7 June 2026

News

Here come the IPOs

SpaceX filed for an IPO last week, and now Anthropic is out too, but unlike SpaceX it filed confidentially, so for now we can’t see the data. LINK

Meanwhile, with the upcoming giant IPOs of SpaceX, Anthropic and (presumably) OpenAI, there’s been debate about whether stock market indices should change their rules (around things like float size, profitability, and waiting period from listing) to include them. Index funds tracking these indices are now over half of the US market (it depends a bit on how you count this), so if the rules had been changed, this would have provided a base of support and put a sliver of all of our pensions into this. In the end, the Nasdaq says yes and S&P says no, and S&P is the important one. LINK

Alphabet capital 

Anthropic and OpenAI don’t have massive legacy cashflows to pay for AI capex and the giant tech incumbents do… but no longer enough: as I pointed out in my latest presentation, Meta, Microsoft and Alphabet all plan to spend close to half of total revenue on capex this year. They’ve been doing debt and SPVs since last year, but this week Alphabet startled everyone by raising equity - $85bn, including $10bn from Berkshire Hathaway. $30bn of this will be used to pay taxes on the vesting of stock options. People have been complaining that options aren’t properly allowed for in the P&L for a while now - the best expression of this is in fact from Warren Buffett, who said “If options aren't a form of compensation, what are they? If compensation isn't an expense, what is it?" LINK

Following this, the FT reported that Meta may also be planning to raise ‘tens of billions’. LINK

The SpaceX neocloud

One footnote in the SpaceX S1 was that in May this year Anthropic agreed to pay $1.25bn/month to rent the xAI data centre capacity - now this week, Google has rented more capacity for $920m/month, which seems to mean that well over half of the xAI compute is now rented out. 

On one level, this means SpaceX has ‘annualised revenue’ from renting AI compute of over $26bn, which is almost as much as OpenAI’s total revenue, and SpaceX’s entire revenue last year was only $18.7bn. And if you have unused infra and someone will pay for it, then well done. 

On the other hand, there’s a difference between a neocloud reselling Nvidia compute and a foundation model lab, and which are SpaceX investors being sold? We already knew that the entire xAI founding team has quit and Elon Musk admitted he’d have to rebuild, and this reflects that - it has no AI business today, and has nothing better to do with all that compute than to rent it out. LINK

Anthropic calls pause again, again

Anthropic released another long piece in which it 1: points out and describes the enormous breakthrough of agentic coding, 2: suggests that this points to a step change in the speed with which we can improve these models due to ‘recursive self-improvement’ (the model can suggest ways to make itself better and then build them), 3: observes that if you think these models can scale to AGI, that means we’ll get there faster, and 4: argues that there should be some kind of ’option’ for ‘the world’ to slow down such development.

Of course, the problems here haven’t changed since the last few times people said this: we don’t have a good theoretical understanding of how these models work or why and how they scale and so cannot predict future scaling except to say that it’s worked so far, and we have no good theoretical understanding of ‘intelligence’ (human or any other kind) and so cannot say when models scaling at a given rate would reach that. Equally, we lack any realistic and practical ‘option’ for ’the world’ to ‘slow down’ anyway, short of someone unilaterally firing ICBMs at data centres, which does not align very closely with generally accepted norms of regulatory best practice. 

The cynical view of all this is to say that this is just marketing and a push for regulatory capture by people who can’t possibly believe any of these claims. I don’t think that’s right - there are plenty of people very deep in the science who really do think this could scale to AGI, and there are plenty of others, just as deep in the science, who don’t. We don’t know. But meanwhile, Anthropic keeps working really hard to build it. LINK

Microsoft’s post-OpenAI reset

Microsoft held its developer conference this week, with three stories to note. First, the renegotiated deal/breakup with OpenAI means it’s free to build its own frontier models, and the quote is that “there are three relevant labs, and we want to be the fourth”. That’s a clear ambition, but I'm not sure that the dozen or so labs that are a few months behind the frontier are entirely irrelevant, especially given that the second story from Microsoft is to address soaring token usage (and hence bill shock) by running older/smaller ‘good enough’ models on the edge, where the user buys the compute and it’s a fixed cost. This will probably be part of Apple’s AI relaunch at WWDC on Monday, and Microsoft announced a new line of Surface laptops starting at $3k, with an NVIDIA SoC (instead of Intel) to run models locally, plus a development stack around that. Look for hardcore dev people to think about the intersection of developer tools, CUDA, and Windows here. Then, the third story is that Microsoft wants to be the everything shop: your models, our models, their models, edge, cloud, agentic platforms, an OpenClaw clone - we’ve got it! The thesis seems to be that this is a commodity tech layer where the hyperscalers will compete to provide the best platforms and tooling. That’s all fine up to a point, but important new technologies are rarely ‘the old thing but more.’ LINK, SURFACE, MODELS

