Cheap AI Models Now Handle 62% of Traffic on Vercel's Gateway, Up From 28% in Two Months

Enterprises are quietly rerouting their AI spending, and the shift happened faster than almost anyone predicted.

Open-weight AI models accounted for 62% of all tokens processed through Vercel's AI Gateway as of August 22, according to Vercel CEO Guillermo Rauch. That's up from 28.4% on June 24. In April, open-weight models held just 11% of the platform's token volume. By June they hit 29%. Two months later they crossed 62%.

Vercel's AI Gateway routes AI traffic for real production applications built by real companies, not lab benchmarks or test environments. That makes the numbers a genuine signal of how businesses are actually spending money on AI, not how they say they might.

The Money Doesn't Match the Volume

Open-weight models running 62% of tokens are NOT collecting 62% of the money.

Anthropic's closed models captured between 61% and 65% of total expenditure on the gateway in recent reporting periods, according to Crypto Briefing's reporting on Vercel's figures. Claude is processing a minority of total tokens while pulling in a majority of the revenue. Premium pricing still commands premium dollars, even as cheaper alternatives eat into raw usage share.

DeepSeek has climbed to the top, or near the top, of Vercel's token volume leaderboard, overtaking Google in processing share. That's a Chinese AI lab out-processing one of the most well-funded companies in Silicon Valley, at least by this one measure on this one platform.

Why Companies Are Switching

The driver here is cost, not ideology. Open-weight models can run at roughly one-tenth the price of closed-source competitors, according to Crypto Briefing.

AT&T reportedly cut inference costs by 56% by routing workloads to cheaper open models, according to figures cited by daily.dev. Coinbase saved around 50% doing the same thing. Neither company made this move because of some philosophical commitment to open-source software. They made it because it's cheaper and the quality gap has shrunk to the point where it doesn't matter for most tasks.

Open models now trail closed frontier models by about 3.3% on Stanford benchmark data cited in a SemiAnalysis chart, or by roughly four months according to Epoch AI's estimate. A few years ago the consensus was that open models could never catch closed labs. That consensus is dead.

The Open-Source Money Problem

There's a real tension nobody has fully resolved: who pays to train these models if they're just going to be given away?

Former Red Hat CEO Jim Whitehurst argues open weights can play the same catalytic role open source software played for enterprise computing, creating what he calls a broader competitive landscape where innovation moves faster and AI's power gets more broadly shared, according to InfoWorld. Nvidia and more than 200 other companies and organizations signed onto an "Open Weights and American AI Leadership" letter in July making a similar case.

But InfoWorld's Matt Asay raises the obvious problem: open source has never generated much direct revenue, and open weights won't either. The bet, according to Asay, isn't that any single company keeps funding expensive model training out of goodwill. It's that enough companies have different, self-interested reasons to keep contributing that the overall supply never dries up.

Meta wants to avoid depending on someone else's AI platform the way it depends on Apple's mobile ecosystem, which Meta CEO Mark Zuckerberg cited in 2024 as a reason for open-sourcing Llama. Alibaba wants cloud consumption. Nvidia wants chip demand regardless of who trains the model. DeepSeek and Moonshot want global attention. Different incentives, same output: cheap or free model weights flooding the market.

What This Doesn't Prove

None of this means open-weight models are about to replace Anthropic or OpenAI. Asay's own framing is useful here: Linux became essential to enterprise computing, but plenty of Windows servers are still running decades later. Open-weight models like Kimi or Nemotron aren't toppling the closed frontier labs. They're carving out the high-volume, lower-stakes workloads where a 10% cost model that trails by 3.3% on quality is an easy call for a corporate budget.

Rauch himself frames this as early-stage growth rather than a ceiling, according to daily.dev, pointing out that much of the developer tooling ecosystem, CLIs, IDEs, SDKs is still hardcoded to specific closed models. That's a technical constraint, not a permanent one. As that tooling loosens up, the open-weight share could keep climbing.

The unresolved question is whether Anthropic and OpenAI's revenue share holds at 61-65% as more routine workloads migrate away, or whether the same cost pressure that pushed AT&T and Coinbase toward open models eventually forces premium pricing down too. Vercel hasn't published forward guidance on that. Neither has Anthropic.