A venture investor argues that AI is advancing exponentially faster than companies can forecast, and that most workloads will soon run on cheap local models rather than expensive cloud services.

What happened: An analysis piece (Edition #44 of Implications) explores how even Anthropic, with better AI data than most, routinely underestimates results in compute requirements and revenue. The piece cites Coinbase CEO Brian Armstrong noting that 80 percent of workloads will run on 99 percent cheaper models within 12–18 months, with the other 20 percent on latest-generation models for high-stakes tasks. Stanford research shared by HuggingFace co-founder Clem Delangue shows local models can answer 71.3 percent of real-world chat and reasoning queries accurately, up from 23.2 percent in 2023, at a fraction of the cost and energy of cloud APIs.

Why it matters: Linear forecasting models are failing to predict AI's pace of change. Companies are discovering they've been overspecifying—sending expensive, high-capability AI to routine tasks the way a cardiac surgeon would take a patient's blood pressure. As cheaper local models improve, businesses face pressure to optimize AI spending and match task complexity to the right tool, or waste significant resources.

What to watch: Early signs of efficient-AI policies are emerging—Uber exhausted its token budget in four months and is now allocating monthly compute budgets to engineers. A routing layer (orchestration software that directs tasks to the most cost-effective model) will become mission-critical, with inference platforms like BaseTen, Modal, and OpenRouter competing to optimize these decisions.