Above the Chips: How the AI Access Layer Quietly Became Worth $7 Billion

The last piece in this series went one layer below the model, into the silicon. This one goes one layer above it, to the LLM API aggregator — a layer that sounds like plumbing and turns out not to be. It decides how a request actually reaches a model.

On 19 August 2026, Stripe announced it was acquiring OpenRouter. OpenRouter’s entire product is deciding which AI model should answer your request. Stripe did not disclose terms. CNBC reported that the New York Times put the price at about $7.5 billion. Of that, $1.5 billion went to OpenRouter’s founders. Bloomberg had reported a signed agreement at more than $7 billion three days earlier. Axios put it above $8 billion in cash and stock.

Take the lowest of those figures. A payments company still paid more for a router than most model companies are worth. That is worth understanding.

The layer above the chips: where an LLM API aggregator sits

Strip the marketing away and there are six layers between a prompt and a watt:

The aggregation layer is the one that has been changing fastest, and it is also the one most coverage describes least accurately. So, plainly: an LLM API aggregator gives you one API key, one billing relationship and one request format that reaches many models across many underlying providers. Stripe’s announcement puts OpenRouter at more than 400 models from over 80 providers. Vercel’s AI Gateway covers 200-plus models across 40-plus providers. Microsoft Foundry, which is a hyperscaler catalogue rather than a neutral router, lists over 10,000 – and made Anthropic’s Claude family generally available in July 2026, which is a quietly remarkable thing for OpenAI’s largest backer to do.

An LLM API aggregator is not an inference provider

This is the distinction worth getting right, because reporters constantly treat the two categories as one.

Groq and Cerebras do not aggregate anything. Neither is an LLM API aggregator. They run models fast on unusual hardware, and they appear inside aggregator catalogues as providers you can route to. Cerebras reports roughly 2,500 tokens per second per user on Llama 4 Maverick against about 1,000 for an NVIDIA DGX B200. Groq has posted figures in the same territory for smaller Llama models. Both numbers are vendor-supplied and workload-specific – the same caveat that applied to every chip figure in the previous article.

Their 2026 has been eventful in opposite directions. Cerebras went public on Nasdaq in May 2026 at $185 a share. The IPO raised $5.55 billion and valued the company at about $56.4 billion on a fully diluted basis. That was the largest US tech IPO by proceeds since Uber in 2019. The stock closed its first day up 68%. Groq went the other way. NVIDIA licensed its LPU inference technology in a deal reported at around $20 billion in late 2025. Groq has since rebuilt as an inference cloud operator under new leadership. Its valuation fell from a $6.9 billion peak to roughly $3.5 billion by mid-2026.

Neither company competes with OpenRouter. Both are things OpenRouter sells access to. Together AI and Fireworks AI sit in the same band – model-agnostic inference clouds, both valued in the billions during 2026 – and they too show up as routes inside gateways rather than as gateways themselves.

Why a payments company bought a router

OpenRouter launched in 2023. Alex Atallah founded it together with Louis Vichy. Atallah co-founded OpenSea and served as its CTO before leaving in 2022. Atallah described the idea, early and often, as the Stripe for AI. That framing aged unusually well.

The company raised $164 million in total, most recently a $113 million Series B in May 2026 at a $1.3 billion valuation, with backers including CapitalG, Andreessen Horowitz, Menlo Ventures, ServiceNow Ventures, MongoDB, Snowflake and Databricks. Three months later it agreed to sell for somewhere between seven and eight billion dollars, depending on which outlet you believe. That is roughly a five-fold step-up in a quarter, and it tells you how sharply the market repriced this layer during 2026.

The scale behind it: around 8 million users, on the order of 25 trillion tokens routed per week, and an estimated $140 million annualised revenue by July 2026 on a take rate of about 5.5% on credit purchases. Stripe named NVIDIA, Zoom and Lovable among the customers.

Stripe’s own explanation was blunt enough. Patrick Collison framed the deal around helping customers spend their tokens efficiently by routing requests intelligently. Which is to say: Stripe already sits on the money coming in. Token spend is the fastest-growing line of money going out for a large class of its customers. Routing decisions are where that spend is won or lost. Buying the router is buying the other side of the ledger.

One thing to watch rather than assume: OpenRouter’s value to developers has always rested on neutrality. If a payments company’s routing preferences ever start to shape which provider serves your request, that is a different product. Nothing announced suggests it will. It is simply the risk a neutral layer acquires when it stops being independent.

What the routing data actually shows

Because OpenRouter publishes its rankings, the aggregation layer has an unusual property: it produces public evidence of what developers actually pay for, rather than what benchmarks say they should.

Where the tokens actually go

US frontier labs – OpenAI, Anthropic and Google combined – accounted for roughly 70% of routed traffic in June 2025. A year later they accounted for roughly 30%. Chinese open-weight models went from under 2% of traffic in late 2024. By April 2026 they made up around half of all platform tokens. By June 2026, DeepSeek on its own was the single largest provider by token volume.

