Vercel: AI Gateway spending up 13-fold since December as customers rotate models

Vercel has published the latest edition of its AI Gateway Production Index, a report built on anonymised aggregate traffic across the company's model-routing service. The index tracks how businesses choose, switch and pay for AI models in production, and its newest figures show spending still climbing quickly even as the price of a unit of intelligence falls sharply.

Spending through Vercel's AI Gateway rose 13-fold since December, while the average cost per token dropped 47% over the same period. Vercel says the gateway routes tens of trillions of tokens a month between production applications and AI labs, which makes the index a view of real workloads rather than of benchmark scores.

The mix of models has shifted just as sharply. Open-weight models handled 71.5% of tokens processed in September, up from 56% in August, and accounted for 23% of spending, more than four times their share in June. Spending on frontier models still grew tenfold since December, but spending on open-weight models grew more than seventyfold, so absolute budgets are rising everywhere while the increment tilts toward cheaper systems.

Customers are not settling on a single supplier. Nearly one in three teams changed their most-used AI lab in September and almost one in two switched their most-used model, with both measures rising for a fifth consecutive month. Among paid teams that processed at least 10 million tokens between December and September, the average customer used 15 different models; that rose to 26 for teams above 100 million tokens and 42 for teams above 1 billion.

The index also traced how customers reach Chinese-developed models. Nearly 70% of September spending on Chinese-origin models went through US-based providers, up from less than 10% in December, meaning more than USD 2.50 moved through US intermediaries for every dollar spent through China-based ones. DeepSeek's models processed 58% of all gateway tokens but represented 6.6% of spending; more than four in five of those tokens were served by DeepSeek itself, yet two in five dollars spent on its models went through US-based providers.

For suppliers, the implication is that revenue growth and customer dependence are coming apart. Businesses are spreading workloads across many models and many intermediaries, so a lab can win more calls without locking in budget share, and rankings can be rewritten within months by a cheaper substitute.

What to watch next is whether the two curves keep diverging: total spending, and cost per token. If open-weight models push their spending share above the current 23%, pricing room for frontier labs narrows further. If teams keep rotating their primary model every month, competition in AI infrastructure shifts from the strongest model to the most flexible routing and cost control. Vercel's next index will provide the first comparable evidence.