Vercel’s AI Gateway routed 17% of its token volume through DeepSeek in May, up from under 1% the previous month, while DeepSeek’s share of spend stayed near 1%. Anthropic’s spend share grew from 61% to 65% over the same period. Both facts are true simultaneously, and the implications split depending on what business you are in.
The AI Gateway production index, published by Vercel on June 9, tracks tens of trillions of tokens monthly between production applications and AI providers. It is not a vendor survey or analyst estimate. It is gateway traffic. That makes the signal unusually clean compared with most public data on AI workload distribution.
The price arithmetic is straightforward: DeepSeek’s per-token cost is low enough that 17 times the volume converts to roughly the same dollar amount as the next-cheapest option. Teams moving workloads to DeepSeek are not paying meaningfully more in aggregate. What Vercel’s framing adds is that this is not procurement experimentation. Engineers tested DeepSeek V4 output and found it good enough to ship for a broad class of tasks.
That distinction matters. Procurement experimentation looks like a handful of proof-of-concept projects. Engineering acceptance looks like 17 points of gateway share in a single month.
The other half of the data is equally significant. Anthropic holds 70 to 80% of spend across AI app generation, back-office agents, and coding agents on the same gateway. The complex, high-stakes, agentic workflows are still running on Claude. The workload mix that is shifting is not the hardest work. It is the bulk inference: summarization, classification, structured extraction, lightweight generation tasks where output quality thresholds are lower and latency matters more than capability ceiling.
This is the procurement thesis in empirical form. Over the past several editions of AI Insiders, the argument has been that cheaper models would absorb the commodity layer of AI workloads while frontier models retained the tasks where errors are expensive. Vercel’s May numbers are the first concrete production-scale confirmation that this separation is already happening in live applications, not just in architectural discussions.
The lab valuation question is the harder one to answer. At current trajectory, DeepSeek is taking volume share without taking revenue share. Anthropic’s spend growing to 65% of gateway dollars while losing token volume is a defensible position, not a deteriorating one. The per-token economics favor Anthropic on complex tasks, and those tasks are not moving.
What the trajectory warning looks like: if the workloads currently running on Anthropic begin to fragment as agentic frameworks get better at routing tasks by difficulty, the volume-to-spend ratio becomes an early indicator of which task categories are next to commoditize. The Vercel data does not show that happening yet. It shows the opposite: Anthropic’s revenue share expanding even as its token share declines. But the mechanism is now visible, and it will be worth tracking month over month.
For operators building on multiple providers today, the Vercel numbers give a concrete calibration point. The tasks where DeepSeek absorbed 17% of token volume are identifiable by category. Routing cheaper models to those tasks and reserving Anthropic for the agentic and high-stakes work is not a cost-cutting speculation anymore. It is what the Vercel production fleet is already doing at scale.
Any team still defaulting all inference to a single frontier provider without a tiered routing strategy is paying for capability headroom they are not using on a large portion of their requests.
Source: Vercel AI Gateway production index (vercel.com), published June 9, 2026.