The obvious objection is that OpenRouter’s users are developers experimenting, not businesses in production. That objection is weaker than it looks: CNBC reported in July 2026 that Chinese-origin models accounted for 46% of US enterprise token usage on the platform. It is not only hobbyists.

Two caveats worth naming

Two caveats still apply, and the first is large. Menlo Ventures’ December 2025 enterprise survey put open-weight models at just 11% of enterprise workloads, down from 19% a year earlier, with Chinese models specifically accounting for roughly 1% of total enterprise LLM API usage. Set that beside OpenRouter’s 46% and you are not looking at a contradiction so much as two populations measured seven months apart: people who choose a gateway are self-selected cost optimisers, and the broad enterprise market is not. It is also worth knowing that Menlo is an OpenRouter investor, which does not invalidate either figure but is the sort of thing a reader should be told. The second caveat is simpler – cheap and fast is doing much of this work, not “better” in any benchmark sense.

But the directional point survives both, and it is the strongest argument in the whole aggregation case: the sensible default eighteen months ago is not the sensible default now. If your access strategy is a single subscription, that shift is a migration. If it is a gateway, it is a configuration change.

The trade you are actually making

The case for aggregation is usually pitched as cost. That is the weakest version of it.

The real case is continuity

The real case is continuity. When a provider has an outage, a gateway can fail over to the same model on another provider, or to a different model entirely. When a provider retires something you depend on, you change a string rather than a codebase – and 2026 gave everyone a live demonstration of why that matters. OpenAI’s GPT-5 launch in August 2025 pushed GPT-4o toward legacy status, the backlash forced a temporary reprieve, and the model was finally retired from ChatGPT in February 2026. Separately, in December 2025 OpenAI had to switch off promotional messages inside ChatGPT after paying subscribers found themselves being marketed to. Neither event was catastrophic. Both were reminders that a single-provider dependency is a dependency on someone else’s product decisions, pricing decisions and ethics decisions.

Where going direct beats an LLM API aggregator

Now the honest other side.

Direct still wins on price. A gateway’s markup is real. OpenRouter charges 5.5% on credit purchases and, for bring-your-own-key usage, 5% of the equivalent OpenRouter cost once a plan-dependent free allowance runs out. At sufficient volume that is a line item worth removing – and it explains the zero-markup wave now forming underneath it.

Vercel’s AI Gateway passes provider list pricing through with no per-token platform fee, including on your own keys. Kilo Gateway – the API layer beneath Kilo Code, the open-source coding agent – advertises zero markup and monetises through team features instead; OpenRouter’s own documentation acknowledges as much when comparing the two. LiteLLM removes the question entirely by being an open-source proxy you host yourself. Portkey and Cloudflare compete on production features rather than per-token economics.

Look at that field as a whole and one thing stands out before any of the pricing does. Eight of the ten are American.

Direct also wins on the data path. Every gateway is one more company between you and the model, one more processing agreement, one more subprocessor entry, one more place a request can be logged. For regulated workloads that is not a rounding error.

Enterprises are not switching, they are stacking

And there is a structural point that cuts against the loudest version of the aggregation pitch. Menlo Ventures’ 2025 research found that only 11% of enterprise builders switched vendors in a year; 66% upgraded within their existing provider and 23% did not switch at all. Enterprises are not fleeing providers. What they are doing is running more of them at once – one analysis of the same data puts 37% of enterprises at five or more models in production. Aggregation is how you run a portfolio, not how you run away.

Europe’s answer: EU-based AI gateway options

For an EU business the calculation changes, and it changes on a dimension that barely features in most coverage of this space: jurisdiction.

Why jurisdiction changes the maths

The mechanics are well-worn but still decisive. GDPR governs where personal data goes. Schrems II made transfers to US-controlled processors a documented risk rather than an assumption. And the US CLOUD Act reaches a US-headquartered company’s data regardless of which region the datacentre sits in – which means an EU region on a US-owned cloud reduces exposure without eliminating it. An EU-headquartered, EU-owned provider is the only configuration that removes it.

What the EU AI Act timing actually changed

Timing sharpens this. The EU AI Act’s Digital Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and pushed the high-risk obligations under Annex III out to December 2027. Most readers took that as breathing room. It is partial breathing room: the Article 50 transparency obligations still applied from 2 August 2026, and watermarking requirements arrive in December 2026. The deadline that moved is not the one that governs disclosure.

The EU AI gateway field

The closest European analogue to OpenRouter is Lithuanian. nexos.ai was founded in late 2024 by Tomas Okmanas and Eimantas Sabaliauskas, the Nord Security co-founders. It ships two products: an AI Workspace for employees and an AI Gateway. The gateway gives developers access to roughly 200 models with guardrails, cost controls, observability and automated retries. It raised an $8 million seed in January 2025 and a €30 million Series A in October 2025 co-led by Index Ventures and Evantic Capital at a €300 million valuation, with Payhawk among its named customers. It is not a unicorn and does not claim to be. It is a credible EU-jurisdiction gateway built with security as the organising idea rather than breadth.

