Chinese open-weight AI models are capturing enterprise workloads, and U.S. compliance teams need a plan

Chinese open-weight AI models processed 29% of all tokens routed through Vercel's production AI gateway in June 2026, up from roughly one-ninth of total volume in April, according to the company's AI Gateway Production Index published this month. Despite that near-tripling of share, those models accounted for less than 4% of total spending on the platform, a gap that is shaping how enterprise teams think about model routing.

The index tracks tens of trillions of tokens monthly between production applications and model providers. Vercel found open-weight models were priced at roughly one-tenth the platform's average token rate, according to Computing's reporting on the index. The implication for operations and IT leaders is direct: a growing tier of production workloads can be handled at a fraction of frontier-model cost, without touching the tasks where accuracy carries the highest business risk.

DeepSeek closes in on the top two

DeepSeek was the clearest beneficiary of the shift, capturing 22.6% of token volume in June to rank third on the gateway, behind only Anthropic and Google. Google's share declined to 24% after a surge in April, leaving DeepSeek within two percentage points of second place, according to Computing.

Anthropic held the dominant revenue position despite that volume spread. The company captured 61% of total spending while processing 32% of tokens, reflecting its concentration in higher-stakes workloads: coding assistants, back-office automation, and application generation, where enterprise teams consistently prioritize accuracy over cost, according to Vercel's data.

Overall AI investment on the platform continued to climb. Token volume rose 29% in June while spending grew 27%, according to Computing. Average cost per token held roughly flat, as the downward pressure from cheaper open-weight adoption offset a roughly 12% price increase among leading closed-weight frontier models.

Specialization by modality: no single leader across the board

Text, image, and video generation are consolidating around different leaders. Anthropic continued to lead text generation. OpenAI held the largest share of image production, generating more than half of all images through the gateway. Google's Nano Banana model followed with nearly 39% of image output, according to Computing.

Video told a different story. ByteDance's Seedance captured nearly half of video-related spending despite generating around one-third of total video volume, a premium-cost position that reflects its hold on quality-sensitive video tasks. The data suggests enterprise buyers are already running a multi-model stack, selecting providers by modality and cost tolerance rather than consolidating on a single vendor.

Vercel's head of agentic infrastructure, Harpreet Arora, described the dynamic plainly in Computing's coverage: when a task does not require the best model, teams are increasingly routing it to the cheapest one that meets the bar, and recent Chinese models are winning that price competition. Arora also flagged that privacy protections and data residency remain critical considerations for any enterprise evaluating open-weight models for production deployment.

Congressional scrutiny adds a compliance dimension

The operational calculus extends beyond cost and performance. The House Committee on Homeland Security and the House Select Committee on China announced in April they would jointly investigate the growing enterprise adoption of Chinese-developed AI models, according to CNBC. Initial letters went to companies including Cursor and Airbnb over their exposure to China-developed AI.

Cursor built its Composer 2 model using Kimi, developed by Chinese company Moonshot AI. The coding tool is set to be acquired by SpaceX for $60 billion, CNBC reported. Airbnb told CNBC its AI activity runs overwhelmingly on U.S.-origin models, and that any China-origin model use is routed exclusively through approved U.S.-based service providers to keep data and operations separate.

Other senior executives have been more open about the cost rationale. Coinbase CEO Brian Armstrong and AI startup Lindy's CEO Flo Crivello have both publicly cited Chinese models as a way to reduce AI operating costs, according to CNBC. Their comments reflect a broader tension enterprise leaders are navigating: real cost savings on one side, and escalating regulatory and geopolitical risk on the other.

A State Department spokesperson told CNBC the growing use of Chinese AI models raises serious concerns, pointing to questions about content design and ideological alignment. The Chinese government, for its part, is reportedly considering restrictions on overseas access to its leading models, according to a Reuters report cited by CNBC. Meanwhile, the four largest U.S. frontier AI companies still accounted for 95% of total gateway spending in June, per Vercel's data, meaning revenue concentration has not yet shifted even as volume share has.

What this means for your team

• Audit your model routing layer now. If your platform or vendor automatically routes workloads to the lowest-cost model, confirm which providers are in that pool and whether any are subject to the House committee investigation or existing government bans.
• Define a data residency policy for open-weight inference. Vercel's own data flags this as a top enterprise concern. Requiring that all inference, including open-weight runs, stays on approved U.S.-based infrastructure is the most defensible near-term posture.
• Separate workload tiers in your AI spend model. Vercel's data shows back-office agents consume 14% of spending despite only 5% of token volume. Document which tasks carry accuracy and risk requirements that justify frontier-model pricing, and which do not.
• Monitor the joint House committee investigation for procurement implications. Letters to named companies are an early signal; potential restrictions on Chinese AI model use by U.S. enterprises could affect vendor contracts and model availability on timelines shorter than typical procurement cycles.