{"slug": "open-weight-models-reach-78-4-of-vercel-ai-gateway-token-volume", "title": "Open-weight models reach 78.4% of Vercel AI Gateway token volume", "summary": "Open-weight models handled 78.4% of token volume through Vercel AI Gateway versus 21.6% for closed models in a September 19th daily snapshot, according to Vercel founder and CEO Guillermo Rauch. Vercel's September production index, covering August traffic, put open-weight models at 56% of gateway tokens and 14% of estimated spending, up from a 7% token share in December 2025, and Vercel said open-weight adoption cut the gateway's average price per token by 23.2% during August. Rauch also said Moonshot AI and DeepSeek ranked third and fourth by estimated spend that day, with their combined spending plus Z.ai exceeding OpenAI's share on the gateway, though he noted those dollars represent inference bought through providers, mostly in the United States, not revenue flowing directly to those labs.", "body_md": "# Open-weight models reach 78.4% of Vercel AI Gateway token volume\n\n**Guillermo Rauch's daily snapshot shows open models dominating volume as Vercel moves closer to the center of the model market.**\n\n        By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)\n        · Published \n\nPrimary source: [Aligned News - AI Intelligence](https://x.com/weeklyclaw/status/2101543376374104314)\n\n## Why it matters\n\nVercel benefits as AI models become easier to swap. If applications route routine work to open models and premium tasks elsewhere, the gateway becomes the durable control point.\n\nVercel founder and CEO [Guillermo Rauch (@rauchg)](https://x.com/rauchg?ref=runtimewire) said open-weight models handled 78.4% of token volume through [Vercel AI Gateway](https://vercel.com/ai-gateway?ref=runtimewire), compared with 21.6% for closed models, in a September 19th daily snapshot. [Aligned News highlighted the figure](https://x.com/weeklyclaw/status/2101543376374104314?ref=runtimewire) on September 20th after Rauch published the underlying data.\n\nRauch has spent his career turning open-source developer projects into infrastructure businesses. The Lanus, Argentina native joined the MooTools core team as a teenager, moved to San Francisco for his first full-time engineering job at 18 and went on to work on Socket.IO, Mongoose and Next.js. His file-sharing service Cloudup was acquired by Automattic before he founded Vercel, originally called ZEIT, in 2015.\n\nThat history makes the gateway snapshot a useful window into Rauch's current bet. Vercel began by helping developers deploy web interfaces. Rauch is repositioning it as the control plane between AI applications and a fast-changing supply of models, where the winning infrastructure may belong to whoever makes those models easiest to swap, route and measure.\n\n### Token volume has moved faster than spending\n\nIn [his post on X](https://x.com/rauchg/status/2101186741042663579?ref=runtimewire), Rauch described the 78.4% share as a possible record day for open models on Vercel AI Gateway. He also said Moonshot AI and DeepSeek ranked third and fourth by estimated spend that day. Their combined spending with Z.ai exceeded OpenAI's share on the gateway.\n\nRauch included an important qualification: those dollars represent inference bought through providers, mostly in the United States. They should not be read as revenue flowing directly to Moonshot AI, DeepSeek or Z.ai.\n\nThe daily figure is also considerably higher than Vercel's latest monthly result. Vercel's [September production index](https://vercel.com/blog/ai-gateway-production-index-september-2026?ref=runtimewire), which covers traffic through August, found that open-weight models processed 56% of gateway tokens while accounting for 14% of estimated spending. Their token share had climbed from 7% in December 2025.\n\nThe gap between volume and spending matters. Token share counts input, output, reasoning, cached-input and cache-creation tokens. It can be pushed upward by long contexts, automated agents and a small number of computationally heavy workloads. Spend reflects list prices rather than the number of requests or the business value of the resulting work.\n\nA single 78.4% day therefore does not establish that open-weight models control three-quarters of the model market. It shows that they controlled three-quarters of tokens inside one managed gateway's self-selected customer traffic on that day. Vercel says the gateway routes tens of trillions of tokens each month, giving the sample real scale without making it representative of all AI usage.\n\nThe August comparison still shows a clear economic pattern. Developers are assigning large volumes of work to cheaper open models while reserving expensive frontier models for tasks where they believe the premium is justified. Vercel said open-weight adoption helped reduce the gateway's average price per token by 23.2% during August.