Vercel's AI Gateway data: DeepSeek overtakes Google in token volume as gateway spend jumps 74%

Vercel's AI Gateway — which routes tokens between production applications and AI labs at massive scale — shows DeepSeek has overtaken Google to become the second-largest lab by token volume, while overall gateway spend and pricing swung sharply in July.

What's new

In an August 11, 2026 post, Vercel said "every month, AI Gateway routes tens of trillions of tokens between production applications and AI labs," giving it a real-time view into production AI usage patterns. The headline finding: "DeepSeek became the second-largest lab by token volume, now running more than twice Google's volume." On pricing, Vercel reported "the average price paid per token fell 13.6% in July, after rising almost 20% in May and holding steady in June" — a reversal that follows two more volatile months.

The data also showed sharp concentration at the high end of the market: "Anthropic collected 65% of gateway spending on 30% of token volume, at 4.4 times the average price of every other lab's tokens," indicating Anthropic's models command a substantial premium even as they account for a minority of raw token throughput. Model churn was also high: "81% of July's tokens ran on models that were not on the gateway six months ago," and overall gateway usage is accelerating — "Gateway spend is up 74% since May and volume has doubled, with July's token growth running at nearly double June's pace."

Context

As a routing layer sitting between a large base of production applications and dozens of model providers, Vercel's AI Gateway index functions as a real-world usage proxy that's harder to game than self-reported labs' claims or synthetic benchmarks — it reflects what developers are actually shipping and paying for, not what labs say their models can do. DeepSeek's climb to second place by volume tracks with a broader pattern this year of Chinese open-weight labs (DeepSeek, Qwen, Z.ai/GLM) taking meaningful production share on cost and throughput even as US labs retain pricing power at the premium end, as reflected in Anthropic's outsized share of spend relative to volume.

Why it matters

The 81% figure — four out of five tokens now running on models that didn't exist on the gateway six months prior — is a concrete illustration of just how fast the frontier model landscape is turning over, and it argues against any single-model or single-lab integration strategy for production AI applications. The volume/spend split between DeepSeek and Anthropic also crystallizes the market's current bifurcation: a high-volume, cost-driven tier increasingly served by Chinese open-weight models, and a premium tier where labs like Anthropic sustain far higher per-token pricing despite lower relative volume. For infrastructure buyers, that split is becoming the central strategic question — optimize for cost-per-token at scale, or pay a premium for capability where it matters.