19 April 2026

News

Anthropic does surge pricing

With software engineers suddenly using thousands of times more tokens for agentic coding,    Anthropic is having capacity problems, with outages increasing, users complaining of suddenly limited responses, OpenClaw being cut off, and a succession of price increases. It was always clear that some of the pricing plans were cheaper than the real cost of heavy users, and Anthropic has invested least in capacity of all of the big labs (partly reflecting the lack of a consumer story), but now, with agentic coding finding massive product-market fit, that’s become a real problem. There are lots of long-term discussions about where this will end up, with the extreme case arguing that token budgets will rise to match headcount budgets, which is all like trying to forecast internet use in the late 1990s, but it seems to me that a good comparison is mobile data in the late 2000s, when the iPhone’s flat rate data swamped AT&T’s network. The natural end-state of all these situations is a segmented range of flat rate bundles. LINK

Allbird AI

Meanwhile, after announcing last week that it would sell the remains of the once-buzzy shoe business for pennies, this week Allbirds (effectively just a cash shell) announced that it would now do something  vague in AI data centres, and the stock went up 5x. That’s the kind of thing that happens in bubbles, but it also reflects a basic shortage of ways for public markets investors to get a clean investment in AI. You can short existing software companies, and buy Nvidia (and to some extent TSMC and ASML) and random Japanese speciality ceramics companies, but what else? LINK

Meta Manus

China’s reaction to Meta’s acquisition of Manus (agentic AI platform) seems to be escalating: the FT reports that it looks like orders came from the very top to crack down, with multiple ministries ordered to take an interest. Moving your co-founders to Singapore to get around Chinese limits on foreign acquisitions might not have been as clever as it seemed, especially in such a strategic sector. Hence, China is only investable if you can get a domestic exit? Would that matter for Beijing - is there a shortage of capital for AI? LINK

The week in AI

OpenAI continues to build out its adtech stack. LINK

Google has joined Anthropic and OpenAI with a desktop Mac app. The shift from the web to local apps is partly driven by practical go-to-market considerations (get your own icon on the dock), but rather more by the scope to interact with local apps and files - using the operating systems as an actual operating system, not just a place to run browser tabs. Meanwhile, it’s also notable there is a Mac app but no Windows app. LINK

The US plans to require data centres to report their energy use. LINK

Back in January, when Elon Musk tried to juice demand for Grok by letting it produce deepfake porn, Apple was concerned enough to threaten to pull the app from the App Store if Grok didn’t add more safeguards. LINK

The Information reports that Apple has sent several hundred developers on the Siri team on a ‘bootcamp’ to learn how to code with AI. I would have guessed that if anyone at Apple considered it an essential part of their job to know everything about this stuff already, it would be the Siri team, but apparently not. LINK

Uber goes back to robots 

Uber got out of robots back in 2018, realising that while ML got us 90% of the way, the next 10% was 50% of the work, and meanwhile Uber was spending both too much money and not enough (also, remember the Levandowski lawsuit). But now physical AI is hot again, Waymo is kind of working, and Uber is back: the FT calculates that it’s committed $10bn over the next few years towards vehicle purchases and equity stakes across more than a dozen different tech companies (it had almost $10bn of FCF last year). That’s a portfolio model, rather than the previous attempt to build the whole thing itself. LINK

Amazon buys Globalstar

As rumoured a couple of weeks ago, Amazon has bought the LEO satellite operator Globalstar to bulk up its satellite Internet project. Apple (which has a stake in) remains an anchor customer for the emergency connectivity in iPhones. LINK

Meanwhile, Jeff Bezos’s Blue Origin rocket company had a semi-successful launch. LINK

Ideas

A long and fascinating interview with Nvidia’s Jensen Huang, conducted by an AI influencer, Dwarkesh Patel. Frankly, he’s a bit out of his depth, throwing challenging questions at Jensen but not really able to handle tough answers: it reminds me a bit of an overconfident undergraduate trying to argue with their professor. That said, we get a lot of good material from Jensen - in general, his elucidation of how Nvidia’s model differs from other chip companies, and in particular (which is where Dwarkesh gets into trouble), on his view on US export controls. Jensen argues that Chinese chips aren’t as good as Nvidia’s (true), but China has an abundance of cheap energy (also true) and so Chinese labs can just use way more chips to get the same result. Hence, withholding Nvidia chips doesn’t slow down Chinese AI - it just cuts off the market for US companies and the US ecosystem and pushes China to build its own tech stack instead. That’s fine as far as it goes - where he runs into trouble (and where Dwarkesh didn’t push him) is that he says we should stop China from using that capacity to build Mythos-style cyber capabilities, by, um, asking them nicely not to be mean to us. We already know that doesn’t work, but the real answer might be ‘we can’t’. LINK

