Vercel: Open-source models account for 71.5% of total tokens consumed, with Chinese models driving significant business for US-based inference service providers.

Beating AI Insight News Flash: Vercel releases its latest AI Gateway usage report. In September, open-source models processed 71.5% of the platform’s tokens, but their estimated call fees (based on public pricing) accounted for only 23%. A month earlier, open-source models’ token share stood at just ~56%. DeepSeek led in volume: its models took 58% of tokens but only 6.6% of estimated fees. DeepSeek-V4.1-Flash alone handled nearly half the tokens, contributing just 2.6% of costs. Conversely, MoonShot AI’s Kimi made up 3.6% of tokens but 13.2% of fees. Low-cost models handle most requests, while higher-unit-price models like Kimi K3 drive more spending. A notable trend: Chinese models are generating huge business for U.S. inference service providers. In September, nearly 70% of estimated call fees for Chinese models went to U.S. providers, up from less than 10% in December 2025. DeepSeek itself processed over 80% of its model’s tokens, but U.S. providers charged higher average rates—~40% of fees went to U.S. firms. These providers don’t need to develop their own models; they simply deploy DeepSeek’s open weights, offer compute power and API services, and charge customers. Meanwhile, total AI call fees on Vercel in September were ~13 times higher than in December 2025, while average per-token prices fell 47%. Among active high-volume paying teams, nearly a third switched their most-used model vendor. Enterprises can call different models via a unified API, eliminating the need to integrate with each vendor separately. Even for the same model, users can choose different inference providers. Model developers no longer monopolize call revenue: inference service providers and platforms like Vercel (which control developer access) are also competing for this business.