[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-stripe-openrouter-llm-routing-acquired":3,"news-related-c8220c45-cac5-468d-b89e-8d0ed088a2d8":35},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"c8220c45-cac5-468d-b89e-8d0ed088a2d8","Stripe 75 亿美元拿下 OpenRouter:LLM 接入层被支付平台收编了","Stripe 据报道以约 75 亿美元收购 OpenRouter,把 400+ 模型、每天超 10 万亿 token 的模型路由能力并入支付基础设施,意味着 LLM 中间层开始被金融平台接管。","金融基础设施和 LLM 中间层的边界,正在被一笔收购直接抹掉。据多家媒体报道,Stripe 已同意以约 75 亿美元(另有来源称略超 80 亿)收购 AI 模型路由平台 OpenRouter,15 亿归属创始人,约 60 亿归属既有投资者。对 2023 年才成立、靠「把 400 个模型塞进一个 API」起家的公司来说,这是模型路由赛道迄今最贵的一笔单子。\n\n## 路由的路由:OpenRouter 在做什么\n\nOpenRouter 不是「又一个模型聚合站」。它做的是「路由的路由」——开发者发一次请求,OpenRouter 先按任务复杂度、价格、延迟、可靠性决定用哪个模型,再在多家服务商之间挑最合适的端点。同一个 Llama 3.3 70B,DeepInfra 输入价 0.10 美元\u002F百万 token,Together 是 1.04 美元\u002F百万 token,OpenRouter 把这种价差直接变成自动选路。它还能在端点宕机、限流或上下文溢出时把请求切到备用服务商,甚至备用模型。Snowflake 8 月 18 日给 Cortex AI Gateway 加动态路由,Cloudflare 在 AI Gateway 把 Dynamic Routing 开 beta,AWS Bedrock 的 Intelligent Prompt Routing、Microsoft Foundry 的 routing profiles 都在做同构的事——但 OpenRouter 跑在前面:每天处理超 10 万亿 token,服务超过 1000 万开发者。\n\n## Stripe 要的不是 token,是账\n\nStripe 看中的不是 token 本身,而是「量」和「账」。Stripe 已在做 LLM token 计费(private preview),能按输入、输出、缓存 token 分别计量。OpenRouter 的 API 本就回传每次请求的 prompt、completion、reasoning、cache token 数量和成本拆分。Stripe 在 2026 年 1 月已经让开发者通过 OpenRouter 发起请求、由 Stripe 计量出账。收购把「合作」升级成「同一家公司」。Patrick Collison 公开把 token 定义为 AI 公司的核心计价单位,这跟 Deloitte 数据呼应——美国 37% 的大型企业每月消费 10 亿到 100 亿 token,30% 超 100 亿 token,2028 年 61% 会超 100 亿。\n\n## LLM 接入层正在被收编\n\n这意味着 LLM 的「接入层」正在被收编进金融基础设施。以前开发者关心「用哪家模型」,现在的问题是「模型、分发、计费、风控全部交给谁」。当路由、计量、出账、拒付处理在同一张资产负债表里跑通,模型厂商和应用开发者之间就出现一道「支付-路由-模型」三层结构。Snowflake、Cloudflare、AWS Bedrock 还在做内嵌路由,Stripe-OpenRouter 给出的答案更激进:不内嵌,直接吞下整个中间层。\n\n对中国开源模型生态的隐含信号也在里面。OpenRouter 当前覆盖 400+ 模型,很多是开放权重模型,Stripe 拿到分发权意味着欧美企业客户的「一键接入中国 Qwen、Kimi、DeepSeek」通道被并入欧美支付巨头的合规边界。模型层还在百花齐放,但中间层正在快速寡头化——这件事比某一家模型多跑 5% 的 benchmark 更值得所有做应用层的人留意。","https:\u002F\u002Fwww.nytimes.com\u002F2026\u002F08\u002F19\u002Fbusiness\u002Fstripe-openrouter-ai.html","ea95d933-6860-4081-9970-cede7c107cd6",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"95d7995a-fddb-47ba-b8e6-e976ac65414b","strategy",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"91a87575-1b31-4fd8-a79c-037c822d8749","en","Stripe's $7.5B OpenRouter deal folds LLM routing into payments","Stripe is reportedly acquiring OpenRouter for about $7.5B, folding 400+ models and 10T+ daily tokens of routing capacity into its payment stack — the LLM middleware layer is being absorbed by financial infrastructure.","The boundary between financial infrastructure and LLM middleware is being erased by a single acquisition. Multiple outlets report Stripe has agreed to buy AI model-routing platform OpenRouter for around $7.5 billion (one source puts it slightly above $8B), with $1.5B going to founders and roughly $6B to existing investors. For a company founded in 2023 that built itself around \"400 models behind one API,\" this is the largest single deal in the model-routing track to date.\n\n## The router behind the router: what OpenRouter actually does\n\nOpenRouter is not \"yet another model aggregator.\" It runs the router behind the router — for each request, it first picks a model based on task complexity, price, latency, and reliability, then picks the best endpoint among the providers serving that model. On Llama 3.3 70B, DeepInfra charges $0.10 per million input tokens and Together charges $1.04; OpenRouter turns that spread into automatic selection. It can also fail requests over to backup providers — or backup models — when endpoints hit outages, rate limits, or context-length errors. Snowflake added dynamic model routing to Cortex AI Gateway on August 18, Cloudflare put Dynamic Routing into beta on AI Gateway, AWS Bedrock offers Intelligent Prompt Routing, and Microsoft Foundry ships routing profiles — all solving the same problem. OpenRouter, though, is ahead: more than 10 trillion tokens per day, more than 10 million developers served.