Intel CEO Lip-Bu Tan said it plainly in February 2026: no relief on the AI capacity crunch until 2028. That is six more quarters of constrained GPUs, power, data centers, memory, and CPUs. Sam Altman said OpenAI was out of GPUs in February 2025. Oracle's Safra Catz said customers were being waved off or scheduled into the future in March 2025. Microsoft's Satya Nadella admitted in October 2025 that chips were sitting in inventory because there was no power to run them. Sundar Pichai called capacity the top question keeping Alphabet up at night in February 2026. This is not a single company's problem. It is a structural ceiling.
The downstream consequence is rationing. Inference prices, which have been largely subsidized and static, will rise. Enterprises will be forced to decide which teams get frontier models and which get something smaller or nothing at all. Not every CRM update requires a trillion-parameter model. Marketing gets this much compute, sales gets that much, engineering probably gets more. That prioritization framework does not exist at most companies yet, and building it will be painful.
The piece is worth reading in full because Tunguz does not stop at the problem. He traces the logical chain: constraint forces optimization, optimization accelerates open source adoption, and smaller models become the default for most workloads. The question he poses is simple and worth sitting with: what happens to your business when your AI does not answer?
[READ ORIGINAL →]