AI Token Demand Grows 250-Fold but 72% Flows to Cheapest Models, Goldman's Molavi Flags "Unplayable" Divergence

Goldman Sachs strategist Bobby Molavi has flagged a structural contradiction in AI markets: token consumption is surging, but most incremental volume flows to the lowest-priced models while GPU costs climb in tandem. He described the market as "unplayable." OpenRouter data show open-weight models handled 72% of token traffic in the second week of September but captured only 27% of platform spending. Closed models' spending share fell from 89% in early January to 73%. About 69% of tokens on the platform in mid-September originated from Chinese AI labs. Token processing on OpenRouter rose from 0.5 trillion weekly in mid-January 2025 to 128.4 trillion in the week of Sept. 14, 2026, roughly 250 times growth. GPU rental prices rose during the year: H200 up 160%, B200 up 120%, H100 up 35%, with older A100 chips now renting above early-year levels—challenging three-year depreciation assumptions. Epoch AI reported Sept. 22 that the cost of reaching equivalent AI performance has fallen about 47% per quarter since 2023, and 66% per quarter at the frontier. Molavi noted AI productivity gains remain unquantified outside the tech sector, with S&P 493 metrics roughly flat versus gains at the seven largest technology companies. Meta recently raised up to $95 billion from shareholders for AI capital spending, with its chip supplier taking equity—shifting vendor financing from debt to equity.