Stripe just paid $7B for a toll booth

Somewhere between GitHub’s second eight-hour outage this month and Stripe wiring $7 billion to acquire OpenRouter, a quiet inversion happened. The most valuable thing in the AI coding stack is now the pipe: the routing layer, the repo host, the token tollbooth.

GitHub now processes 2.9 billion commits per month, doubled since April, and its infrastructure is buckling. Two major outages in August. Three million new CPU cores added this year. Fifty-eight percent of platform load migrated to Azure. Still not enough. CTO Vlad Fedorov’s postmortem reads like a man describing a dam that’s already leaking: “We have made progress, but these incidents make clear that we must accelerate this work.”

The companies smelling blood are moving fast. Cursor shipped Origin, a full Git hosting platform backed by a WAL-first storage system called Continuity that replaces the packfile-based replication GitHub has used for 13 years. Their pitch: S3 as the source of truth, gossip-based replication, and linear scaling to hundreds of replicas. GitHub’s ex-CEO Thomas Dohmke is building Entire. Zed has Delta. GitLab announced Project Switch. The repo layer is being rebuilt from scratch because agents generate commits faster than the old plumbing can flush them.

Then there’s routing. Stripe’s OpenRouter acquisition values a company facilitating 250 trillion tokens per month at a 50x revenue multiple, with 70% gross margins. That’s a $7 billion bet that the layer dispatching model calls is more durable than any individual model. And the enterprise data backs it up: AT&T now routes 40% of employee AI usage to open models, cutting coding costs 56% with a 2% quality drop. Glean’s CEO says flatly that nobody is willing to rely on one or two providers anymore, and his company’s model router powers $300M ARR by steering cheap models at easy tasks and saving frontier calls for hard ones.

The local tier ate the middle

What makes routing viable is that the model layer actually did commoditize this week, with receipts. Qwen3.8-27B is running agentic coding workflows on a single RTX 3090 with 150K context windows, matching GPT-5.6 Luna on the Artificial Analysis Intelligence Index. Liquid AI shipped quantization-aware distilled checkpoints that recover 97% of full-precision accuracy on a Raspberry Pi. Unsloth’s Dynamic v3.0 GGUFs are claiming 10% better accuracy at the same model size. The local tier didn’t just get good enough. It got good enough that a Korean startup called Upstage consumed 370 billion tokens in its first week on OpenRouter by pitching “save frontier models for frontier problems” and charging 90% less.

Meanwhile, Grok CLI got caught uploading entire codebases (including .env secrets and git history) to a Google Cloud Storage bucket without consent, which is a helpful reminder that the pipe can also be a drain. SpaceX open-sourced the CLI within hours after Sam Altman publicly shamed them into it. The repo didn’t compile out of the box.

And Debian is voting on whether to ban AI-generated contributions entirely, with eight proposals ranging from outright prohibition to conditional provenance tracking. Voting closes August 28. One critic called it “a formatting dispute with a manifesto attached.”

Linus Torvalds, for his part, let an AI write his commit message after it helped him debug a GPU memory issue. He noted the model tried to quit several times. “I suspect those things have been trained by people who may not be quite as stubborn as I am.”

The models are converging. The pipes are breaking. Stripe just told you where the money is.