5 April 2026
News
TrillionAI
This really, really isn’t a venture newsletter or a capital markets newsletter, but sometimes that’s the macro story. This week OpenAI closed its latest funding round, raising $122bn at a $852bn valuation, and SpaceX filed for IPO, reportedly aiming at a valuation of $1.75tr (as I pointed out a few weeks ago, the largest ever IPO was Saudi Aramco in 2019, raising $29.4bn at a $1.7tr valuation). Both of these, of course, are bets on an essentially unknowable future. OPENAI, SPACEX
Meanwhile, the New York Times reports that Elon Musk is insisting that banks that want to be included in the IPO (given the size, they will need a lot of banks in the syndicate) need to buy subscriptions to the ‘Grok’ also-ran LLM that he merged into SpaceX earlier this year. Pay to play. LINK
OpenAI buys a media company
OpenAI bought TBPN, a tech video talkshow that gets views of 75-100k, and apparently had ad revenue of $5m last year and aimed at $30m this year: the price was apparently ‘low hundreds of millions’. I guess the new focus on focus starts next week.
Friendly, multi-hour video interview podcasts are a big thing in the valley, and when they work (and not all do) they can find a space for wide-ranging discussions between sycophancy on one side and performative attacks on the other. These are tech people talking to tech people for an audience of tech people, and bypassing news brands and news audiences. Interestingly, this one is small on YouTube and seems to be rather bigger on Twitter itself, which remains important for discussions of AI even as the steady stream of racist posts from the owner drove a large portion of the users away (including me). Last month one of the co-founders said they aim at a core of about 200k people in tech (this newsletter has about 170k subscribers, and my presentations get 3-400k downloads).
That said, while OpenAI might well need a better comms strategy, I don’t know why owning this particular channel could help it affect the public narrative. This isn’t the Washington Post (which disappointed Bezos), nor Twitter (which gave Musk a lot of influence with Trump), and there are many other voices. That said, I’d be happy to give a16z a friends and family discount - if they want to buy me out for $100 million, they know where to find me. LINK, PRICE, BACKGROUND
This week in AI
Fidji Simo, OpenAI’s head of product (and arguably de facto CEO, wth Sam Altman handling funding and research), is taking a medical leave of absence to manage a chronic illness. LINK
A chunk of Anthropic’s source code (but not model weights) leaked, due to human error in a software configuration. Mostly, this revealed a lot of inside info about product mechanics. LINK
A few weeks ago Ramp corporate spending data (which OpenAI dismissed) suggested that Anthropic is gaining enterprise share, and this week TechCrunch has credit card data which suggests the same for consumers. As I wrote a few weeks ago (see above), for most consumers, these models are commodities with very few switching costs. LINK
Remember Mistral, once the great French hope for AI (other than Yann, obviously)? It just raised $830m to build a data centre in Paris. LINK
Oracle layoffs
Oracle has a cash-generative legacy business that’s in long-term decline, and it’s borrowing heavily against those cashflows to buy its way into data centres instead. Now it’s doing layoffs in that legacy business as well to fund the DCs - certainly thousands, and possibly up to 30k (on a 160k base). LINK
Dating privacy
OKCupid settled with the FTC over a claim it gave personal data for 3m users to Clarifai, a facial recognition ID provider. Much as GDPR is a mess, this kind of thing is why it exists. LINK
Amazon satellites
Amazon did a deal with Delta to deploy its satellite internet, starting in 2028 (Starlink is already deploying on a bunch of airlines - getting a real 150 meg connection on Qatar is very nice). Meanwhile, Amazon is apparently also looking at buying Globalstar (market cap ~$10bn) to bulk up its effort to compete with Starlink. (NB Globalstar powers Apple’s emergency satellite connection on recent iPhones, and Apple has a 20% stake.) LINK
Remember Allbirds?
Allbirds was a flagship of the D2C boom a decade ago, and at one point it seemed like every Tesla Model S in Silicon Valley came with a pair. The valuation peaked at $4bn, but it was just bought for $39m, making it a symbol of the D2C bust as well. LINK
Robot IPOs
Unitree, maker of those viral humanoid robots, filed for IPO (PDF in Chinese). It’s profitable and it did about $170m of sales in the first 9m of last year. About half the revenue is from the humanoids, but most of the sales are ‘R&D’ (other researchers) and apparently most of the rest are for trade shows. Still very, very early, but it’s not very clear how far humanoid robots are useful without much more generalised AI than we have now. LINK
Ideas
This week’s viral AI paper uses satellite imaging to claim that greenfield construction of AI data centres can raise temperatures in the local area by ten degrees or more. Wow.
