Stripe bought the neutral layer for $7.5B — and Ramp gave it away free the same day

What happened, in order

Sunday 16 August, Bloomberg reported Stripe was nearing a deal for OpenRouter at over $7 billion. Wednesday 19 August, Stripe confirmed it, describing OpenRouter as “a leading AI model gateway and routing platform” that helps businesses route and optimise token usage across more than 400 models from over 80 providers. Stripe disclosed no financial terms. The $7.5 billion figure, and the split of roughly $1.5B to founders and $6B to investors, comes from the New York Times; TechCrunch adds that Stripe outbid Databricks.

The same Wednesday, Ramp launched Router at router.com — an OpenAI-compatible endpoint that routes across models from OpenAI, Anthropic, SpaceXAI, DeepSeek, Moonshot, Minimax, Nvidia and Z.ai, with Google’s Gemini listed as coming soon. Free routing through the end of 2026. US-only. A waitlist, being drained as capacity allows.

Ramp also publishes the AI Index, the most-cited public dataset on which labs are winning corporate spend. It released the August edition the day after Router launched, arguing that businesses switch labs readily and that enterprise AI spending is less sticky than investors assume — which is also the strongest available argument for buying a router.

Ramp did not build Router last month to spoil Stripe’s announcement. It has been running the thing internally for about three years, across the AI products serving its 70,000 customers. What Ramp chose this week was the date it stopped being internal.

The valuation is the tell

Three months. That is the interval between OpenRouter’s $113 million Series B at a reported $1.3 billion valuation in May and a $7.5 billion sale in August. OpenRouter did not become 5.8 times better at routing between May and August. What changed is what the seat is worth.

The obvious reading — that Stripe overpaid in a hot market — is the boring one and probably wrong. Read Patrick Collison’s framing instead: “Tokens are the central currency for companies building with AI, and it’s clear that the real-world economic potential will depend on making good use of scarce compute resources.” Then read Stripe’s supporting note that 88% of the Forbes AI 50 already use Stripe products.

Stripe’s business is standing between a company and its money and taking a small, legible cut. OpenRouter’s business is standing between a company and its models. For a growing set of firms, those are converging into the same flow. Every token routed is a purchase decision, and the gateway is the only component with a complete real-time view of it: vendor, model, volume, latency, cost, fallbacks. Nothing else in the stack sees all of it. Not the cloud — Bedrock sees only what runs on Bedrock. Not the lab — each sees only its own traffic. Not the IDE. The gateway sees everything.

That is what $7.5 billion buys: not routing software, which is not especially hard, but the position.

Ramp’s answer: give it away

Which is precisely why Ramp’s move is the more interesting half of the week. Ramp sells corporate cards and spend management. It raised $750 million at a $44 billion valuation in June. It does not need routing revenue, and Router’s pricing says so: free through 2026, $26 in launch credits, 50% off GPT-5.6 Sol through 18 September, and no announced 2027 price.

Free is a position, not a discount. If the gateway’s value is visibility into AI spend, and your existing product is a system of record for company spend, then routing is not a product line — it is an instrumentation play for the product you already sell. Ramp does not need you to pay for Router. It needs your token spend to appear in Ramp.

The feature set reads accordingly. Router ships four routing strategies: a Flex tier that shifts requests to discounted service tiers when latency permits; shadow models that test a candidate on production traffic without affecting users; benchmark routing weighted across up to three benchmarks of your choosing; and NVIDIA Switchyard escalation, which sends the hard steps of a workflow to a capable model while routine steps stay cheap. The dashboard reports token spend, cost, latency and fallback attempts. That is a cost-control console with an inference API attached, not the reverse.

Ramp’s founding engineer Veeral Patel puts the demand-side case plainly: the best model changes constantly across GPT, Claude, Gemini, Grok, Qwen, DeepSeek, Kimi and GLM, with prices and capabilities moving every week. He is right, and that volatility is exactly what created the opening — as the August price war showed when it flipped direction mid-month, and as Gemini 3.7 Flash’s introductory pricing showed on the way down. When prices were stable, a gateway was a convenience. When they move weekly, the ability to switch without a rebuild is worth money — and whoever grants it holds something.

The sentence to sit with

From OpenRouter’s own post announcing the sale:

There are few companies on earth we would have considered selling to; our mission, our neutrality, and our lead in the market make the story for independence strong.

