The model labs are finally reporting something that looks like quarterly revenue rather than "annualised run rate". The Journal says OpenAI did $5.7bn in the first quarter and $6.7bn in the second, while Anthropic went from $4.7bn to $11.6bn and overtook it for the first time, with a small operating profit against OpenAI's $12.3bn operating loss. Bloomberg puts Anthropic's run rate at more than $65bn by the end of July. All of this is one story: agentic coding found product-market fit faster and harder than anything else in AI so far, and Anthropic got there first.
Stripe bought OpenRouter, reportedly for somewhere between $7.5bn and $8bn. If models end up as interchangeable commodities for most tasks, which is one plausible path, then there is a layer in the stack for routing and brokerage: take a task, match it to the model with the right characteristics at the right price, clip a fee. Taking a percentage on everything that flows through is exactly what Stripe does. It runs a roughly 3% tax on a large slice of the internet, with a great deal of complexity hidden behind the friendly simplicity, and a few basis points on every AI call is a natural extension. What interests me more is that this is strategic M&A, which we have barely seen in AI. Everyone has been building bigger models and doing financing deals and tuck-ins. This is someone looking up the stack for the point of leverage. That is a more interesting idea than redefining the singularity as an uptick in company formation.
Nvidia had three deals in one week. It confirmed it will guarantee up to $105bn of lease and power payments for OpenAI's 8-gigawatt Ohio site, and put $1.5bn into SB Energy, the SoftBank subsidiary building it. Then it paid $6bn for a non-exclusive licence to Poolside's model-building infrastructure, made offers to 109 of its engineers, and took a $1bn stake at a $12bn pre-money valuation, which the investor letter insists is not an acquihire. The pattern underneath is this: Nvidia has a torrent of cash coming in from the hyperscalers, who get it from ads and software, and from the labs, who get it from investors. It is pumping that cash straight back into the market in three directions. Cheap capital for the labs, via guarantees, which weakens the hyperscalers' hold over them. Reinforcement of the GPU ecosystem, which hurts Google's TPUs and everyone else building their own silicon. And money for open models, which is bad for the closed labs and much less bad for the hyperscalers. Nvidia is using its customers' money to help its customers, to set some of them against others, and to make all of them more dependent on Nvidia.
Smaller AI notes. Google paid $10m in a bankruptcy auction for Spirit Airlines' internal data: a hundred million emails, five hundred million Teams messages, code, operational records, all de-identified, all training material. Apple accidentally shipped a support video for camera-equipped AirPods inside a macOS release candidate: look at a book, ask Siri to remember it. Meta is dealing with a backlash about what people do with camera glasses; Apple's privacy branding is stronger, and the form factor and battery suggest these are for recognising things rather than photographing them, let alone filming. Higgsfield, an AI video tool for creative professionals, raised $400m at $5.4bn and claims a $700m run rate. People pay for this stuff when it is verticalised properly.
Apple and Europe. There are YC founders who were not born when the App Store launched and the argument grinds on. Apple's response to regulators has been trench warfare: give a little ground, trim percentages in the right places, keep the friction high enough that most of the revenue stays. The new EU terms look like more of the same. The per-install fee is gone, the headline rates drop to 26% for in-app purchase, 20% for alternative payments, 15% for link-outs and 5% outside the App Store, and it is not obvious how many developers will actually move. The underlying problem has always been that Apple ranks Apple first, users second and developers third, while the alternative on offer ranks developers first, Apple second and users third. It has never been clear that Apple's rules stopped anything users would have wanted; most of this is an argument about who gets the loot-box money. Hard to have much sympathy for Tim Sweeney's campaign when Epic charges developers much the same. The more interesting development is that Germany's competition regulator made Apple rewrite its tracking-consent prompts, having found that the scary warnings it put on third-party ad tracking did not appear on its own. Neutral wording for everyone, from now on.
Unitree listed in Shanghai and the stock rose as much as 629% intraday before closing up 460%, around $50bn, which is a very 1999 kind of day. Humanoid robots are hot and this is one of the few clean ways to own the theme. The actual market is murkier. Morgan Stanley expects China to ship 50,000 units this year, but a large share of that demand is government procurement and research centres rather than validated commercial customers, and a lot of what happens there is humans teaching robots through telepresence. I remain on the perplexed side. Intelligence is the hard part, not walking, and most of the new use cases I can think of need a brain and a manipulator and no legs at all.
Drone delivery is the cleanest example. Do you want a humanoid to climb out of the UPS truck and walk the parcel to your door, the same job the same way but with robots, or do you want something to fly over and drop it on the porch? Uber partnered with Zipline, targeting a million drone deliveries a day by 2029, and Amazon says Prime Air will reach nearly 500 US towns by December.
Two retail notes. Walmart finally started accepting Apple Pay in the US, twelve years after it launched. The holdout was partly about owning the customer relationship through its own wallet and partly a very technical fight over basis points, which matters when you remember that American retailers pay roughly as much in card fees as they make in profit. And here at home, the government is back with its revived news bargaining scheme: if Google, Meta or TikTok decline to sign commercial deals with Australian publishers, they pay a levy on their local ad revenue instead. The idea that a click on Google, but not on any other site, to a newspaper, but not to any other site, should trigger a payment has never had any commercial logic. If you want to tax one group of companies to subsidise another, say so and do it.
Finally, two data points. Pew finds that 55% of Americans under 30 are now more concerned than excited about AI, up from 31% in 2021, which puts them ahead of the 30-to-64 cohorts on worry. That is the opposite of how attitudes to new technology normally run, and it is worth sitting with. And Etched, a chip startup founded by three Harvard dropouts, has raised close to $2bn, booked $1bn of orders, and says it took 44 days to get its chip back from TSMC and into production where that usually takes six months. I am not a semiconductor analyst, but every industry with long, boring, time-consuming steps that double as a barrier to entry is now in play.