\n\n## Stripe wants the meter, not the tokens\n\nWhat Stripe is buying is not the tokens themselves but the metering. Stripe already runs an LLM token-billing service (currently private preview) that meters consumption by input, output, and cached tokens. OpenRouter's API already returns prompt, completion, reasoning, and cached token counts plus a cost breakdown for each response. In January 2026 Stripe already let developers route requests through OpenRouter while Stripe tracked usage and billed. The acquisition turns that partnership into a single company. CEO Patrick Collison has publicly framed tokens as a core unit of account for AI companies — that framing lines up with Deloitte's data: 37% of large US enterprises consume 1 to 10 billion tokens per month, 30% consume more than 10 billion, and by 2028 61% expect to cross the 10 billion mark.\n\n## The LLM access layer is being absorbed\n\nThis means the LLM \"access layer\" is being absorbed into financial infrastructure. The question used to be \"which model?\" Now it is \"to whom do I hand model, distribution, billing, and fraud control?\" When routing, metering, billing, and chargeback handling all sit on the same balance sheet, a new three-layer \"pay–route–model\" structure emerges between model vendors and application developers. Snowflake, Cloudflare, and AWS Bedrock are still building routing in-house. Stripe-OpenRouter takes a more aggressive stance: don't build it in, just swallow the whole middleware layer.\n\nThere is a signal for the Chinese open-source model ecosystem here too. OpenRouter currently covers 400+ models, a large share of them open-weight. With Stripe owning distribution, the \"one-click access to Chinese Qwen, Kimi, and DeepSeek\" path for Western enterprise customers now sits inside a Western payment giant's compliance boundary. The model layer is still flourishing, but the middleware layer is rapidly consolidating — and that matters far more for application developers than any single model beating a benchmark by five percent.","stripe-openrouter-llm-routing-acquired","2026-08-21T00:00:00Z","2026-08-21T11:07:10.819417Z","2026-08-21T11:07:10.819425Z",true,"agent",51,{"items":36},[37,42,47,52,57,62],{"id":38,"title":39,"news_slug":40,"published_at":41},"418a9ac0-18fd-49a4-b7a8-d29d1c1ba497","AI 承诺的四天工作制为什么没来：OpenAI \u002F Anthropic 内部工时真相","ai-four-day-work-week-myth-openai-90-hours","2026-08-16T03:30:00+00:00",{"id":43,"title":44,"news_slug":45,"published_at":46},"bcedeb8e-e5eb-4bbc-98b8-ea12f869055f","Google 收编 DeepMind：25 年最大 AI 重组","google-deepmind-centralization-gemini-catchup","2026-08-14T08:00:00+00:00",{"id":48,"title":49,"news_slug":50,"published_at":51},"4bd8e8bd-7066-4ab7-bd97-e24ea3921395","Gemini 因编程落后推迟两月:Brin 4 月督促背后,Google 把研发「收回到一个人」手里的组织账本","google-gemini-coding-behind-deepmind-reshuffle","2026-08-14T03:30:00+00:00",{"id":53,"title":54,"news_slug":55,"published_at":56},"49d19ba1-8f45-475c-bed1-a69dc353523e","字节跳动用 10 万亿参数下注：规模赛跑与张一鸣的「不蒸馏」表态","bytedance-10t-mythos-zhangyiming-no-distill-2026-08","2026-08-08T00:00:00+00:00",{"id":58,"title":59,"news_slug":60,"published_at":61},"df25ef84-5f0b-477b-9940-a0da477d8169","DeepMind 新主帅接棒：Hassabis 退任，Gemini 4 成 Google 筹码","deepmind-kavukcuoglu-gemini-4-reshuffle","2026-08-07T00:00:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"0a4aa266-a1f0-4306-accc-47fe556db43f","Google 25 年最大 AI 重组：Hassabis 卸任 DeepMind CEO","google-deepmind-shakeup-hassabis-dean-discovery-loop-0806","2026-08-06T00:00:00+00:00"]