Vercel report shows that open-source models Token rose to 71.5%; Chinese models drive revenue growth for U.S. inference providers

AT A GLANCE

The report, citing Vercel data, stated that September open-source models processed 71.5%of the platform's Token, while estimated call fees accounted for 23%。

Article

Vercel: Open-source models consume 71.5% Token, Chinese models are creating substantial business for U.S. inference providers

According to Beating AI news, Vercel released its latest AI Gateway usage report.September, open-source models processed 71.5%of the platform's Token, but based on public prices, the estimated call fees accounted for only 23%。1 months ago, open-source models'Token share was only about 56%.

Among them, DeepSeek handled the most volume. Its models accounted for 58%of the platform's Token, but the estimated fees accounted for only 6.6%. Kimi, from Moonshot AI, was the opposite: it accounted for only DeepSeek-V4.1-Flash processed nearly half of the Token, while its fees accounted for only 2.6%. Kimi from Moonshot AI, by contrast, accounted for only 3.6%of Token, yet corresponded to 13.2%of the fees. Low-cost models handled a large number of requests, while higher-priced models such as Kimi K3 contributed more in spending.

More notably, Chinese models are creating substantial business for U.S. inference providers.September, nearly 70%of the estimated call fees for Chinese models corresponded to U.S. providers, while 2025 in December was less than 10%. DeepSeek itself handled more than of the models'Token, but calls handled by U.S. providers were billed at a higher average rate, with about of the fees corresponding to U.S. companies. These companies do not need to develop models themselves; they only need to deploy DeepSeek's open weights and provide computing power and API services to charge customers.

At the same time, the total call fees for September on the Vercel platform were about AI higher than in 2025 by December times, while the average 13 price fell by Token. Among continuously active, high-volume paying teams, nearly one-third switched their most frequently used model provider. Enterprises can use a unified 47%to call different models without integrating separately with each provider. Even when using the same model, they can choose different inference providers. Model developers no longer monopolize call revenue; companies providing inference services, as well as platforms such as Vercel that control developer access points, are also competing for this business.

Original link API https://m.theblockbeats.info/flash/372212

Key points

According to Beating AI news, Vercel released its latest AI Gateway usage report.
The report stated that September open-source models processed 71.5%of the platform's Token, while call fees estimated at public prices accounted for 23%；1 months ago Token their share was about 56%。
The report stated that DeepSeek's models accounted for 58%of the platform's Token, while estimated fees accounted for 6.6%。
The report stated that DeepSeek-V4.1-Flash handled nearly half of the Token, while fees accounted for 2.6%。
The report stated that Kimi accounted for 3.6%of Token, corresponding to 13.2%of the estimated fees.
The report stated that September Chinese models' estimated call fees corresponding to U.S. providers were nearly 70%, while 2025 in December was less than 10%。
The report stated that total September AI call fees on the Vercel platform increased by about 2025 compared with December in 13 times, while the average Token price fell 47%。
Among continuously active, high-volume paying teams, nearly one-third switched their most frequently used model provider.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Token This can be understood as the unit used to measure the text processed by a model. The report's point is that open-source models handled most of the processing on the platform, but based on public prices, their fee share was significantly lower than their usage share. Different models also have different usage and fee structures; for example, DeepSeek had a high usage share in the report, while Kimi's fee share exceeded its usage share. Enterprises can also switch models or inference providers through a unified API.

Why it matters to readers

If accurate, these data indicate that model call volume, model-development revenue, and inference-service revenue do not necessarily accrue entirely to the same company; providers that deploy open weights and offer computing power and API services, as well as platforms controlling unified access points, may also capture related business.

Beginners should focus on distinguishing concepts: who develops the model, who provides inference for calls, and which platform developers use to access it. The same model may be provided by different providers, and high usage does not necessarily mean a high fee share.

Risks and unknowns

The evidence is a Chinese media account of the Vercel report; no link to the original Vercel report was provided, so the specific data cannot yet be independently verified.
“September” was not explicitly marked with a year in the main text, so the comparison period with “2025 year December” cannot be fully confirmed.

Vercel report shows that open-source models Token rose to 71.5%; Chinese models drive revenue growth for U.S. inference providers

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center