Vercel CEO: Open Source Models Account for 62% of AI Gateway Token Traffic
Vercel CEO Guillermo Rauch announced that as of August 22, open source models accounted for 62% of the company's AI Gateway Token traffic, up from 28.4% on June 24, more than doubling in just over two months; closed source models dropped from 71.6% to 38% during the same period.
This data reflects the actual usage on the Vercel platform and does not represent the overall AI market. The platform's AI usage has surged, with total Token volume increasing by 59% month-over-month in July, and open source models now making up more than one-third of the total.
DeepSeek has surpassed Google to become the second-largest model vendor by Token consumption. The combined Token volume of OpenAI and Anthropic continues to grow at an accelerated pace.
The expenditure structure is markedly different. In July, open source models accounted for about 36% of Tokens but contributed only 8.6% of expenditures; Anthropic took 65.1% of expenditures with about 30% of Tokens.
In the future, closed source cutting-edge models may only account for 15% to 25% of Tokens but could capture 60% to 90% of economic value.
Market mechanisms are skewed towards low-cost open source usage, while funds still flow towards high-premium closed source models. Beneficiaries include cutting-edge closed source vendors like Anthropic and infrastructure providers, while mid-tier closed source models and high-cost inference scenarios face pressure.
Vercel emphasizes that enterprise adoption is still in its early stages, and toolchains need further adaptation to model-agnostic architectures.
Source: Public Information
ABAB AI Insight
Guillermo Rauch has long promoted an open-source ecosystem for front-end and edge computing, with Vercel AI Gateway serving as a window to observe real production usage. Open source has risen from 11% in April to 36% in July, reaching 62% in a single day, driven by the weight of Chinese open source models like DeepSeek, Kimi, and GLM, reflecting a rapid migration of cost-sensitive workflows.
In terms of capital flow, closed source vendors lock in high-value scenarios (coding agents, high-risk decisions) with high prices, while open source handles a massive volume of low-profit Tokens. In Vercel's data, Anthropic captures most expenditures at about 4.4 times the average price, forming a "barbell structure" of "volume-price separation": open source captures volume, closed source captures money.
A similar pattern was seen in the early days of cloud computing: open source Linux dominated server volume, while commercial distributions and managed services captured profits. Currently, AI is transitioning from a "model capability race" to a "layering of inference costs and scenarios," with open source accelerating infrastructure demand rather than replacing premium cutting-edge models.
Essentially, this represents a transfer of pricing power. The volume of Tokens spreading towards open source lowers the average unit price, but cutting-edge intelligence still holds scarce premiums. The mechanism lies in the layering of task value: commodifiable tasks flow to low-cost models, while high-risk, high-reward tasks continue to pay premiums, leading to a high concentration of economic value.
ABAB News · Cognitive Laws
Victory in volume does not equal victory in money
Open source captures traffic, closed source captures premiums
Cutting-edge models always account for a minority of Tokens but capture the majority of value