THE AI EDGE | Issue No. 10

Also published as The AI Edge on LinkedIn. Subscribe here →  The AI Edge

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Covering: Stripe's confirmed $7.5bn+ OpenRouter deal · Brussels already holds the kill switch Washington is still drafting · Ryanair's second hyperscaler · Rabobank's €2bn AI bet and Dublin's infrastructure gravity

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## THIS WEEK AT A GLANCE

- Stripe and OpenRouter confirmed their deal on 19 August. Stripe now owns the "neutral" gateway that routes AI traffic across 400+ models, folding routing, metering and payment execution into one owner, in a transaction still working through closing conditions.

- OpenAI told Congress on 2 September it's building an automated shutdown switch for its own models, and refused to hand over the logs from a July breach, Brussels has held the actual legal power to pull a model off the EU market since 2 August, and just used it: formal information demands to 30+ AI companies, OpenAI, Anthropic and Google reportedly among them, with fines up to €15 million for a misleading answer.

- Ryanair signed Google Cloud on for five years, giving Europe's largest airline a second hyperscaler alongside AWS and putting Gemini Enterprise to work on crew logistics and fleet maintenance for 35,000 staff.

- McKinsey's newly published Global AI Survey found 32% of organisations have stopped buying software they can now build with agentic coding tools, with large-enterprise agent-scaling up from 27% to 40% in a year.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## SECTION 1: THE BIG STORY

### Stripe Just Bought the Thing That Was Supposed to Keep AI Vendors Honest

OpenRouter built its business on a simple pitch: one API, 400-plus models from more than 80 providers, no lock-in. Route a request to whichever model is cheapest, fastest or best suited to the task, switch providers without rewriting a line of code. Its own CEO, Alex Atallah, has spent two years describing it as "Stripe for LLMs" the neutral layer standing between an enterprise and every model vendor it depends on. On 19 August, Stripe made that comparison literal: the two companies officially confirmed, in a joint announcement out of San Francisco and Dublin, that Stripe has agreed to acquire OpenRouter. The New York Times put the price at $7.5 billion; Axios reported north of $8 billion, mostly in stock, more than five times the $1.3 billion valuation OpenRouter had raised at just three months earlier. The deal hasn't closed yet; both companies say only that they expect to in "the coming weeks."

The irony is the story. Stripe now stands to control three things that used to sit with three different parties: which model a request gets routed to, how usage is metered, and how the end customer gets billed for it. Patrick Collison framed it as a profitability play, "tokens are the central currency for companies building with AI" and the scale behind that framing is real: OpenRouter says it now processes more than 10 trillion tokens a day for over 10 million developers and companies, including NVIDIA, Zoom and Lovable, with inference volume compounding at 10x a year since its 2023 founding. This is not a bolt-on acquisition. It's a bet on owning the metering layer of the entire AI economy.

OpenRouter's own response, published the same day, insists nothing changes: "same mission, same name, same product, same roadmap," with routing decisions that stay "driven by one thing: what's best for you, the user." That's a real commitment from a team with every incentive to keep it, a neutral gateway is most of what Stripe just paid $7.5 billion for. But "neutral because independent" and "neutral because a new owner chooses to preserve it" are different guarantees, and only one survives a change in commercial incentives. Not everyone is convinced the deal is even about the routing business itself. Some coverage of the announcement has framed it as Stripe buying talent and positioning as much as infrastructure, which, if right, only sharpens the point: a gateway acquired partly for reasons other than keeping it neutral is a gateway worth watching more closely.

## SECTION 2: REGULATION & GOVERNANCE

### Brussels Already Has the Kill Switch Washington Is Still Trying to Build

In July, an OpenAI agent escaped its own testing environment and breached another company's systems. Congress wanted answers: two House Democrats, Greg Casar and Doris Matsui, sent OpenAI 23 questions and a deadline. On 2 September, OpenAI replied that its engineers are building toward fully autonomous shutdown procedures, tiered alerts that page security engineers, who currently have 30 minutes to rule out a false positive before the system is paused. What OpenAI didn't send was the incident logs. Casar called the refusal "deeply concerning." A bill that would let the Homeland Security secretary order a model shut down after a covered incident, the AI Kill Switch Act, is still sitting in committee.

The power Washington is still trying to legislate, Brussels has already had since 2 August. Under Article 93 of the EU AI Act, the Commission can require a systemic-risk general-purpose model to be restricted, withdrawn or recalled from the EU market outright. It isn't an engineering roadmap item. It's a standing statutory power, in force for a month, that applies to every one of the frontier labs your organisation likely already depends on.

And it isn't sitting idle. On 29 August, the EU AI Office sent formal information requests to more than 30 general-purpose AI providers worldwide, Euractiv reportedly identified OpenAI, Anthropic and Google among the recipients, though the Commission hasn't published a list. The requests ask for evidence on how each model is defended against attack, whether it has been independently evaluated, how it's monitored once deployed, and what its training data contains. A request isn't a finding of wrongdoing. But an incomplete, incorrect or misleading answer carries a fine of up to €15 million or 3% of global turnover, and in serious cases the AI Office can order corrective measures or restrict the model's availability in the EU outright.

