31 June 2026

News

Anthropic’s product-market fit

Anthropic raised another round, collecting $65bn at a $965bn valuation, and reported that earlier this month it passed $47bn ‘run-rate annualised revenue’ (i.e. previous 28 days multiplied by 13), up from $30bn in April and $9bn at the end of last year. OpenAI hasn’t given a public number recently (which speaks for itself) - apparently it recently passed $30bn, but note that Anthropic is giving these numbers  gross, before payments to its inference partners, whereas OpenAI gives them net. A few people are still trying to persuade themselves that this revenue is all made-up, but that would be securities fraud, and we’ll get the audited GAAP numbers in the IPO filings soon enough. 

A few observations - first, agentic coding really works and really has product-market fit: companies are willing to pay hundreds of thousands of dollars per programmer for this, and collectively tech companies are paying Anthropic 3-4bn a month. And this is only the first generalised big use-case (though even at the end of last year, before agentic coding started working, OpenAI and Anthropic were at a combined $30bn run-rate). Second, the revenue model, pricing, and margin structure for this is very far from being worked out yet. However, it is clear that inference itself has positive gross margins.

Also... $965bn is more than the total market cap at issue of every single venture-backed IPO in the USA from 1995 to 2000. LINK

Tokenmania

The other side of the agentic coding revenue surge is a lot of bill-shock, as CFOs everywhere get a nasty surprise in their latest Claude Code bill. Last month, Uber said it had already spent its entire 2026 budget (unsurprising given that would have been set in late 2025, before this worked), and now that’s a trend: Amazon pulled an internal leaderboard that drove more use, and the WSJ points to a lot of other companies cracking down. On the other hand, Salesforce increased its budget buy 10x. LINK, AMAZON

Back at Uber, the COO said on a podcast that while the spend is high, it’s not always as easy to point to specific gains, which speaks to a broader challenge in early deployment: it’s easier to  say you’re using this a lot and love it that prove how many more features you’re shipping. LINK

More capex

Softbank announced a plan to spend ‘up to’ €75bn ($87bn) to build 5GW of AI data centres in France, leveraging ‘data sovereignty’ on one hand and France’s nuclear-generated electricity on the other. Of course, given that Softbank also announced ‘Project Stargate’ to build $500bn of infra with OpenAI at the beginning of last year and almost none of that seems to have happened, we should regard this with caution. LINK

Meanwhile, and perhaps more consequentially, Bloomberg reports that Bytedance is considering increasing its capex from $25bn last year to as much as $70bn in 2026 to invest in AI. Up until now, Chinese tech giants have been notable in their absence from the AI capex boom, partly because they can’t get enough Nvidia chips anyway, and hence the second penny to drop is that Bloomberg also says Bytedance is doing a server chip deal with Qualcomm. But, would that be covered by the same sanctions as Nvidia? LINK

Spending flow-throughs 

Dell shares jumped 40% after it said its server revenue is booming: quarterly revenue for the segment was $16.1bn, up from $1.9bn a year ago, and it expects to sell $60bn of AI servers in the full year, up from a previous estimate of $50bn. LINK

Meanwhile, the SaaS apocalypse is getting some nuance. Earlier this year, the entire software market was radically re-rated, not because anyone seriously (or anyone serious) believes people will vibe-code their own Stripe or SAP, but because it’s obvious that AI will lead to a lot more competition, competition in new forms, new pricing and margin environments, and a lot of general swirl as everything recalibrates, and there’s no real way to be sure which companies will suffer. The other side of this, though, is that some companies will get a lot of new business from the change, and this week Snowflake’s stock went up by 36% (and it’s a $50bn company) on good numbers and a thesis that companies migrating to AI will use Snowflake to get there. LINK

Meta does enterprise?

Last quarter, Meta raised its CY2026 capex outlook from $115-135bn to $125-$145bn, which is well over 50% of expected revenue, and it’s the only one of the big four hyperscalers that doesn’t have any kind of cloud or enterprise business to load onto that (let alone the coding tools that are working so well at Anthropic). That might change: apparently now it’s setting up an Enterprise Solutions group. Easier said than done (just ask GCP, which has been in a distant third place in cloud despite pretty much inventing cloud) - enterprise infra is a very different business. LINK

Meanwhile, at the AGM this week Mark Zuckerberg said that if Meta finds it’s overspent and has excess capacity, it would be willing to resell in competition with AWS, Azure, and GCP. Yes, xAI is doing that now with Anthropic, but that’s because it screwed up its models: if, on the other hand, it turns out that AI in general doesn’t need this much capacity, then everyone else will be dumping their capacity onto the market too. LINK

Nvidia AI PCs

Axios says that Microsoft will have a second try at ‘AI PCs’ this week with ARM devices using an Nvidia chip. There’s a whole platform development story here (and recall that Microsoft first tried ARM over a decade ago with the original Surface), but the end-game is that if you can run a good-enough small model on the user’s compute, it doesn’t have marginal cost. We should expect Apple to spend a lot of time talking about this at WWDC on 8 June. LINK

Blue Origin blows up on the pad

Congratulations to Jeff Bezos on one of the coolest explosions ever. Every FX lab in the world has this as a reference. (Seriously - bad luck, but no-one was hurt, and this is how we build stuff.) LINK, ANALYSIS

Ideas

This week’s viral paper on AI and employment suggests that the correlation in change is around work-from-home, not exposure to AI. LINK

Meanwhile, this survey says US companies think AI means they’ll need more juniors. LINK

A long interview with Boris Cherny, creator of Claude Code. A bit like Gmail, it was a side project that worked. LINK

Salesforce’s CTO on how they’re using agentic coding. LINK

Why Trump’s Green Card changes are moronic - a lobbying site generated automatically with Claude Code, with options for each kind of MAGA. (Via China Talk.) LINK

