Open-Source Models Hit Record 78.4% of Vercel Gateway Tokens

Episode 1 · Open-Weight Models Hit Record 78.4% of Tokens on Vercel AI Gateway (2026-09-19, 10 posts)

Vercel CEO Guillermo Rauch disclosed that open-weight models accounted for a record 78.4% of daily token volume on Vercel AI Gateway, versus 21.6% for closed models. Shares were widely circulated and discussed as fresh evidence that open models are penetrating production traffic.

Confirmed

• Daily open-weight token share of 78.4% vs. 21.6% closed (gateway data shared by Rauch, relayed by @GaryMarcus, @charlesirl and others)
• Per @SumitGup, open models were around 40% in June, nearly doubling in three months; DeepSeek V4.1 Flash alone accounted for 59.3% of gateway token volume
• By spend, open models collectively exceeded OpenAI for the top position; Moonshot AI and DeepSeek ranked 3rd and 4th by daily spend, and together with Z.ai the three open-model providers led overall
• Vercel published its AI Gateway Production Index report for September 2026 (relayed by @cramforce, based on tens of trillions of tokens routed monthly between production apps and AI labs), showing open-weight models at 56% of production token volume, with Astro among the cited data points
• Polymarket also cited and amplified the record 78.4% figure

Unconfirmed

• Whether the single-day 78.4% is sustainable or a record peak; the monthly figure of 56% shows a notable gap

Why it matters

• Commenters such as @GaryMarcus and @charlesirl noted token volume and spend are different measures, yet both show open-model share rising sharply
• In a post relayed by @markjeffrey, Chamath predicted the top three models will all be open within 12 months, with cloud providers carrying the data as winners, naming Nebius, Baseten, Together and Fireworks as inference infrastructure beneficiaries

• Open-weight models now take 56% of production token volume, per Vercel index — cramforce · 2026-09-19
• Open models hit record 78.4% of token volume on Vercel AI Gateway; Kimi, DeepSeek, Z.ai spend beats OpenAI — charles_irl · 2026-09-19
• Open models now outspend OpenAI on Vercel gateway, taking 78% of token volume — soumitrashukla9 · 2026-09-19
• Open-weight models hit 78.4% of token volume on Vercel AI Gateway — GaryMarcus · 2026-09-19
• Open models hit 78% of Vercel token volume; Chamath predicts top 3 models will be open source — markjeffrey · 2026-09-20
• Open Models Hit 78% of Vercel AI Gateway Token Volume; Moonshot and DeepSeek Spend Tops OpenAI — markjeffrey · 2026-09-20
• Open-weight models hit 78.4% of token volume on Vercel's gateway, up from ~40% in June — SumitGup · 2026-09-20
• Open-source AI models hit record 78.4% of token volume on Vercel's AI Gateway — Polymarket · 2026-09-20
• Open models hit 78% of Vercel token volume; Moonshot, DeepSeek, Z.ai spend tops OpenAI — SumitGup · 2026-09-20
• Open models hit 78.4% of Vercel AI Gateway tokens; DeepSeek and Moonshot spend top OpenAI — realmeetjames · 2026-09-21

Episode 2 · Open-Source Models Hit Record 78.4% of Token Volume on Vercel AI Gateway (2026-09-21, 5 posts)

Vercel founder Guillermo Rauch disclosed that open-source models accounted for a record 78.4% of daily token volume on Vercel AI Gateway, versus 21.6% for closed models. By inference spend, Moonshot and DeepSeek ranked third and fourth that day, and together with Z.ai the three Chinese vendors outspent OpenAI. The figures were widely reshared and sparked debate about the eroding moat of closed labs.

Confirmed

• Open-source models hit 78.4% of daily token volume on Vercel AI Gateway (closed: 21.6%), per Rauch, a likely single-day record.
• Moonshot and DeepSeek ranked 3rd and 4th by inference spend; combined with Z.ai, the three Chinese vendors outspent OpenAI.

Why it matters

• Resharer ccerrato147 titled the story "I Have No Moat And I Must Scream," arguing closed vendors' moats are failing.
• Alex Banks rounded up reactions: Emad Mostaque asked what labs do once open models match the frontier after DeepSeek's V4.1 Flash, estimating its training cost at $10M, far below comparable GPT models; the remaining moat is brand.
• gerardsans argued firms are moving AI pipelines partly or fully to self-hosted open models, breaking the per-token business model—open models run locally on desktops, pinning token prices near zero as lab monopolies keep eroding.

• Open-source AI models hit record 78.4% of token volume on Vercel's AI Gateway — ccerrato147 · 2026-09-21
• Open models hit 78.4% of token volume on Vercel AI Gateway; Chinese labs' spend tops OpenAI — burny_tech · 2026-09-21
• Open models hit 78.4% of Vercel token volume; Chinese labs' spend tops OpenAI — AccBalanced · 2026-09-21
• Open-source models running locally have pegged token prices near zero — gerardsans · 2026-09-21
• Open Models Hit Record 78.4% of Vercel Token Volume—Frontier Labs' Last Moat Is Brand — thealexbanks · 2026-09-22

Episode 3 · Open-Source Models at ~10% of AI Revenue? Analysis Slammed as Outdated (2026-09-21, 6 posts)

KonstantinPilz published an estimate claiming open-source models accounted for only 10% of global AI revenue as of September 2026, with OpenAI, Anthropic and other closed-source vendors taking the vast majority, implying serious software engineering still relies mostly on closed models. The analysis spread widely but was challenged the same day by multiple posters, led by GutenLeinad and xeophon, for relying on badly outdated data.

Confirmed

• Core figure of the original analysis: open-source models at 10% of global AI revenue, closed-source vendors dominant, data as of September 2026.
• The analysis covered open-model providers TogetherAI, Fireworks, Baseten and OpenRouter with July-August revenue shares (per max16065).
• GutenLeinad argued the analysis used data from Q1 or even earlier, while actual growth far exceeded what its charts showed.
• xeophon cited specifics: TogetherAI's token volume grew over 10x from Q1 to Q2, Fireworks roughly 3x, with Baseten and other inference providers also growing significantly.
• xeophon argued "tokenmaxxing" concerns product forms like Claude Code and Codex rather than open-source usage, and may even accelerate open-source revenue as routers like Databricks grow more important.

Unconfirmed

• The true revenue and usage share of open-source models under updated data; neither side offered a corrected full estimate.
• The direction of tokenmaxxing and router growth's impact on open-source revenue remains xeophon's view, not a data-verified conclusion.

Why it matters

• The dispute centers on whether the basis for judging the market landscape is sound: if open-source inference volumes grew tenfold within months, inferring "open-source is only 10%" from stale data could systematically understate the open-source camp's momentum, affecting developer choices, enterprise procurement and investment decisions.

• Open models are ~10% of AI revenue, but the analysis used badly outdated data — xeophon · 2026-09-21
• Open model revenue analysis uses stale Q1 data: TogetherAI tokens up 10x, CRWV beats them all combined — xeophon · 2026-09-21
• Analyst's AI market data outdated: TogetherAI tokens up 10x from Q1 to Q2 — xeophon · 2026-09-21
• Open model inference is booming: TogetherAI tokens grew 10x from Q1 to Q2, analysis used stale data — GutenLeinad · 2026-09-21
• Tokenmaxxing debate: routers may accelerate open model revenue, says xeophon — xeophon · 2026-09-21
• Open model providers' revenue data badly outdated: TogetherAI tokens up 10x+ since Q1 — max16065 · 2026-09-21