Also, there was a new quantum chip. LINK

More Meta ad automation

Google and Meta’s ad revenue is surging as they push AI through their recommendation and optimisation systems (Meta was up over 30% last quarter). But the structural problem for SMB advertising (a large share for both and especially Meta) has always been mapping the sophistication and potential of the product against busy small business-people who aren’t ad-tech experts. This week Meta launched a new set of agentic tooling to try to automate all of this, building the campaigns for you. A while ago, Mark Zuckerberg said he wanted advertisers to be able to give Meta a business objective and a budget and have his systems do the rest - this is one step. LINK, EVENT

AI hacking

It turned out to be not just possible but really easy to get Meta’s AI chatbot to give you control of someone else’s Instagram, bypassing all security controls including 2FA: over 20k accounts were affected before Meta fixed this. LINK

Meanwhile, someone used Anthropic’s Opus 4.8 (not even Mythos) to spot a major bug in the Zcash crypto token: it had a $10bn market cap and promptly fell by a third. Someone once said that all cryptocurrencies are self-funded bug bounty schemes. LINK

AI versus the web

Google’s ‘IO’ event signalled a structural shift in search from links to answers, synthesised and summarised from everything published on the web. For publishers, that gives you no traffic, so you might want to block it, but Google didn’t let you block AI overviews etc without also blocking search. Now that’s changing: Google will let publishers opt out of AI summarisation features affecting search, driven by pressure from the CMA in the UK. GOOGLE, CMA

Codex for X follows Claude for X

Following Anthropic’s ‘Claude for X’ products (or templates, perhaps), OpenAI has launched ‘Codex for X’. As with ‘Claude for X’, a lot of the output comes as micro sites, and I hear from friends in professional services that this is replacing PowerPoint with extraordinary speed for tasks like ‘take this ten-page memo and turn it into an interactive status update.’ LINK

The week in AI

Trump’s AI regulation executive order was signed this week, much reduced in scope after a strong reaction from Silicon Valley: most of the important parts are optional. As ever, there’s such a big gap between what Trump says (or signs), what execution is then attempted, and how that works out, that there’s little one can predict at this stage. LINK

Continuing the chaos theme, the White House has picked up a suggestion, apparently originating from Sam Altman, that the US government should take strategic stakes in ‘AI’ companies. Um, isn’t that going to be most big tech companies? And also, um, isn’t that what MAGA people call ‘socialism’? LINK

JP Morgan worked out that 60% of US data centres planned for completion in 2027 haven’t started construction yet, mostly because of difficulties in getting power. LINK

Bloomberg says that China now has foreign travel restrictions on top AI talent at private Chinese AI companies, extending its reaction to the Manus deal. Not only is ‘Singapore-washing’ over - you can’t even get on a flight there. LINK

Apparently, Deepseek is in talks to raise $7.5bn at a $50-60bn valuation. Cheap models aren’t cheap anymore, and agentic is profitable but needs capital. LINK

AI trust & safety

Getting its backlash in early, Florida is suing OpenAI and Sam Altman personally over mass-shooters’ use of ChatGPT. Some of this is about guardrails or the lack of them, but some is also about the nature of an information tool - if you ask an AI system whether a shooting at a school would get more press than a shooting at a factory and it says “yes”, is that really ‘advice’ that should have been blocked? LINK

The return of autonomy 

The autonomy winter is over, with dozens of Chinese companies deploying things, and BYD is now making its own chips to power this, in rivalry with Huawei. Meanwhile, Uber continues its return to the space: Reuters thinks it’s invested $500m in Nuro. BYD, UBER

Ideas

Uber has somehow become everyone’s favourite case study for tokenmaxing, first spending its entire 2026 AI budget in Q1, then saying this wasn’t that useful, then saying it was, and now it’s capped employee usage of coding tools at $1500/person/month. Dara Khosrowshahi explains. DARA, CAPS

Bain has a new study out surveying CEOs, saying, in effect, that it’s a lot easier to ‘give everyone AI’ or ‘launch a project’ than to get everyone to use it, or to create the right incentives, cultural alignment, and workflow redesign to make it useful and successful. Of course, this was also the case when we deployed ERP, or CRM, or the web, or anything else. Bain is talking their book - these problems are why people hire strategy consultants! - but that doesn’t mean it isn’t true. If you want a new technology to transform your company, you can’t just put it on everyone’s desk. You do actually, well, have to go about and think about how to transform your company, and that’s a whole new business for designers and consultants as much as it is for engineers. LINK