Around it, a real field has assembled. OVHcloud AI Endpoints serves 40-plus open-weight models out of Gravelines with zero data retention. Scaleway’s Generative APIs run about ten open models from Paris, ISO 27001 and HDS certified, with SecNumCloud under review. Nebius Token Factory, Amsterdam-based and Nasdaq-listed, offers serverless and dedicated endpoints across 40-plus open models with EU regions in Finland and France and a 99.9% uptime SLA. Eden AI operates a multi-provider gateway with an EU-only endpoint. And Aleph Alpha has pivoted from frontier models to a sovereign enterprise stack, with a merger into Cohere announced in April 2026 and still pending approval.

One clarification worth making because it comes up constantly. Hugging Face’s Inference Providers is an excellent unified gateway across fifteen-plus inference backends at pass-through pricing. French nationals founded it, but the company sits in New York. It does not clear the sovereignty test, whatever the accents in the founding team.

Mistral is becoming an aggregator

The most interesting European development is that the continent’s flagship model lab is quietly turning into a platform.

Mistral raised €1.7 billion in September 2025 at an €11.7 billion post-money valuation, led by ASML with a stake of roughly 11%. In August 2026 it announced Mistral Compute. The plan is 200MW of European capacity by the end of 2027 and one gigawatt by 2030. European Compute Units finance it forward. Anchor customers buying in advance include ASML, Amadeus, CMA CGM, Caisse des Dépôts and Capgemini. Arthur Mensch has put the cost of a gigawatt at around $50 billion. Reports of a further raise near a €20 billion valuation remain unconfirmed, and Mistral has not published ECU capacity, pricing or delivery dates – worth noting before treating the gigawatt as booked.

The structural news is smaller and more telling. Mistral has begun hosting third-party open-weight models on its own platform, starting with GLM-5.2 from the Chinese lab Z.ai. Regional endpoints let customers pin execution to the EU or the US. A model lab that hosts other labs’ models, with jurisdictional routing and an uptime tier, is not really a model lab any more. It is a sovereign aggregator with its own weights included.

The layer below the layer

Here is where this stops being an AI story and becomes an infrastructure one.

Latency: every AI gateway is a hop

A gateway is a hop. Your request leaves your application, reaches the aggregator, and is forwarded to whichever provider wins the route. For text streaming the added latency is usually small. For agentic workloads that chain dozens of calls, or for multimodal payloads, it compounds – which is why inference is drifting toward regional zones rather than centralising.

Egress: the invisible bill

Then there is egress, which is where the invisible bill lives. Hyperscaler egress runs roughly $0.05 to $0.12 per gigabyte. Pure text streaming is light, on the order of 180GB a month per billion tokens a day. Retrieval-augmented and multimodal traffic escalates quickly. A 70-billion-parameter model streaming to ten thousand concurrent users can move about a terabyte an hour. For data-heavy deployments, egress can reach something like 30% of total AI infrastructure cost. Providers with zero egress fees change gateway economics materially, and that is a networking decision, not a model decision.

Address space and peering

And underneath all of it sits address space. Every inference provider, every neocloud, every regional endpoint an aggregator routes to is a set of datacentres that need IPv4 addresses, ASNs and peering to exist on the public internet. Europe’s buildout makes this concrete: Mistral’s gigawatt ambition, Nebius’s contracted capacity across EMEA. Cerebras plans to open its first European capacity before the end of 2026, with expansion across France and the Nordics. All of it adds up to, among other things, substantial demands on IP resources. IPv4 in that region ran out more than a decade ago. For operators standing up capacity across multiple EU sites, bringing your own IP space is the difference between portable infrastructure and infrastructure that has to be renumbered every time the footprint changes.

The aggregation layer is only as sovereign as the network beneath it. That part rarely makes the announcement.

So what should you really do?

Not “pick a side,” because the framing is wrong. The useful questions are narrower.

If you are shipping a product on top of models, an LLM API aggregator is close to a default. The failover, the model portability and the single invoice are worth 0–5.5% for almost everyone who is not at serious volume, and the 2026 traffic data is a strong argument for keeping your options open.

If you are running high-volume production on one model you have already tuned, direct is probably cheaper and definitely simpler. Keep a gateway configured as a fallback path rather than the primary route.

For an EU business handling regulated data, jurisdiction outranks both. The question is not which gateway has the most models. It is which entity is subject to which legal system, and the answer narrows the field to a handful of names .

The thread running through all three, and through the chips article before it, is the same one. AI competition keeps moving downward into the stack, and the layers that looked like plumbing keep turning out to be strategy. A router just sold for more than most model companies are worth. That is not a story about models at all.

FAQ