Pulling up the ladder - two important pieces about high-tech manufacturing. First, the FT argues that after the ‘China shock’ of cheap low-value manufacturing, there’s now a growing second China shock of high-value, high-tech manufacturing, where the same model of ferocious, Darwinian competition, backed by subsidies and cheap energy, produces a handful of very efficient and capable winners in each space, plus a lot of overcapacity, that then moves to exports. Second, Bloomberg says that Chinese export controls in those high-tech industries are crippling India’s attempt to build its own tech manufacturing base. INDIA, CHINA

The debate within China about AI and data privacy. LINK

Two analyses of the report that OpenAI is hoping for $100bn of ad revenue by 2030. The current global ad market, excluding China, is close to$1tr, but new things create new TAMs. LINK 1, LINK 2

Someone leaked an internal memo from OpenAI talking to employees about Anthropic. Mostly predictable, but note the confirmation that OpenAI is reporting revenue net and Anthropic is reporting gross, which means their public revenue statements are not comparable (though the growth clearly is). LINK

The UK’s AI Security Institute assesses Anthropic’s Mythos impact on cyber. LINK

OpenAI should buy Snap? LINK

Email and Slack archives from dead companies are the new source of AI training data. LINK

Mondelez on how they’re planning for agentic commerce. Mostly, best practice for SEO. LINK

How accurate is AI health advice? It depends a lot on small differences in precisely how you ask the question. LINK

Outside interests

The album of tributary peoples. LINK

Data

The 2026 Stanford AI report: a huge number of charts on the state of AI. LINK

Newzoo’s latest games industry report. LINK

Emarketer forecasts that Meta’s ad revenue will overtake Google’s this year (but only by subtracting Google’s TAC). LINK

Google’s latest ad safety report shows a surge in attempts at placing scam ads using AI. LINK

Gallup’s latest survey of US public opinion, like many other surveys, shows rising worry around ‘AI’. It’s almost as though the heads of two of the big labs have spent the last three years trying to get regulatory capture by saying loudly that people should worry. LINK

Bain survey data on US consumer use of generative AI for search. LINK

Column

What’s the TAM for AI?

Way back in 2014 an NYU finance professor got a lot of attention with a piece arguing that Uber was massively overvalued, on the basis that the global ‘taxi and car service market’ was worth $Xbn and Uber would get X% share at Z% margin. He estimated that the total market in 2024 would be 183bn and Uber would have 10%: in fact, Uber’s gross bookings for mobility in 2024 were $83bn, and total gross bookings were $162bn - off by almost 10x. 

This wasn’t just wrong in hindsight. As a few people pointed out at the time, it was based on a fundamentally flawed premise -  sometimes a new thing can just take share of the existing market, but often it finds an entirely new market. 

What does that mean, though, if we’re investing half a trillion dollars a year and growing in AI data centres? The simple answer is that demand for tokens is far ahead of supply, and that’s what drives the investment. The more complex might be that demand for tokens is ‘infinite’ in the same way that demand for bandwidth was, and that in a few years, token demand, token supply, and capex (plus compute in the edge) will get into equilibrium just as they did for telcos, which have spent (say) $300bn a year on infrastructure for 25 years to give us more and more bandwidth. 

But of course, the real question is how much money people will save or make by buying tokens, or new software powered by tokens, in which industries, doing what? Narrowly, you can look at the size of the current software industry itself and wonder how much that might be replaced or accelerated. You can wonder what it’s worth to a law firm or a bank to replace some of their associates with machines that work 10x as fast and never get tired. 

Then, of course, you ask about the Jevons paradox , which is really just applied price elasticity - if you make it much faster and cheaper to do something, do you do the same for less money, or more for the same money, or indeed more for more money, since the ROI has changed? The answer could be any of these - it depends on the industry. 

But looking back to Uber, where does the new thing not just unlock new demand but create an entirely new thing? 

Conversely, where was the cost base that just collapsed the moat that defended your entire industry? There’s an old joke that newspapers thought the internet would be great, because they’d save on printing, not realising that the cost of printing was what gave them their monopolies. 

Taking that one step further, though, what happens when something that used to be really expensive becomes free? What if it was impossibly expensive? Imagine you lived in 1750 and you wanted to build a ‘train’ from London to Scotland? You could buy every horse in England and still not make it work. 

Here, I think it’s intriguing to think about the last 25 years of the recorded music industry. As most people know, from 2000 to 2015 recorded music industry revenue roughly halved, but since then it’s increased by half, taking it back to about 75% of where it was in 2000 (all adjusted for inflation), and that increase is entirely due to streaming. I think you could see this as posing two different questions: in the first 15 years the question was “what if I don’t need to spend $15 to get that song?” But since then, for the last ten years the question instead has been ”what if I can get every song there has ever been for $15 a month?” Before the internet, that question would have been impossible. 

The best summary of all of this might be Yogi Berra: “prediction is hard, especially about the future.“ You can construct plenty of elegant frameworks to try to predict where this will go, and some of them will even be directionally correct, but imagine how you’d have done if you’d tried that exercise in 1997. Would you have realised that the TAM for advertising included retail rents? The real constant is that the TAM is almost always bigger than it looks, because it covers things that you’re not analysing at all.