OpenRouter’s 10T-Token Order Book

Ten trillion tokens a day—and a reported $7.5 billion price. In 2026, Stripe agreed to acquire OpenRouter, a three-year-old AI model marketplace that says it routes more than that volume across over 400 models. A payments company was buying the software that chooses an AI supplier for each request.

Founded in 2023, OpenRouter began as a marketplace that helped companies switch model providers. By Stripe’s agreement, OpenRouter could choose a provider on every request.

The inference router is the durable asset: the layer that allocates each unit of AI work across models and capacity, then authorizes, governs and meters that work for settlement.

The request has replaced the vendor as the unit of choice

A model gateway gives developers one interface for many providers. Routing adds a decision to that interface. The application can select a model for each query rather than bind the entire product to one supplier.

Developers once made that choice during integration. They negotiated access, wrote against an API and changed providers when cost or quality justified the migration. OpenRouter instead directs prompts according to factors such as cost and speed. The comparison now happens request by request—Model A for this query, Model B for the next.

That change moves model selection into the model procurement stack. A coding task can favor capability. A classification request can favor price. A background process can tolerate delay. A production failure can trigger a fallback. Developers specify the constraints, and software makes the purchase on each call.

Before its latest daily figure, OpenRouter reported 25 trillion tokens a week across more than 400 models, up from 5 trillion six months earlier.

At that scale, no team can switch suppliers manually. OpenRouter makes the choice repeatedly as prices and capacity change, turning procurement into a scheduling problem.

Agents put allocation on the critical path

Human-facing applications can expose a preferred model, and users can tolerate occasional switching. AI agents instead assemble work from chains of automated calls, often without a person waiting for each answer. Cost, latency, capacity and failure handling become programmatic inputs.

Agents also change the value of speed. A user notices every second spent staring at a cursor. An unattended agent can accept some latency for lower cost, then spend the savings on more calls or longer work. As Stratechery argued in May, speed matters differently when humans leave the loop. The agent’s deadline, not the user’s patience, becomes the constraint.

OpenAI’s Agent Plugins standard bundles skills and MCP servers, with Amazon, Cursor, Microsoft and Vercel on its steering committee. Each new tool expands the set of calls an agent can make and the failures its operator must handle. One visible prompt can now conceal a sequence of model requests, tool invocations and retries.

Agents multiply those decisions, forcing a router to enforce budgets, select fallbacks and react to unavailable capacity while the work continues.

Each route can become a purchase order

Stripe’s interest is clearest at the level of the transaction. Each model request has a buyer, a supplier, a quantity and a price. The route determines the supplier. Metering determines the charge. Settlement closes the loop.

AI labs keep effective prices in motion. OpenAI, Anthropic and other providers have offered startups token credits, promotions and one-time bonuses while competing for durable business revenue. A gateway can compare nominal prices, apply credits and aggregate spending across suppliers. The routing decision serves as both an execution instruction and a purchase order.

OpenAI and Anthropic reported in investor materials that inference costs exceeded half of revenue. At that share, a small routing gain can outweigh a large saving in a peripheral cost.

Models now compete across cheap capacity, specialist performance and changing availability. Providers can optimize their own endpoints, while applications can optimize their own workloads. A multi-model gateway can compare both sides at once and make fragmented demand legible to fragmented supply.

Each OpenRouter record identifies the route, provider and volume. Stripe brings the institutional machinery for turning that usage into settlement.

An order book sees demand before model vendors do

A model company sees traffic to its own endpoint. OpenRouter sees substitution across hundreds of endpoints. It can observe which workloads move after a price change, which providers gain when a rival loses capacity and which models attract sustained use rather than launch-day curiosity.

Beginning February 8, Chinese models accounted for more than 30% of weekly token use by US companies, peaking at 46%. Over the preceding 12 months, their share had been 11%. US companies shifted workloads quickly even though the models’ national origin had not changed.

Ox Alpha exposed the mechanism. The free, anonymous model rose to the top of OpenRouter’s usage leaderboard before Z.ai identified it as GLM-5.3-Flash. Users pushed it to the top before they knew the model’s maker; distribution and a zero price supplied what branding usually does.

A leaderboard measures usage rather than value, and free capacity attracts experimentation. Even so, Ox Alpha let OpenRouter record demand before users knew whose model they were choosing.

OpenRouter can also move the market it measures. A leaderboard directs developer attention toward rising models. More testing generates more usage evidence, which improves OpenRouter’s position in provider negotiations. Once the observer ranks the field, observation becomes distribution.

Autonomy makes policy part of the route

Palo Alto Networks reportedly agreed to acquire Portkey for $120 million to $140 million. Portkey develops gateway technology that manages and secures AI agents. The buyer was a cybersecurity company, not a model lab.

Vercel supplied the cautionary case. The company said an employee’s Google Workspace account was compromised through a breach at the third-party AI platform Context.ai, contributing to unauthorized access to Vercel’s internal systems. The breach made an AI dependency a security path, though it did not involve OpenRouter or show that all gateways are unsafe.

Companies can enforce identity checks, approved-model lists, data-handling rules and audit records where a request leaves the application. The gateway can also send a query to retrieval, a model or a human operator. Human escalation assigns accountability when software cannot safely make the next decision.

As agents encounter more context-dependent choices, operators need one place to coordinate them. A gateway already sees the request, destination and response status, making it a practical location for deployment accountability. The reported prices bracket two gateway businesses: $120 million to $140 million for Portkey’s security controls, and $7.5 billion for OpenRouter’s transaction volume.

The price assumes plurality survives

A few models could become capable, cheap and reliably available enough to reduce the benefit of constant selection. Routing can add latency and produce subpar results for some queries. More choice can leave an application with a larger menu and a worse answer.

The reported price leaves OpenRouter little room to become a commodity gateway. The company reportedly generated about $140 million in annualized revenue in July, nearly triple its April level. Against the reported $7.5 billion purchase price, that implies a multiple of roughly 53.6 times annualized revenue.

Even at that growth rate, gateway fees do not explain the price. Stripe’s bet requires OpenRouter to retain economic value as model prices fall and provider capabilities converge.

OpenAI engineers reportedly found a way to more than halve inference costs. If providers pass such gains into pricing, routers have less cost dispersion to exploit. If one provider combines lower cost with adequate quality and reliable capacity, applications have less reason to split traffic.

If a universal model emerges, applications select less often; if efficiency gains flatten prices, routers have less to optimize. At the reported multiple, Stripe has little room for either outcome. Routing volume could keep growing while OpenRouter captures less value from each token.

Ten trillion tokens a day first looks like throughput. Across more than 400 models, it records a choice: which supplier got the work, at what price and under which rules. At a reported $7.5 billion, that traffic reads as order flow.