The rail is the product now

Nobody paid $60 billion for a VS Code fork

SpaceX closed its all-stock acquisition of Cursor maker Anysphere on August 15, the largest disclosed devtools acquisition on record, first optioned in an April partnership. Read Cursor’s own announcement for what was actually traded: compute flows in (“the largest fleet of GPUs in the world”, in Cursor’s words), and a large base of professional developers flows out, a ready-made rail for routing Grok models into working editors. They paid for the rail.

Everyone else was building rails too. GitHub opened an “agent apps” category on its Marketplace, nine inaugural partners, Amplitude, LaunchDarkly, PagerDuty and SonarQube among them, plugged into the same harness that runs Copilot’s own cloud agent. Vercel added Exa to its Agent Marketplace. And the MCP roster stopped being a devtools story: MongoDB shipped a fully hosted Atlas MCP server and Nutanix shipped one for hybrid-cloud ops. Data platforms and enterprise infrastructure incumbents now ship the agent integration as a product feature, not a hackathon artifact.

Every one of these rails is curated. Someone decides who is listed, and last week’s stranded-integrator timeline put numbers on what that can cost: a four-month queue at one directory against a 29-day approval at another. More rails is good news for a devtool marketer, more places to be found and more competition among the gatekeepers. It also means your channel mix now has a line item most playbooks do not carry.

Your move: inventory which rails your product can be installed from, dual-list everywhere your buyers’ agents actually run, treat each review queue as a risk with a date on it, and keep an install path you own so no launch waits on someone else’s queue.

Vercel’s meter says your model is a one-line change

Why is the rail worth more than the model? Vercel published the cleanest answer yet, from its own AI Gateway routing data: DeepSeek hit 25% of July token volume, past Google’s 11%, and Google ran nearly 40% in April. Open-weight models went from 11% to 36% of volume in four months. Three in four high-volume teams changed at least a tenth of their model mix. Anthropic, meanwhile, held 65% of spend on 30% of volume, priced at 4.4x the field. The caveat is structural: Vercel is both the source and the router, so this is one gateway’s traffic, not the market. The shape is hard to argue with. Switching models is a one-line change, teams do it constantly, and premium pricing survives only as long as the capability gap does.

That is why Z.ai is giving GLM 5.2 away free on the gateway through August 27. When defaults are this loose, a time-boxed free window is a land grab for default-model share.

Your move: if you sell anything model-adjacent, assume your users re-decide the model monthly, and position on what survives a model swap: the workflow, the data, the integration surface. Strike every “powered by” line from your homepage that a competitor could copy by changing a string.

Meta is reading the web for itself

The quietest big move: Promptwatch’s crawler logs show Meta-WebIndexer going from ~2.2% to 37.8% of tracked AI-crawler requests in under a month, now the heaviest AI crawler they track, more than one in three requests, coinciding with reports that Meta is building a web index of its own rather than renting Google’s. Whether or not a Meta search product ships, the crawl is real today, and it says the crawler population your docs serve is unstable. An allowlist or a robots policy tuned to last year’s bot names quietly goes stale.

Your move: re-read your own crawler logs this month, not the folklore. Twenty minutes: list the top ten user agents hitting your docs, check which ones your robots policy names, and fix the gap.

Quick hits

• Capital repriced the category. Lovable doubled to $13.3B eight months after its last round; Databricks wanted $1B, was offered $15B, and took $5B at $190B; two-month-old River AI raised $1.1B; Blacksmith’s valuation rose ~10x in under a year. Your competitors’ marketing budgets just changed; assume theirs did.
• CodeRabbit raised $143M at $1.5B and moved up-stack from reviewing PRs to governing them. The review tool becoming a change-control layer is the same rail logic applied to the merge queue.
• Debian opened a binding project-wide vote on AI/LLM contributions, eight ballot options from outright ban to disclosure-conditioned acceptance, closing August 28. The largest project yet to decide the question by vote rather than maintainer fiat.
• Stripe’s transaction data put numbers on localisation: local payment methods lifted conversion 7.4% and revenue 12% on average, and the fastest-growing AI companies use twice as many of them. Vendor data with a product attached, but the sample, 88% of the Forbes AI 50, is hard to dismiss.
• Two honest-content plays earned the swipe file: Val Town rebuilt its docs on its own platform and printed the cost (UX over AX over DX, ~500ms cache misses named next to the ~100ms edit loop), and Netlify ran one prompt through 11 models and published the failures alongside a 216x credit-cost spread. Both are in the examples gallery.

Watch

1. Cerebras’ free-tier transition lands today, as this issue ships: existing free accounts move to a credit-based tier requiring a payment method. The test set last week stands: a clean landing returns deprecation-named errors against a stable model catalog; a dirty one is silent 404s.
2. Debian’s vote closes August 28. Which of the eight options passes is the first binding, project-scale answer to “are agent contributions welcome”, and every maintainer weighing their own policy will cite it.
3. Two clocks run out August 31. No incumbent has shipped or ruled out a data-for-discount tier answering Meta’s, and no second stranded-integrator account or published review SLA has landed for Anthropic’s MCP directory. Silence past the month’s end is itself the answer: the entrants are noise, and the queue is the policy.