Except, if you actually read the paper, you discover that 1: the data set is from 2004 to 2024, and so it covers almost no actual ‘AI’ data centres and 2: the analysis does not control for other kinds of construction, nor for time of day. So, this looks at the increase in surface temperatures of replacing fields with data centres and the associated roads and parking lots, but does not compare that to construction of warehouses, office parks or anything else, which, of course, also have hotter surface temperatures than fields - remember walking on a road with bare feet in the summer? That means that this report is in principle only able to tell us that concrete and asphalt tend to have higher surface temperatures than grass, and that’s without even checking the maths.
There’s a teachable moment here. I’m not an expert in data centres, nor geospatial imaging. But I know that if you want to analyse something that starts in 2024 or maybe 2023, a data set that runs from 2004 to 2024 is a problem, and I know that if your thesis is that construction of X has a particular effect then you need to compare it to other kinds of construction. Of course, that depends on your real objective. LINK
The tech industry signals that a feature uses AI by adding purple sparkles (this is a draft ISO standard), but Microsoft also adds a little label that says ‘Copilot’. This has created a problem - Copilot is not actually a product, just a marketing term, rather like Watson at IBM, but the branding suggests that all of these ‘Copilots’ are one thing, that are integrated and connected and act in some coherent predictable way - they aren’t’ and they don’t. How many things are we talking about? Well, Tey Bannerman did the work and counted at least 78. (Surely this is an AI use case? ChatGPT says anything from 10 to 50+ depending on your definition, and Gemini says over 30. Let’s settle on 42?) LINK
A useful interview with Mark Prichard of P&G, (no longer) the world’s largest advertiser. LINK
Amazon’s pitch to advertisers for its Rufus chatbot. LINK
Using AL to analyse Chinese state propaganda at scale. LINK
OpenAI’s latest attempt at an app store has fizzled so far. LINK
Meta (like Google) continues to deploy generative AI into its ad and ranking systems, and publishes some interesting papers. LINK
Outside interests
Zeppelin lamp. LINK
Related: Zeppelin cloud cars. LINK
Nepalese police say they uncovered a plot to give Everest tourists food poisoning to force expensive helicopter evacuations. 4,700 people affected? LINK
NASA launched a test of its new moon rocket program (a sublimely pointless exercise), and the astronauts’ email crashed because they use Outlook. Douglas Adams is much missed. LINK
Data
EU data on AI and productivity, suggesting early effects are mostly on labour productivity rather than employment. LINK
YouTube’s revenue passed Disney, making it the world’s largest media company. LINK
SimilarWeb says ChatGPT is running ads on 0.1% of prompts in the USA. Still purely experimental. LINK
A new version of the ‘humanity’s last exam’ AI eval was just released: humans get 100% and the latest frontier AI models get less than 1%. All of these things are wrestling with the fact that we have no good theory of intelligence and no good theory for why LLMs work so well: we can make empirical measurements, but don’t really know what we’re measuring (one theory, or perhaps definition, is that we will have ‘AGI’ when we can’t think of a new benchmark). LINK
Column
What is an AI job?
As we all try to work out what AI will mean for industries, companies, and jobs, the obvious question is to ask which tasks can be automated and which jobs are made up of those tasks, and by how much. But it seems to me that a much more important question is how much those tasks are where the real value was, and how much they might be incidental to the real job to be done.
We had a very similar kind of question for the internet: the obvious question was how you still need physical distribution or a physical product anymore and how much that could be made digital, but again, the real question was whether that was where the value really was.
There were some markets where physical was so much the nature of the product or the business, and the internet was such a direct substitute, that those businesses just disappeared. At the other extreme, there were markets where it was very difficult to substitute online, or where the switch from physical to online distribution was peripheral or irrelevant (luxury goods, say, or airlines). However, I think the important part to this about today was the middle case, where the value and the product wasn’t about physical at all, but physical was how the business worked - most obviously, newspapers selling lumps of paper and record labels selling small pieces of plastic. No-one got into the music business or journalism because they wanted to work in specialised light manufacturing, but that was the basis of a business model, and it was the moat, and once it was possible to break that apart both industries came close to collapse.
