AI Model Routers Surge as Stripe Buys OpenRouter Amid GPU Costs

The middleman moment

In the summer of 2026, the most closely watched layer of the AI stack was not a chip or a frontier model. It was a router: software that decides which model handles a given request. Stripe announced in August that it had agreed to buy OpenRouter, a gateway that sends token traffic to more than 400 models from over 80 providers.31 Within days, expense-management company Ramp released its own routing service, called Router. Ramp says it has used the tool internally for three years.21 Cursor, Meta and even video startup Runway have also been reported as building or shipping routers, and Salesforce and Databricks are adding routing to their platforms.23

The timing is the story. The industry is spending record sums on data centers, and Nvidia is promising large drops in the cost of inference. At the same moment, enterprises are piling into software meant to cut spending on that same compute. These trends do not conflict. Together they show where AI economics stands right now: tokens per unit of hardware are getting cheaper, but total bills keep rising.

What happened, and what it cost

Reports on the OpenRouter price differ, which is common for a private deal. The Wall Street Journal first reported in July that the companies were in talks at a valuation near $10 billion.38 Bloomberg later put the final price above $7 billion, according to TechCrunch.35 The New York Times reported $7.5 billion, citing a person familiar with the deal: $1.5 billion to the founders and $6 billion to investors.40 Neither company disclosed terms.40 Every version points to a large markup. OpenRouter was valued at $1.3 billion in a funding round in May.35

One report adds a twist. Nvidia's executives reportedly wanted to bid but needed more time, and the founder would not wait.39 That account comes from a single outlet and has not been widely confirmed, so it is best treated as unverified. What is on the record is that Nvidia invested in OpenRouter40 and appears on Stripe's list of OpenRouter customers.31 The leading GPU maker is both a backer and a user of the software that decides how its customers' compute gets used.

The growth figures explain the price. Menlo Ventures, an investor, says OpenRouter's token volume has grown about 30,000-fold since launch, to an annual run rate above 4.5 quadrillion tokens. By that count, volume has doubled roughly every 11 weeks.32 OpenRouter's chief operating officer told Fortune that heavy use of coding agents such as Anthropic's Claude Code has driven demand.23

The hardware paradox driving demand

On the chip side, Nvidia's message all year has been cheaper inference. At CES in January, the company launched its Rubin platform: six chips designed together that it says can cut the cost per inference token by up to 10x compared with Blackwell.3 Coverage of that claim varies in emphasis. HPCwire reported Jensen Huang describing a 5x gain in inference performance per GPU and noted that a 1.6x increase in transistors alone cannot produce a 10x result.7 One startup-focused analysis said the 10x figure was measured on specific model setups and suggested real-world gains could land closer to 3x to 5x.8 Rollout timing has also moved. Nvidia originally said partners would get systems in the second half of 2026. One later account says production shipments were expected to start this fall.10

Even a 10x cut in unit cost would not solve the problem routers address, because cheaper tokens have so far meant more tokens. One analysis estimates inference took $23.3 billion of a $42 billion AI cloud market in 2026, overtaking training for the first time. It argues that unit costs fall about 50% a year while volume grows faster, so bills keep rising.17 The same analysis found that on identical H100 hardware, the cost per million output tokens ranged from $0.21 to $15.25 depending on load.17 Those figures point to utilization and workload matching as the variables that matter. Hardware generation is only part of the picture.

Prices for the hardware itself are just as scattered. One cost survey puts H100 rental between roughly $0.30 and $14.90 per hour depending on the provider, a 23x spread for essentially the same chip. It also estimates that AI-first software companies spend 40% to 50% or more of revenue on cost of goods sold.19 Another price index places Azure's on-demand H100 rate at $12.29 per GPU-hour against $1.49 on a marketplace provider.12 When the same compute can cost wildly different amounts, a layer that can steer traffic across providers is valuable.

The buildout behind it all

All of this sits on top of the largest corporate spending cycle in technology. Estimates of 2026 hyperscaler capital spending vary widely. CreditSights puts the top five at around $602 billion, J.P. Morgan has cited $697 billion42, and other analyses run to $775–800 billion.43 UBS's projection is far higher: about $1.009 trillion in 2026, and roughly $4.1 trillion from 2026 through 2028.49 The spread reflects different groups of companies and different definitions. Every estimate points the same way.

The financing is changing too. According to one analysis of Epoch AI data, aggregate hyperscaler capex is on track to exceed operating cash flow around the third quarter of 2026. It also finds that debt rose from 9% of capex in fiscal 2024 to 32% by mid-2026.42 That puts real pressure on the companies at the end of the chain, the ones buying tokens. If the infrastructure is increasingly debt-financed, the people renting it will want proof that each token is earning its cost. Patrick Collison, Stripe's chief executive, linked the deal to "making good use of scarce compute resources."31 Coming from a payments executive, the reference to scarcity is telling.

Why routing, and why now

The case for routing is simple. Not Diamond's chief executive told Fortune that most companies send everything to the most powerful model, which wastes money.23 A router sends easy work to cheaper models and saves frontier models for hard tasks. The supply of capable cheap models has grown. The New York Times noted that strong open-source models from China, such as Moonshot AI's Kimi, have made switching between models more practical.40

Cost is not the only motive. Dataiku's chief executive said recent access restrictions on Anthropic's Fable model pushed enterprise customers to reconsider relying on a single AI provider.23 Read this way, routers work as insurance against an outage, a price change or an access policy at one lab. Menlo's investors note that OpenRouter's ability to sign contracts across many providers gave it higher uptime than individual frontier models.32

The crowded field shows how much the category is still being defined. Microsoft's Foundry offers a trained router with Balanced, Cost and Quality modes.24 Amazon Bedrock's prompt routing is limited to two models from the same family.27 LiteLLM launched a tiered Auto Router in beta in July.24 Palo Alto Networks acquired Portkey and renamed it Prisma AIRS AI Gateway, a sign that security vendors want a role in governing AI traffic.24 Pricing ranges from OpenRouter's 5.5% pay-as-you-go platform fee to Bedrock's $1 per 1,000 routed requests.25

The limits of the pitch

There are reasons for caution, and some come from OpenRouter's own backers. Menlo Ventures argues that routing on the prompt alone is "a bit of a fool's errand" for agents. In a multistep task, a wrong routing choice early on carries forward and degrades the final result.32 Menlo also says most developers use OpenRouter mainly to reach many models, not for automatic routing.32 That undercuts the simplest version of the cost-saving pitch.

Neutrality is the second issue. Consolidation reduces dependence on any single model lab but creates dependence on the gateway. Stripe now sits between developers and AI providers that compete for the same traffic, while already processing payments for many of those labs.37

The reading

Routers are not a sideshow to the GPU boom. They are what customers do in response to it. Nvidia is lowering unit costs, and hyperscalers are borrowing to add capacity. Agents are consuming that capacity as fast as it arrives. Routers sit where those forces meet. If Rubin delivers even part of its promised savings, demand will likely rise to absorb them, and the question of which model a request deserves will matter more. Stripe paid a high price for a company that was worth $1.3 billion in the spring. Its bet is that whoever controls that decision ends up controlling a share of the money flowing into AI compute.