Open-weight AI models' token-to-spend split

Gateway data from multiple platforms shows open-weight models handling the majority of AI token volume while representing a small share of actual spending. On Requesty's platform for the week of September 21, open-weight models served 54.6% of tokens and 34.3% of requests but only 14.9% of spend across the 40 largest models. Vercel's AI Gateway Production Index showed 78.4% open-weight token share on September 19, a figure that rose from 11% in April to 29% in June before continuing upward, while open-weight models stayed under 4% of spend. OpenRouter showed open-weight models at roughly 78% of token volume.

Requesty argued that token share is the most flattering metric for cheap models because "the cheapest models attract the longest, most repetitive agent loops." Cache dynamics reinforce the divergence: Flash-tier open-weight models on Requesty ran 81-96% cache hit rates against a platform-wide rate of 78.3%. Artificial Analysis found a 76x cost spread across models in its September 3, 2026 snapshot, with rankings that did not match per-token price lists.

Within open-weight volume, Chinese models hold a disproportionate and growing share. On OpenRouter, Chinese models processed 36.39 trillion tokens in the week ending July 19, up from 4.37 trillion in late April, compared to 7.39 trillion for top US models in the same week. Vercel's data marked the week of February 9-15, 2026 as the first time Chinese systems processed more tokens than American ones.

One source characterized the open-weight volume shift as "a HUGE problem for both companies" referring to OpenAI and Anthropic, framing it as direct pressure on their pricing, margins, and revenue growth projections, while also calling it bullish for AI infrastructure because "even if OpenAI and Anthropic capture a smaller share of AI spending, the demand for compute isn't going anywhere." Jefferies strategist Chris Wood stated his base case is "massive capital destruction in the US" from the AI investment cycle, with the four largest US hyperscalers projected to spend approximately $695 billion on capex in 2026 and approximately $870 billion in 2027, and said his trigger for an investment unwind would be credit markets cutting off funding.