Stripe buys OpenRouter
As rumoured, Stripe bought OpenRouter, reportedly for about $8bn. Hmm. It’s clear that one possible pathway for AI is that models are mostly commodities and interchangeable for most tasks. If that happens, we could very easily have a layer in the stack filled by routing and brokerage, perhaps taking real-time bids on tasks, matching them against models with the right characteristics, and of course taking a fee. That’s OpenRouter (though you’d also need evaluations of some kind.), and taking a percentage on routing, of course, is what Stripe does - it’s a 3% tax on half the internet (with an awful lot of hidden fees behind the friendly simplicity). Taking a few basis points on every AI transaction by making everything easy would seem like a cultural fit.
Stepping back, it’s also interesting to think about this as a piece of strategic M&A, because we haven’t seen much of that yet. Everyone wants to build a bigger, faster model, and there are lots of financing deals and plenty of tuck-ins, but this is an example of someone looking further up the stack, and thinking about the points of leverage and value capture beyond this month’s product roadmap. That’s more interesting than redefining the singularity as an uptick in company creation. LINK, SINGULARITY
Real numbers?
The model labs are starting to shift from ‘annualised run rate’ to real quarterly numbers: the WSJ says OpenAI did $5.7bn in Q1 and $6.7bn in Q2: Anthropic did $4.8bn in Q1 and $11.6bn in Q2. Meanwhile Bloomberg says Anthropic’s annualised revenue was ‘more than $65bn’ by the end of July. These numbers, obviously, reflect the explosive product-market fit of agentic coding, which Anthropic got to first. REVENUE, RUNRATE
The Bank of Jensen
There were no less than three deals from Nvidia in the news this week. First, Nvidia confirmed (as previously reported) that it will guarantee $105bn towards an OpenAI/SoftBank data centre in Ohio. LINK
Second, as part of that deal Nvidia will invest $1.5bn in SB Energy, the power provider to the project, while the Information reports that Nvidia will also invest $3bn in Lacium, which is providing power for an OpenAI/Oracle data center project in Texas. SB ENERGY, LACIUM
Third, Nvidia is doing a $7bn deal with Poolside, a US startup building open models. Nvidia will buy a $1bn stake at a $12bn pre-money valuation, and then license the tech and hire most of the engineers for a further $6bn (so, an acquihire). After all that I hope their bankers can get some sleep. LINK
The underlying story is that Nvidia has a torrent of cash, from the hyperscalers (who get if from ads and enterprise software, and from model labs) and from the model labs (who get it from investors), and it’s pumping that cash back into market, by enabling cheaper capital for the model labs (with all these guarantees), which is bad for the hyper-scalers, by strengthening the Nvidia GPU ecosystem, which is bad for Google’s TPU business and everyone else who wants to build their own AI chips, and by pushing into open models, which are bad for the model labs, but much less bad for the hyperscalers. In other words, Nvidia is using its customers’ money to help its customers, but also to build up some against others, and to make them all more dependent.
The week in AI
Today in training data: Google paid $10m for a chunk of business data from the bankrupt airline Spirit - 100m emails, 500m Teams chats and a lot more besides. LINK
Apple accidentally leaked a support video for camera-equipped AirPods, apparently launching imminently, showing them doing a visual lookup (‘where can I buy this book?’). Meta has a backlash around people doing sleazy things with its camera-equipped glasses: Apple has strong privacy branding, but given the form factor and battery constraints I’d also suspect that these won’t be aimed at taking pictures in the same way, let alone video? LINK
Following the Hugging Face hack fiasco, OpenAI is keen to signal caution: new monitoring systems for its development process will have a 20% compute overhead (note of course that things like this could become a barrier to entry for smaller labs). LINK
Higgfield, an AI video generation tool aimed at creative pros, raised $400m on a $5.4bn valuation and claims a $700m revenue run-rate. Yes, this stuff is really useful and people are willing to pay, when it’s sufficiently verticalised. LINK
Speed-running the lessons of social media (and full of ex-Meta employees), OpenAI launched a Teen account product. LINK
Apple’s EU lawfare
There are founders at YC who weren’t born when Apple launched the App Store, and yet the arguments still drift on: at this stage anyone paying attention knows all of the issues on each side, and I won’t bother repeating them yet again. Apple’s model has come under increased pressure from regulators and court cases in the last couple of years, and its response has been trench warfare: giving just enough ground, cutting percentages just enough in just the right places while keeping the friction just high enough, that it's hung onto most of the revenue. This week it announced a new deal with the EU that looks much the same: the headline percentages have dropped, but it's not clear how many developers would actually move.
The underlying issue, I think, is that on this issue, as in everything, Apple's policies put Apple first, users second, and developers third, whereas is the alternative position would simply put developers first, Apple second and users third. It's never been clear that Apple's policies were actually precluding things that would be good for users, as opposed to just arguing about the division of spoils from loot boxes. Certainly it’s hard to have much sympathy for Epic Games CEO Tim Sweeney’s quixotic campaign against Apple when Epic charges much the same commissions itself. LINK, ANALYSIS
A more interesting development, I think, is the German regulator (echoing the UK’s Ofcom) pointing out that Apple’s required disclosures on ads are grossly hypocritical, putting scary languages on third party iOS ad disclosures that don’t apply to its own ad targeting. Apple will now have to ‘align much more closely’ the first party and third party prompts and use ‘more neutral’ language. LINK
Unitree!
Unitree IPOed this week, and the stock popped by nearly 630% at one point (which is very 1999), taking it to a $50bn valuation. Humanoid robots are very very hot and this is one of the only clean ways to invest. However, the real market is a lot less clear - the FT points out that the majority of shipments (25-50k units this year, Morgan Stanley estimates) are going to Chinese government ‘research centres’ - mostly local government boondoggles where humans ‘teach’ the robots through telepresence. I am on the ‘perplexed’ side of the argument on humanoid robots. Intelligence is the hard part, not walking - how many new potential use cases need intelligence and a manipulator but don’t need legs at all? IPO, GOVERNMENT
Drone delivery
Drone parcel delivery is the perfect example: do you want a humanoid to get out of the UPS truck and carry the parcel to your door, doing exactly the same thing in the same way ‘but with robots!’, or is it better for a drone to fly overhead and drop it on your porch? This week Uber partnered with Zipline for local delivery (which is now half of Uber’s revenue), while Amazon’s ‘Prime Air’ is expanding service to 500 US towns and claims to be doing ‘thousands’ of deliveries daily. UBER, AMAZON
Hopefully, not too many of them are being dumped in the pool, as in this viral video. LINK
In other news
Walmart finally bit the bullet and started accepting Apple Pay in the USA. The delay was partly about wanting customer ownership with its own wallet, and partly a lot of very technical arguments about basis points on low-margin transactions (reminder that in the US overall, retailers pay roughly as much in credit card fees as they make in profits). LINK
Apparently, Apple cut several hundred people working on content for the Vision Pro (Remember that? I have one in a drawer. I take it out for a few minutes every couple of months.) LINK
Australia is revisiting the amazingly silly idea that if you click a link on Google or Meta, but not on any other websites, and that link goes to a newspaper, but not to any other website, then Google and Meta should pay the newspaper a fee. The new idea is that if they don’t make a ‘commercial agreement’ (that has no commercial reality), then they’ll pay a tax of 5% of their ad revenue. As I’ve said before, if you want to tax one group of companies to subsidise another, you should be honest and do that. LINK
|
