Chinese open-weight AI models grab 29% of Vercel traffic as US costs bite
Synopsis
Chinese open-weight AI models now command 29% of token volume on Vercel's AI Gateway — nearly triple their April share — with DeepSeek's V4 Flash alone controlling 20% of traffic and Zhipu's GLM-5.2 surging 50-fold since mid-June, signalling a structural cost-driven exodus from US premium AI.
Key Takeaways
Open-weight models reached 29 per cent of token volume on Vercel 's AI Gateway , nearly tripling their share since April 2026 , according to the company.
DeepSeek 's V4 Flash became the single largest model by volume on the gateway, accounting for more than 20 per cent of Vercel 's total traffic as of Wednesday, 16 July 2026 , up from about 15 per cent a month prior.
Zhipu 's GLM-5.2 saw its daily token volume surge 50-fold on Vercel since mid-June , running at approximately one-fifth the cost of Anthropic 's Claude Opus 4.8 .
The shift is driven by near-frontier performance parity between Chinese open-weight models and premium closed-source US alternatives from OpenAI and Anthropic .
Open-weight models allow enterprises to download code for free and run inference on local hardware, eliminating per-token subscription fees charged by proprietary providers.
Chinese open-weight AI models are rapidly displacing premium US counterparts on enterprise platforms, with Vercel — the San Francisco-based cloud platform for AI web development — reporting on Wednesday, 16 July 2026 that cost-driven migration is accelerating sharply across its global developer base. Open-weight models now account for 29 per cent of token volume on Vercel's AI Gateway platform, nearly tripling their share since April, according to the company.
The numbers driving the shift
Since mid-June, the daily token volume of Zhipu's GLM-5.2 — which operates at roughly one-fifth the cost of Anthropic's Claude Opus 4.8 — surged 50-fold on Vercel's platform, according to the company. Separately, DeepSeek's V4 Flash, a streamlined version of the firm's flagship V4 Pro, emerged as the single largest model by volume on the gateway, capturing more than 20 per cent of platform traffic on Wednesday, up from approximately 15 per cent a month ago.
Why it matters
The economics are stark: proprietary closed-source models from OpenAI and Anthropic require premium cloud subscriptions and charge per token — the basic unit of data an AI model processes. Open-weight models, by contrast, allow businesses to download code at no cost and run it on local or third-party hardware, dramatically compressing inference bills. Until recently, enterprises tolerated the premium because open alternatives lagged too far behind frontier performance — a gap that is narrowing fast.
The competitive backdrop
DeepSeek and Zhipu are among a cohort of Chinese AI developers whose models have achieved near-frontier benchmarks at a fraction of the training and inference cost of their US rivals. Their open-weight release strategy — publishing model weights publicly — has allowed global developers to integrate them without vendor lock-in or recurring subscription fees, a structural advantage that closed-source providers cannot easily replicate.
Who is most exposed
OpenAI and Anthropic face the clearest near-term revenue pressure as developer-facing token consumption shifts toward free-to-download alternatives. Cloud hyperscalers that host and bill for proprietary model inference are also exposed if enterprise workloads migrate to self-hosted open-weight deployments. The trend also raises questions for US policymakers weighing export controls on AI model weights.
What's next
The pace of adoption on Vercel's gateway suggests the migration from closed to open-weight models is no longer a niche developer experiment — it is becoming mainstream enterprise behaviour. Analysts will be watching whether OpenAI and Anthropic respond with aggressive price cuts, and whether US regulators move to restrict open-weight model distribution from China as a national-security measure.
Point of View
Not a lagging one — developer platforms are where enterprise AI spending patterns form before they show up in quarterly earnings. The 50-fold surge in Zhipu GLM-5.2 usage and DeepSeek V4 Flash's 20 per cent traffic share suggest that cost, not capability nationalism, is the primary enterprise decision variable in 2026. What mainstream coverage underplays is the structural asymmetry: US closed-source providers monetise through recurring token fees, while open-weight Chinese models erode that model by design. The real policy question — whether Washington moves to restrict open-weight model exports from China the way it restricted chip exports — is now squarely on the table, and the answer will shape the global AI market more than any single product launch.
NationPress
16 Jul 2026
Frequently Asked Questions
What share of Vercel AI Gateway traffic do Chinese open-weight models now hold?
Chinese and other open-weight models account for 29 per cent of token volume on Vercel 's AI Gateway as of July 2026 , nearly tripling their share since April , according to the company. DeepSeek 's V4 Flash alone controls more than 20 per cent of the platform's traffic.
Why are businesses switching from OpenAI and Anthropic to Chinese AI models?
The primary driver is cost: open-weight models can be downloaded for free and run on local hardware, avoiding per-token subscription fees charged by OpenAI and Anthropic . Zhipu 's GLM-5.2 , for instance, operates at roughly one-fifth the cost of Anthropic 's Claude Opus 4.8 , while delivering near-frontier performance.
What is DeepSeek V4 Flash and why is it significant?
DeepSeek V4 Flash is a streamlined, cost-optimised version of DeepSeek 's flagship V4 Pro model. It became the single largest model by volume on Vercel 's AI Gateway , capturing more than 20 per cent of platform traffic as of Wednesday, 16 July 2026 , up from about 15 per cent a month earlier.
Which companies are most at risk from the rise of open-weight AI models?
OpenAI and Anthropic face the most direct revenue exposure as enterprises redirect token consumption toward free open-weight alternatives. Cloud providers that bill for proprietary model inference are also at risk if workloads shift to self-hosted deployments.
Could US regulators restrict access to Chinese open-weight AI models?
It is a live policy question. The rapid adoption of models like DeepSeek V4 Flash and Zhipu GLM-5.2 by global enterprises has intensified debate in Washington over whether open-weight model weights from China should face export-control-style restrictions, similar to measures already applied to advanced semiconductors.