Hence, for AI: there are some markets and some jobs that we will be able to automate very easily, and there will be some where AI is irrelevant or peripheral. Just as for the internet, sometimes this will be unexpected: it seemed obvious that airlines didn't face a competitive threat from the internet, but in 1998 you would have said exactly the same thing about hotels and taxis.
But then there’s the middle question: again, there is a broad category of industries and products where there is a thing that can be automated and a thing that the customer is actually buying, and those might not be quite the same thing, and may now be split apart. Right now this question is especially concentrated around professional services: accounting, law, consulting, some kinds of medicine, and of course software development.
But what is it exactly that you pay software developers or lawyers or consultants to do, and is that the same as the things you can automate?
Back to the internet: what is it exactly that you pay a retailer to do? Well, what do you pay a bookshop for? Sometimes you just want the best seller: sometimes you know exactly what book you want and you just want that book - that SKU - delivered to your home tomorrow. If you know the SKU, Amazon is almost always a better way to get it, and Amazon has peeled off a large part of that market. But sometimes you don't know what you want, and some book buying is a leisure activity, and so US physical bookshop sales have stabilised, at about half of what the peak was. Part of the ‘job to be done’ of a bookshop is to be the end-point to a logistics network and Amazon generally does that better, but part of it is about something else, around curation, taste, suggestion and serendipity, which are all things that Amazon is much less good at and indeed are inherently far harder to scale (at least until AI, perhaps).
So too with professional services. We’ve automated accounting in successive waves throughout the last century, through adding machines, punch cards, mainframes, Excel, ERPs, and SaaS, and yet the number of accountants grew every decade in the 20th century. So what is the ‘job to be done’ that you’re paying the accountants for? Is it to make the spreadsheet? The same for strategy consultants: when a big company hires Bain, BCG, or McKinsey, are they paying for a PowerPoint? Well, it depends. Sometimes. But often no, that’s not where the value is, any more than you were paying the accountants to count. The narrow criticism of the idea that LLMs will replace McKinsey is to say that the LLM can’t make a good deck, which is true, but transitory. The important question is how often the deck was really the product. The deck isn’t the hard part.
This, of course, also applies to software. Back in the 1980s, when Microsoft and IBM created a partnership to build a new OS, Microsoft was amazed to discover that IBM thought you could measure software productivity by counting lines of code. When you build a software company, writing the code isn't the hard part - the hard part is knowing what the code should be and who should be using it, and then getting them to use it - the search for product/market fit.
Then, sometimes there’s another step. As most people know, from 2000 to 2015 or so, recorded music revenue roughly halved in real terms, but since then, it's grown by 50%, back to about 75% of the number for 2000, due to streaming. I think that's because we changed the question. The first half of that period was asking “what happens if you don't have to buy a piece of plastic for $15 to get that one track you want?” In the second half, the question was “what if you could have all the music that there is for $15/month?”
I think this is a question about what happens when the price elasticity curve goes vertical. Whenever anyone asks about AI’s impact on jobs, it’s now a cliché to talk about the Jevons Paradox, which is really just applied price elasticity: if you make it cheaper to do something, do you do the same for less money or with fewer people, or do you do more work with the same number of people, or indeed more work with more people? This is what happened to accounting. But what happens if the cost goes to zero, or near zero? That’s what happened to newspapers: that cost base of people and trucks and manufacturing was your moat. But when that went to zero, that didn’t mean more newspapers: it completely changed the nature of the information environment or, for better and worse. And when we made it possible to get music for no marginal cost, the first step was ‘free’ music, but the second step was allthe music, for a flat monthly fee.
Back to AI: what things become possible now for the first time, because that would have needed millions of people, and now you can automate it? If you'd wanted to make an express train from London to Edinburgh in 1700, you couldn't have done it no matter how many horses you bought. You could have bought a million horses, and it still wouldn't have worked. So, once again, the narrow question is to ask what tasks make up that job, and whether those tasks can be automated, and the more important question is whether those tasks the reason you had a business, and whether that was what people were actually buying. But the more important question again might be to ask: what can you do if you aren't bound by needing that cost at all?