\n\n### Rauch is building the layer above the models\n\nVercel AI Gateway reached [general availability on August 21st, 2025](https://vercel.com/changelog/ai-gateway-is-now-generally-available?ref=runtimewire). It gives developers one API for hundreds of models, with usage analytics, automatic failover and routing based on cost, latency or availability. Vercel says it adds no markup to token prices.\n\nThat makes neutrality part of the sales pitch. Applications can move between OpenAI, Anthropic, Google and open-weight alternatives without rebuilding every integration. In July, Vercel added [gateway-level routing rules](https://vercel.com/changelog/ai-gateway-routing-rules?ref=runtimewire) that let teams rewrite requests to a different model or block an unapproved model across every application using their credentials.\n\nThe feature also gives Vercel a strategic position that model labs cannot easily occupy. A lab wants developers to stay inside its own model family. Vercel benefits when developers treat models as interchangeable suppliers and place an independent routing layer between applications and those suppliers.\n\nRauch's original thesis was that infrastructure built internally by Amazon, Google and Facebook should be available to ordinary developers. By 2025, he had reframed that mission around AI applications, writing that Vercel was moving from pages to agents. AI Gateway applies the same playbook to inference: hide operational complexity, standardize the interface and become the default layer through which developers consume a fragmented market.\n\nVercel has been filling out the rest of that stack. RuntimeWire reported in June that [Vercel added persistent storage to its managed sandboxes](https://runtimewire.com/article/vercel-sandbox-persistence-ga-agent-state), giving long-running agents a place to retain state across compute sessions. In July, [Vercel Labs released a TypeScript-to-native compiler](https://runtimewire.com/article/vercel-scriptc-typescript-native-compiler-no-javascript-engine), another attempt to pull developer workloads deeper into Vercel-controlled infrastructure.\n\n### The scoreboard is part of the product\n\nVercel made its [AI Gateway leaderboard data openly downloadable](https://vercel.com/changelog/open-data-and-shareable-charts-for-ai-gateway-leaderboards?ref=runtimewire) under a CC BY 4.0 license in July. The leaderboards rank models, labs, applications and inference providers using daily data aggregated across trillions of tokens.\n\nPublishing the data turns Vercel into an observer of the market as well as a vendor inside it. Each model launch, price cut and workload migration gives developers another reason to consult Vercel's charts. The reports also advertise the gateway's scale and reinforce the argument for placing Vercel between an application and its model providers.\n\nThe direct economics remain less visible. Vercel says it charges no token markup, and it has not published AI Gateway customer counts or gateway revenue. The broader commercial logic is integration: a gateway customer can also consume Vercel's deployment, compute, sandboxing and developer tools.\n\nInvestors have already backed that expansion. On September 30th, 2025, Vercel [raised a $300 million Series F at a $9.3 billion valuation](https://vercel.com/blog/series-f?ref=runtimewire), co-led by Accel and GIC. BlackRock, StepStone, Khosla Ventures, Schroders, Adams Street Partners and General Catalyst joined the round, alongside existing investors GV, Notable Capital, Salesforce Ventures and Tiger Global. Vercel also announced an approximately $300 million tender offer for employees and early investors.\n\nThe 78.4% snapshot supports Rauch's wager that model choice will remain fluid. Open-weight models are already absorbing the bulk of tokens in Vercel's sample, while premium closed models retain disproportionate spending. Both trends make routing valuable. Developers need a way to push routine work toward cheaper models, preserve expensive capacity for harder tasks and change the mix again when the next model arrives.", "url": "https://wpnews.pro/news/open-weight-models-reach-78-4-of-vercel-ai-gateway-token-volume", "canonical_source": "https://runtimewire.com/article/vercel-ai-gateway-open-weight-models-78-percent-token-volume", "published_at": "2026-09-20 08:14:49+00:00", "updated_at": "2026-09-20 08:22:26.835995+00:00", "lang": "en", "topics": ["ai-infrastructure", "large-language-models", "ai-products"], "entities": ["Vercel", "Vercel AI Gateway", "Guillermo Rauch", "Moonshot AI", "DeepSeek", "Z.ai", "OpenAI", "Aligned News"], "alternates": {"html": "https://wpnews.pro/news/open-weight-models-reach-78-4-of-vercel-ai-gateway-token-volume", "markdown": "https://wpnews.pro/news/open-weight-models-reach-78-4-of-vercel-ai-gateway-token-volume.md", "text": "https://wpnews.pro/news/open-weight-models-reach-78-4-of-vercel-ai-gateway-token-volume.txt", "jsonld": "https://wpnews.pro/news/open-weight-models-reach-78-4-of-vercel-ai-gateway-token-volume.jsonld"}}