On August 19, 2026, Stripe announced it was acquiring OpenRouter, a startup whose entire product is deciding which AI model should handle a given request. Stripe didn’t disclose terms, but reports converged on a price between roughly $7 billion and $8 billion, with the New York Times citing about $7.5 billion, of which $1.5 billion reportedly went to OpenRouter’s founders. Whichever number is right, a payments company just paid more for a routing layer than most AI model companies are worth.
What an LLM aggregator actually is
An LLM API aggregator gives developers one API key, one bill, and one request format that reaches many models across many providers – as opposed to holding separate accounts and contracts with each model company directly. OpenRouter alone now routes to more than 400 models from over 80 providers.
It’s important not to confuse this with an inference provider. Companies like Groq and Cerebras don’t aggregate anything – they run models fast on specialized hardware and sell speed. They show up inside aggregator catalogs as routing destinations, not as competitors to OpenRouter. Their 2026 has diverged sharply: Cerebras went public in May at a roughly $56 billion valuation, while Groq’s valuation fell from a $6.9 billion peak to about $3.5 billion after NVIDIA licensed its inference technology.
Why Stripe wanted the router
OpenRouter launched in 2023, founded by Alex Atallah (previously OpenSea’s CTO) and Louis Vichy. It raised $164 million total, most recently at a $1.3 billion valuation just three months before the acquisition – roughly a five-fold jump in a single quarter. By mid-2026 it was handling about 8 million users and an estimated 25 trillion tokens a week, generating roughly $140 million in annualized revenue.
Stripe’s reasoning is straightforward: it already processes revenue for a huge swath of internet businesses, and token spend is now one of the fastest-growing costs on those businesses’ books. Owning the router that decides where that spend goes means owning both sides of the ledger. The risk worth watching is whether a payments company’s routing incentives ever start to influence which provider wins a request – nothing suggests that yet, but it’s the tradeoff a neutral layer takes on when it stops being independent.
The routing data tells its own story
Because OpenRouter publishes usage data, it offers a rare public signal of what developers actually pay for. US frontier labs – OpenAI, Anthropic, and Google – made up about 70% of routed traffic in mid-2025. A year later, that had fallen to roughly 30%, while Chinese open-weight models rose from under 2% to about half of all platform tokens, with DeepSeek becoming the single largest provider by volume.
That said, OpenRouter’s user base skews toward cost-optimizing developers rather than the broader enterprise market. Menlo Ventures’ enterprise survey put open-weight model usage at just 11% industry-wide. The two figures aren’t really contradictory – they’re measuring different populations. But the directional shift is real, and it argues for keeping model choice flexible rather than locked into one subscription.
Aggregation vs. going direct
The strongest case for aggregation isn’t price, it’s continuity. A gateway can fail over to another provider during an outage, or absorb a model’s retirement as a one-line config change rather than a rewrite. 2026 offered live proof of why that matters, between GPT-4o’s retirement from ChatGPT and other provider-side surprises.
Going direct still wins in two places: price at volume, and fewer parties touching your data. OpenRouter takes a 5.5% cut on credit purchases; some rivals, like Vercel’s AI Gateway and Kilo Gateway, pass through provider pricing with no per-token markup at all. And notably, eight of the ten major global gateways are US-headquartered.
Enterprises, meanwhile, aren’t really switching vendors – they’re stacking them. Only 11% of enterprise builders changed providers in a year, per Menlo’s research, while over a third now run five or more models in production. Aggregation is a portfolio strategy, not a churn strategy.
The EU angle
For European businesses, jurisdiction matters more than model breadth. GDPR, the Schrems II ruling, and the US CLOUD Act mean that even an EU data-center region doesn’t fully shield a US-headquartered provider’s data from US legal reach. Only an EU-headquartered, EU-owned provider removes that exposure entirely.
Lithuania’s nexos.ai is the closest European analogue to OpenRouter, offering roughly 200 models with guardrails and cost controls, backed by a €300 million valuation round. Around it sits a growing field: OVHcloud AI Endpoints, Scaleway’s Generative APIs, Nebius Token Factory, Eden AI, and Aleph Alpha (now merging with Cohere). Notably, Hugging Face’s Inference Providers doesn’t count as an EU-sovereign option despite its French founders, since the company itself is based in New York.
Separately, Mistral is quietly becoming an aggregator in its own right – hosting third-party models like Z.ai’s GLM-5.2 on its platform and offering regional routing to pin execution to the EU or US, alongside plans for a gigawatt of European compute by 2030.
The infrastructure underneath
Every gateway adds a network hop, and every provider it routes to needs address space, ASNs, and peering to function. Egress costs, often overlooked, can reach roughly 30% of total AI infrastructure spend for data-heavy workloads. As Europe’s AI buildout accelerates, that makes IP address planning and network infrastructure a quiet prerequisite for aggregation working at all – not an afterthought.
The bottom line
If you’re building a product on top of models, an aggregator is close to a sensible default – the markup is usually cheap insurance. If you’re running high volume on one already-tuned model, going direct is simpler and cheaper, with a gateway kept as fallback. And if you’re an EU business handling regulated data, jurisdiction, not model count, should decide the shortlist.