That is an unusually honest sentence, and worth reading twice. It concedes the argument for staying independent was strong — on the merits, on the market position, on the mission. And then reports the outcome anyway.

CEO Alex Atallah’s accompanying framing holds up: “We believe intelligence will be multi-model: no single model will be optimal for every task, and developers need a neutral layer to orchestrate and manage them all.” The diagnosis is correct. Multi-model is how production systems are actually built now — it is the premise under Warp’s model-agnostic agent fleets and under every serious harness on the best AI harnesses list.

The gap is between a neutral layer exists and a neutral layer stays neutral. Neutrality is not a property of software. It is a property of a business model, and it survives exactly as long as the business model does not reward tilting. OpenRouter has committed that its product, mission and current commitments remain unchanged, and that it will operate independently after closing. Believe that for the near term. Just do not confuse a commitment of intent with a term you can hold anyone to.

What buyers should actually do

Nothing urgent. Both gateways work; both keep working next week. But three habits are worth adopting now, while switching is still cheap.

Keep the gateway a config value. Both OpenRouter and Ramp Router expose OpenAI-compatible endpoints, which means the swap cost between them is deliberately near zero. That is a gift from a competitive market and it will not necessarily persist. Do not let gateway-specific features creep into application code you would have to rewrite. The reason multi-model orchestration works at all is that the seam stays thin.

Log your own spend. If the gateway’s dashboard is your only record of what you spend on models, you cannot price a migration without asking the incumbent for the numbers. Keep an independent tally of tokens by model and by workload. This costs almost nothing to start and is very annoying to reconstruct later.

Read the retention default before the pricing page. Ramp Router retains inputs, outputs and tool calls for one year by default, PII stripped before use for product improvement, opt-out. That may be entirely fine for your workload. It is not fine for all of them, and it sits oddly against the direction the labs are moving — OpenAI’s private safety processing and zero-data-retention tiers exist because enterprise buyers pushed hard on exactly this. A gateway is one more hop that sees plaintext prompts. Price that hop the way you price the 10% AWS charges for data residency: as a real, itemisable cost of a compliance position, not an afterthought.

For teams whose model traffic runs mostly through a coding agent rather than a bespoke application, this week changes less than it looks. Cursor, Claude Code and Codex each manage model selection inside their own subscription, and the gateway question does not arise until you build something yourself. If you are choosing between agents rather than building on raw APIs, the coding tools comparison is the more useful place to start — and the Cursor vs Claude Code breakdown covers how each handles model choice.

The uncomfortable version

Strip out the deal mechanics and this week says something plain about the shape of the AI stack.

The labs are consolidating upward into products. The clouds are consolidating downward into distribution. And the thin neutral layer in between — the part that made switching cheap, that kept the multi-model promise honest, that existed precisely because nobody with a stake owned it — has been bought by a payments company and cloned for free by its competitor, inside 24 hours.

Neither outcome is villainous. Stripe paid a fair price for a real asset. Ramp built a good product and priced it to win. But buyers who treated the gateway as the load-bearing guarantee of model portability were relying on a structure that turned out to be one funding round deep. The guarantee was never the vendor. It was the thin seam in your own code, the one you can point anywhere.

Keep that seam thin. It is the only part of this you own.

And watch for the same shape one layer up. The day after Stripe confirmed the OpenRouter deal, Salesforce launched Slack Code — a free container that hosts Claude, Devin, GitHub Copilot and Vercel agents on equal terms and sells none of them. Same play, different layer: be the shell, be free, be where the work already happens.

And watch the variant that makes no announcement at all. Days later Nvidia agreed to pay Poolside $6 billion for a non-exclusive licence to its model-building pipeline and to hire 109 of its roughly 115 engineering and research staff — without acquiring the company, and so without triggering anything in any customer’s contract. Stripe buying OpenRouter is consolidation you can see and plan around. That one is consolidation that leaves the org chart intact and moves the capability out from under it.

Update, 22 August 2026: the case for owning your routing layer got a concrete price tag the next day. OpenAI cut GPT-5.6 Sol to $4/$20 per million, taking the frontier price crown from Claude Opus 5 — but only until 21 November 2026, because the rate is an explicit three-month promotion. That is the exact scenario a router is for: a rate that is worth capturing, on a clock that guarantees you will want to move again. Teams that can re-point traffic with a config change collect the discount and hand back the reversion; teams that cannot get to do a migration twice.

Questions or corrections? Email Pick Right. Want the full list? See all news.