trade priced compute as permanently scarce. Every hyperscaler capex announcement was read as demand. But if hyperscalers overbuild and then rent out the surplus, capex stops being pure demand and starts being future supply. Scarcity with a lag is not scarcity, it is a cycle.
This does not kill the AI trade. Compute demand is real and still growing. But it changes who captures the margin, and markets reprice who captures the margin much faster than they reprice whether the technology works. July's tape was that repricing happening in real time.
Watch the second-order effects: neocloud pricing, GPU depreciation schedules, and whether other hyperscalers follow. If selling surplus compute becomes standard practice, the moat was never the chips. It was the customers.