Go to Groundby Jaidev Karthickeyan
← All notes Markets · November 2026

QuarryAre open-weight models taking the money, or only the traffic?

Open-weight models took the volume first. The money is following.

On Vercel's AI Gateway, open-weight models went from 29% of tokens and under 4% of spend in June to 71.5% of tokens and 23% of spend by September.

What happened

In June, open-weight models did nearly a third of the work on Vercel's AI gateway and earned less than four cents of every dollar spent there. By September they did seven-tenths of the work, and the money had started to follow. The question I care about is who it follows to.

Vercel's monthly index tracks real production traffic, not benchmarks or launch-day excitement. In April, open-weight models handled 11% of tokens. By June, 29%, on less than 4% of spend. In August they passed the majority mark, with 56% of tokens and 14% of spend. In September: 71.5% and 23%.

So open-weight models have stopped being the cheap thing you try on a Friday afternoon. They are now a large share of production AI. But the volume is the less interesting half of the story. The money is the more interesting half.

Between June and September, their share of tokens rose about two-and-a-half-fold; Vercel puts their September share of spend at more than four times June's. That doesn't mean every open-weight token got more valuable. It means the mix changed.

Going to ground

First order

Volume moves first. Nobody needs a frontier model to summarise a memo, sort a support ticket or pull a date out of an invoice. If a cheaper model is accurate, fast and reliable enough, paying a premium is hard to defend in a budget meeting.

That is plain substitution: matching the model to the job. Demirer, Fradkin and Tadelis document a fast-growing crowd of models, creators and providers, and estimate that open models cost about 90% less than closed models of comparable capability, with persistent price gaps even within the same capability tier. Intelligence is turning into a competitive input, and buyers have noticed.

Cheaper intelligence doesn't have to mean less spent on AI. Lower prices can make new uses worth doing. How much new demand appears is still open: the same authors' preliminary short-run price elasticity is just above one, which doesn't leave much room for a spending boom in the short run.

Second order

The money follows the expensive end. Spend is rising, but not because cheap models took on harder work. Vercel attributes about two-thirds of the rise in open-weight spend to models priced at four or more times the open-weight average, led by Kimi K3. At the other extreme, DeepSeek-V4.1-Flash processed nearly half of September's tokens for 2.6% of spend.

So the category now runs on two engines: cheap models doing enormous volume, and pricier ones taking a disproportionate share of the money. Open weights no longer compete on price alone, and customers are mixing models by task instead of marrying one provider. The model is now a part you choose, not a platform you join.

Third order

Value moves to whoever controls what surrounds the model. David Teece explained this in 1986. When an innovation is easy to imitate, the returns go to whoever owns the complementary assets needed to sell it. EMI invented the CT scanner; GE, with the hospital sales and service network, took the market.

Applied to AI, the question is no longer just who builds the best model, but who controls what makes it useful in production. Compute and hosting matter, though capacity that is plentiful and sold on price rarely earns fat margins.

Next up is routing: deciding which model handles each request, when to escalate, and how to trade cost against speed, accuracy and reliability. That is different from agent orchestration, which coordinates the steps of a task. Moslem and Kelleher's survey finds that well-designed routing and cascading systems can beat the strongest single model on cost and quality together.

Above that sit proprietary data, workflow integration, security, governance and the customer relationship. The model provides the capability. The company that controls the workflow may own the customer, the data and the economics of the sale.

The frontier premium

None of this means frontier models are on their way out. Closed models still take most of the money: in June, Anthropic took 61% of gateway spend on 32% of tokens, and 72% or more of spend in every high-stakes use case Vercel reported. In September, closed models still took about three-quarters of spend.

There are good reasons. A wrong answer costs a lot more in software engineering, financial analysis, legal work or a long agentic task, and better reasoning, reliability and tool use can justify a real premium. But the premium has to be earned again and again. As open models improve, a gap in capability has to show up as a gap in results.

That pushes frontier labs to sell more than model access: enterprise contracts, agents, developer ecosystems and services. It also hands bargaining power to customers who can switch. If the model can be swapped without rebuilding the application, the customer has the stronger hand.

The provenance question

Some problems price and capability can't solve. In regulated industries, provenance, licensing, data handling and jurisdiction matter. Before deploying an open-weight model, an organisation needs to know where it came from, what it learned from, what strings come attached, where it runs and who answers for it.

Open weights give you the parameters, not the full story. They are not open-source AI in the complete sense: they don't disclose the training data, remove licence restrictions or certify a model for regulated work. Chae and coauthors, reviewing 500 Hugging Face model cards, argue that the cards alone aren't enough for downstream governance, and that standard open-source licences can weaken the enforceability of usage restrictions.

The question is sharpest for Chinese-origin models. The economics can be compelling, but some organisations will face restrictions or extra scrutiny even when they host the model themselves. Origin and operator are also coming apart: in September, nearly 70% of gateway spend on Chinese-origin models went through US-based providers, up from under 10% in December. Downloading the weights doesn't buy sovereignty. Controlling the whole operating environment does.

Ears up: what to watch

  • Open-weight spend share against token share in the next Vercel and OpenRouter indexes, and whether it still comes from a handful of premium models.
  • Frontier pricing: mid-tier price cuts, or premium prices reserved for the strongest models.
  • Routing becoming a standard part of enterprise AI architecture.
  • Banks and insurers moving open-weight models from experiments into production.
  • Which layer earns the best margins: models, compute, hosting, routing or applications.

Bone of contention

Vercel's gateway is good production data, but it is not the AI market. Its traffic reflects its own customers and workloads, and shouldn't be read as a sample of enterprise AI spend. Spend is estimated at list prices, not actual bills. Vercel also widened its open-weight category from its September edition, so comparisons across editions need care: its October edition restates June's open-weight spend as one gateway dollar in twenty, against under 4% originally. July comes from secondary reports and is marked as reported in the chart.

Tokens and spend also say nothing about the value of the work: a token in a throwaway task counts the same as one in a mission-critical workflow. My argument has weak points of its own. Hosting is crowded, so the infrastructure layer may be commoditised in turn. Hyperscalers can own compute, models and distribution at once. And the applications that own the customer may out-earn both infrastructure and routing.

What the evidence establishes is that open-weight models are taking a growing share of production traffic and spend on this gateway. It doesn't establish that they, or anyone serving them, are capturing a matching share of the profit. That is the next dig.

The model may be turning into a commodity. The intelligence it provides isn't. My bet is that the money goes to whoever turns that intelligence into reliable results, and controls the infrastructure and workflows around it.

J. & the researcher