The reasons why it’s hard to sell AI to lawyers. LINK

Conversely, Kirkland & Ellis, the world’s highest-grossing corporate law firm, plans to spend $100m this year and $500m overall on building its own legal AI tools, to get away from depending on the same tools as everyone else. After all, do you want to give the client the same advice as anyone else? Maybe, but did banks build their own spreadsheet software? How much is this differentiation and how much is it just a condition of entry? LINK

And the FT on strategy consultants starting new firms that use AI to replace the associates at the bottom of the pyramid and avoid that cost-base. I’m not sure that the associates are really the barrier, nor what the client is paying for. LINK

A useful Bain primer on the state of agentic commerce. LINK

Also from Bain, a presentation in partnership with OpenAI on experiences in enterprise deployment. LINK

Apparently Apple is working on a feature for iPhones to detect if someone grabs them out of your hand and automatically lock into ‘stolen’ mode. LINK

The Pope published a 40k-word paper on AI and religion. LINK

Outside interests

Ferrari revealed its first fully electric car, designed in collaboration with Jony Ive’s LoveFrom design firm (also involving Marc Newson), and reaction to the styling is… mixed. (I have no strong opinion, though I would note that electric poses new weight and size challenges.) However, when LoveFrom showed off the interior a while ago, it all felt clever, elegant, and full of thoughtful ideas and appealing little touches… but it felt like an Apple interface. The aesthetic said Cupertino. It didn’t feel like a Ferrari. LINK

Posting this again: drowning doesn’t look like drowning, and half of children who drown do so within 25 yards of an adult, often without them realising what’s happening. So, you should read this one. LINK

Transcribing and translating ancient Greek papyri and inscriptions with LLMs. (There are hundreds of thousands of these that haven’t been read for a thousand or two thousand years, and now search and statistical analysis will be possible.) LINK

Data

An estimate that podcast revenue in 2025 was $9.3bn, with most of the growth coming from video. LINK

GDC thinks a third of US game industry workers were laid off in the last two years. LINK

Drexel University does an annual survey on job prospects for US graduates. LINK

Column

Pricing crunches and value captures

It seems pretty obvious, deterministically, that the current supply crunch in AI is temporary, in some form. Agentic coding has driven a vast increase in demand, but there's a trillion dollars of capex coming down the pipe, we have 100x efficiency gains every year, edge and open source are floating around, and there are too many players to get much price discipline. On the other hand, of course, we don't know what the next model will be to drive usage up (or down!), and we also don't know what the next use cases will be - all of this demand is only driven by AI coding and not by anything else. So, expect change! 

But stepping back from that, once we get to a different state, and to some kind of longer-term equilibrium of supply, demand, price, capacity, and capex, what might that look like? It seems to me that there are four or five different analogies, at least, to think about. 

The first one that we should mention, if only (mostly) to dismiss, is a comparison with the fibre bubble 25 years ago. As I think most people will understand, that was driven by people building out fibre capacity massively ahead of demand, whereas here demand is massively ahead of supply. 

A much more useful comparison, which I made in the presentation I published a few weeks ago, is to look at mobile data. Mobile data has marginal cost; it charges in bits, which like tokens are a pretty opaque measure that doesn’t map well to use-cases or value, and of course, it’s had several waves of demand surge and pricing shock. Over the last 15 years, mobile data traffic globally has risen by over 1500x, and telcos dealt with this on one hand by moving to flat-rate capped bundles (sometimes implicitly capped by ‘fair use’ or throttling) that align the price with capacity, and on the other by spending about $200 billion a year on capex. But the net result of that is that though mobile networks are a big business, with annual revenue of about a trillion dollars a year, the stocks have been flat for decades because all the value is further up the stack. They’re commodity infrastructure, and all the cool stuff is built by other people. 

This is, of course, the existential question for foundation model labs: will they capture value up the stack, or does that get separated out and built by other people?

This prompts a comparison with cloud: an infrastructure layer that is, to some extent, differentiated on service offering but doesn't capture much value up the stack either - there's no network effect. This contrasts with operating systems like Windows or iOS, which do have network effects and do capture a lot of value, but neither foundation models (for now) nor cloud work like that - you don’t typically buy a SaaS app based on which cloud it runs on (and it might use several for different things anyway). 

The problem in comparing with cloud or telecoms infrastructure, though, is that neither is really on the continuous escalating cost curve that drives LLMs at the moment. This is why I sometimes suggest a comparison with semiconductors: Rock’s Law says that the cost of a semiconductor chip fabrication plant doubles every four years. This dynamic is why the industry has gone from dozens of companies at the cutting edge to just one, TSMC. TSMC today has a de facto monopoly of current-generation chip technologies because of that escalating cost curve. So, it might be that LLMs follow the same path: there's no network effect and no fundamental differentiation, but it just gets so expensive and so difficult that you have a small number of winners and everyone else drops off the ladder. 

However, TSMC, even with that monopoly, sits only at the bottom of the stack - it doesn’t control what Apple or Meta do, and it had net income last year of $56bn where Apple alone, all of whose products are powered by TSMC, had more than double that. TSMC gets a pretty small slice of the overall tech hardware pie. 

Of course, none of these comparisons have predictive value - they don’t tell us how LLMs will play out, and of course, each of them played out differently to the others. But they do suggest that you can build something very big, important, difficult, complex, and expensive, powering things used by everyone on earth, without necessarily having much control or value capture over what people do with it. You can’t presume that having a foundation model by itself gives you leverage. There has to be some other dynamic that lets you reach up the stack and take control. Windows and iOS had that; TSMC, telcos and TSMC do not, and it’s not clear what that would be for an LLM.