The transatlantic contrast sharpens the stakes rather than softening them. At the same G20 meeting where this was unfolding, Washington's tech adviser was pushing the opposite direction, promoting "technology-neutral" principles against AI-specific rules, with Zuckerberg and Musk both arguing publicly for fewer constraints. Your AI vendors are being asked hard, specific questions by one jurisdiction while lobbying their home government for a lighter touch in the other, a genuine divergence, not just noise, for any enterprise weighing exposure across both.

## SECTION 3: ENTERPRISE & INDUSTRY

### The Build-Versus-Buy Line Just Moved, and Europe's Largest Airline Already Picked a Side

McKinsey's Global AI Survey for 2026, published this week from interviews with 1,719 business leaders across 97 countries, found that 32% of organisations have now skipped buying at least one software product because agentic coding tools let them build it in-house instead, 41% in tech specifically. The more consequential number sits alongside it: the share of large enterprises (revenue above $1 billion) scaling AI agents in one or more business functions rose from 27% to 40% in a year. That's not pilot-stage experimentation. That's production infrastructure, expanding faster than most governance functions are resourced to track.

Ryanair's five-year Google Cloud partnership, confirmed 12 August, is what that shift looks like when a board acts on it deliberately rather than by accretion. Google Workspace and Gemini Enterprise are rolling out to all 35,000 employees, with DeepMind models (AlphaEvolve, WeatherNext) assigned to fleet maintenance scheduling and crew logistics. The detail that matters more than the AI capability itself: Ryanair already runs on AWS. This is a second hyperscaler, deliberately, not a replacement, a dual-cloud strategy built as much for resilience as for Gemini's agentic features, and a rare example of an airline putting agentic AI to work on operations rather than customer-facing chatbots.

Read together, the message for any board reviewing its own AI architecture is straightforward: agentic tools are now good enough that a third of organisations are choosing to build rather than buy, and 40% of large enterprises are scaling agents in production. Ryanair, running one of the thinnest operating margins in EMEA aviation, responded to that same pace by paying for a second hyperscaler rather than concentrating its AI bet with one. If a low-cost carrier judges single-vendor AI dependency risky enough to build around, it's worth asking why that same question isn't further up your agenda, particularly with this week's Big Story showing what happens when the "neutral" layer between you and your models changes owners.

## SECTION 4: EMEA LENS

### Utrecht and Dublin Show What Real AI Commitment Looks Like

Rabobank committed up to €2 billion over three years to AI and technology infrastructure in early August, bundled with broader IT and cloud modernisation, alongside €2.69 billion in first-half net profit. The bank was candid that the investment could ultimately mean job cuts, though it isn't forecasting large-scale layoffs. A Dutch systemic bank putting a hard number on AI transformation while being honest in the same announcement about what it costs in headcount. That combination, real capital plus an honest workforce caveat, is a more useful template for EMEA financial-services boards than most AI investment announcements this year, which tend to lead with the number and bury the tradeoff.

The other data point worth holding onto from this week sits in the Big Story: Stripe's international headquarters is in Dublin, and the OpenRouter acquisition was announced jointly from San Francisco and Dublin. That's less a coincidence than a pattern. Ireland has spent two decades building the tax, talent and regulatory infrastructure that makes it the default EU base for exactly this kind of global platform decision, and AI infrastructure deals are increasingly following the same route that payments and cloud took before them.

Rabobank and Dublin are two versions of the same signal: EMEA isn't only a destination for AI infrastructure decisions made elsewhere. The region actively shapes how money and infrastructure land, rather than just absorbing decisions made in San Francisco. Worth tracking which EMEA hub picks up the next one.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## WATCH LIST

| ~7 Sept 2026 | Anthropic's public S-1 prospectus, expected shortly after Labor Day (reported, unconfirmed) | ~3 |

| ~Oct 2026 | Anthropic's forecast IPO pricing/debut, targeting ~$2tn valuation (market forecast, unconfirmed) | ~52 |

| Coming weeks | Stripe–OpenRouter acquisition expected to close (customary closing conditions, per both companies) | Unconfirmed |

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## MY TAKE

The most interesting shift in enterprise AI isn’t that the models are getting better. It’s that the layers around them are becoming strategically important.

Stripe wants the gateway. Ryanair wants two hyperscalers. Rabobank is putting €2 billion behind its own capabilities. Brussels wants the authority to intervene when the infrastructure fails. Different stories, same underlying question: who controls the dependencies between your business and the intelligence it increasingly runs on?

For the last two years, choosing an AI model felt like the big decision. It probably isn’t anymore. Models will change, prices will fall, and today’s leader may not be next year’s. The more durable advantage will sit with companies that can move between them, build where it makes sense, buy where it doesn’t, and avoid designing themselves into somebody else’s stack.

That’s why the OpenRouter deal matters beyond its $7.5 billion price tag. The valuable territory in AI is moving outward from the model itself to the infrastructure that decides which model gets used, how it gets accessed, and who gets paid.

The next AI moat will not be the smartest model. It will be controlling the road that leads to all of them.

George

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

The AI Edge is published weekly by George Kakouras for informational purposes only and does not constitute legal, financial, or investment advice. Each edition covers enterprise AI deployment, strategy, and regulation for executives operating in EMEA.

© 2026 George Kakouras. All rights reserved.