This study, based on GitHub data, shows a substantial increase in developer activity resulting in a much lower increase in shipped product. Is this because AI is generating lots of busywork, or because it only accelerates one part of the build and deployment process while the others are unchanged? LINK

How Coca-Cola uses AI to manage marketing spend. LINK

Outside interests

The word ‘genocide’ has a fluid definition these days, but this long and painful report from the FT describes China’s ongoing effort to erase the Uighur people: a university’s Uighur-language library was shredded, 90% of children in some areas are in Mandarin-only boarding schools, there are state targets for sterilising Uighur women, and up to 2% of the entire population may be imprisoned. In one county, 4% of the entire population is in prison on ’terrorism’ charges. Many of those prisoners are used as forced labour, and they’re trucked to work all over China to bypass sanctions on ‘Uighur cotton’. Meanwhile, there are AI cameras everywhere, your phone really is listening to you, and the next step will be an LLM deciding if your conversation is a crime. There’s a trend now for tourist trips to China to ‘see the future’, and the future is robots, but it could also be things like this. LINK

Data

Headline: “Nearly 20% of young people use used AI for mental health advice” - almost as many as see a professional. Actual story: 20% have used it at any time for advice if they were feeling down in any way. Framing is everything. LINK

Doordash has an interesting new survey of US restaurant buying behaviour and AI. LINK

US sentiment has swung really strongly against AI data centres near them. This reminds me a bit of the great 5G panic a decade ago. LINK

Column

Improvised software

“For half of my jobs I tell people who use Excel to use a database, and the other half are the other way around” 

I think you could group all enterprise software into three categories. There are the big iron systems of record, like SAP, Oracle, Workday, and so on. There are dedicated vertical applications that take some specific workflow or problem and turn it into software: the typical big US corporation has 3 to 400 SaaS applications, and then a thousand (or many thousands of) other legacy apps, tools and databases that have accumulated over the last couple of decades. 

And then in the middle, there's a fuzzy intermediate space of email, shared folders, Google Sheets, network drives, Excel and macros, where people improvise the things that haven't been institutionalised, and codified as software. This is improvised software. 

And then, for any task, where does this fit? Is SAP built to solve this? Can you force SAP to do it? Should you give up, export a CSV and do it in Excel? Or do you buy a new piece of software?

If you're PWC and you hire thousands of graduates each year, you have software for this that you've bought or built. If you hire 5 graduates a year, then you’re either forcing Workday to do this (using Workday may be the ultimate AGI eval), or you're using email and a Google sheet, or perhaps someone built a Notion database. And as the number of graduate hires rise, at some point you might scratch your head and say “it's time to buy software”. 

This is a process of continuous bundling and unbundling and swirl, as people try to work out the right task and the right tool, and the right place to do this. 

Now we lay AI onto this, and people ask “does this mean more software or less software?” I think it's clear that the answer is ‘yes’. Do you do this task in a chatbot instead of wrestling it out of SAP? Yes. Does SAP have a bunch more flexible tools? Yes. Is it much easier now to do this with a CSV and ChatGPT rather than Excel, and now you don't need to build a vertical tool or buy a vertical tool? Can AI make you a beautiful new tool to do this? Yes. But will there be now be way more vertical tools that are more powerful and much cheaper, and address use cases that were uneconomic before, because now it's so cheap, and efficient to build software - and are all these tools maintained and insured and backed-up and considered by people who want to spend their lives thinking about what this tool should be? Yes, that too. 

I mentioned Notion, and, of course, Notion was part of that story before. Notion itself is a reason that you don't need to go out and buy dedicated tool until you do. A generation earlier, the same was true for Access. I used Access to build an ERP for my mother's publishing business in the early 1990s; my former colleague Steven Sinofsky did this for his father's retail business before he got a job at Microsoft in the early 80s. This has always been the flow.

The more important questions, it seems to me, are not about where and in which substrate you solve each task, but what entirely different kinds of tasks you can achieve. Narrowly, LLMs enable new features. Salesforce can draft you an email or build you a dashboard. But that’s doing the old thing more and better - new things start out doing the old things more and better, but then they do something else. 

So, what happens if you ask “go look at our app telemetry, and all of the AI-generated transcripts of every client call, review our competitors, and then suggest what we could do to improve our churn.” That would have been science fiction a few years ago, or a big analytics project that would take six months. But now it feels like that's a product question, and it's a different category of software, not just more and better versions of the software we already have.