21 June 2026
News
SaaS apocalypse
German PE firm Thoma Bravo had the second-biggest PE loss ever, writing off Medallia after buying it for $6.4bn in 2021. Medallia manages customer feedback and service, which of course is a really clear obvious thing to do much better and cheaper with AI (plus they probably overpaid). There will be more of these.
Only idiots think people will vibe-code their own SAP, but it looks like 1: software will get way more competitive and quite possibly lower-margin; 2: for some use-cases AI will replace existing solutions with a better cheaper technology (funny though it is to call AI cheap in June 2026); and 3: there will be lots of swirl as things get bundled and bundled, value gets peeled off and moved around (this is where you say ‘headless!’ and ‘harness!’ to sound clever), and the entire industry is held upside-down and shaken. Just as with any industry shift, including the emergence of SaaS itself, that means the death of some companies that don’t manage the transition, some of which look obvious (like this one - call centre companies are also heavily shorted right now) and some of which are inherently hard to to predict. Meanwhile, PE had 15 years where interest rates were low and you knew what effects the tech would have, and now neither of those statements are true. LINK
A UK social media ban?
The UK is looking at banning social media for people under 16, following Australia and a growing trend globally. Enforcement will be managed by Ofcom, the UK’s TMT super-regulator. Setting aside the frankly fuzzy evidence of harm, this is a tough thing to implement: you’re forcing tech companies to collect IDs, which they hate for privacy and security reasons, and we saw in Australia that it’s not that hard for children to get round this (especially if their parents let them use their devices). There are also definitional questions, which come down to your theory of harm - are you worried about content, or recommendation systems, or messaging? Do you ban SMS too, or Google? More pragmatically, this is a costless headline that sounds good, from a lame-duck prime minister with an instinct for state intervention to solve every problem and a pathological inability to make hard decisions, and here all the hard decisions will have to be taken by someone else. LINK
Meanwhile, India blocked Telegram because people used it to cheat in medical exams, which is just silly. LINK
Amazon’s AI cards
Amazon isn’t really on the radar for foundation models (though it hosts all of them on AWS), but AI will clearly transform the Amazon business. This week it had two big lists of announcements; on the one hand expanding product recommendation and discovery, and on the other talking about what this means for advertisers (reminder - with $70bn of ad sales in the last 12 months, Amazon is the world’s third-largest media-owner after Alphabet and Meta). ADVERTISERS, SHOPPING
RAM pricing
One effect of the surge in investment in AI is a crunch in memory supply: manufacturers are giving priority to high spec memory used in AI data centres, which means that in the last couple of months supply and hence pricing for the memory used in consumer electrics has been squeezed. This week Tim Cook said that Apple, which normally holds prices fixed over product lifecycles, will have to raise prices later in the year: this will probably be hundreds of dollars for many products. LINK
Snap Spectacles
I sometimes point out in presentations that all the stuff we were excited about before ChatGPT is still there. E-commerce is 20% of US retail (and 40% of UK non-food retail), TV is streamed (see the next item), the entire car industry is being overturned, and people are still working on smart glasses. This week Snap unveiled its latest attempt to stay relevant, $2195 ‘glasses’ that are really so bulky that we should probably still call them an HMD or headset. Lots of people have mocked just how big they are (although this is actually kind-of on-trend), but like Apple’s Vision Pro, this still all feels years from being ready. And of course, as we could have said 5 or 10 years ago, we don’t really know if people will want to do this even once all the engineering is solved. LINK
Fox buys Roku
This is one of those stories that you wish you’d said was obvious before rather than after it was announced. As TV shifts to streaming, the screen actually knows what you watch, and ads can be just as dynamic and personalised in TV shows as they are online, so how does that happens and who has the data? Roku is one answer, and Fox just bought it for $22bn. LINK
Midjourney pivots to medical
Midjourney pioneered prosumer AI image generation, but never really built a product around it and is now mostly left behind by multi-modal models from the big labs. There are many ways it could have reacted to that, but I don’t think anyone expected this: an attempt to disrupt medical imaging, selling a new full-body ultrasound scanner. The idea is that you can do frequent, cheap and (apparently) harmless full-body scans to look for changes. This is very far from my field, but the thesis is that we all have many masses but most of these are harmless, so the danger of doing an invasive biopsy on all of them to check for cancer is greater than the benefit, and now you can monitor all of them cheaply at no risk to see if any are growing. That sounds good to me, but 🤷🏻♂️. LINK
New cars
Uber continues to expand its AV partnerships. Meanwhile, Waymo paid $200m to buy the vehicle proving ground in Arizona that Apple built for its cancelled car project. UBER, WAYMO
On the other side of the shift, BMW issued a profit warning as its sales in China continue to shrink, like all foreign car car companies there. For all that people seem to hate the new Ferrari EV, I find it far easier to see a future for Ferrari in electric and automaton, as an eternal luxury brand, than for BMW or Mercedes. LINK
Shenanigans
Three amusing stories to group together. First, Substack is launching an ad product. The founders were very vocal in declaring that things like recommendation systems, news feeds and ads were stupid and evil, and now they’re doing all of them. LINK
Second, the $500 ‘Trump Phone’, which is sold as ‘American-made’, is of course an old HTC (in MAGA terms, 'CHINESE') from 2024 that’s been spray-painted gold. LINK
And finally, Amazon’s studio has dropped a nearly-complete and apparently very good movie about Sam Altman, which of course has no connection at all to Amazon’s deals with OpenAI. LINK
Ideas
More stories this week on how unhappy lots of people are at Meta. Zuck, Boz, repeated hire/layoff cycles, Manus, endless boring AI work… ALL-HAND, CULTURE
Life outside AI: Uber ended 2025 with $2bn in run-rate ad revenue, which would have been 4% of revenue but probably more like 10% of cashflow given the different margin structures. Now, naturally, it’s building an ad network. LINK
Meanwhile over at Instacart, ‘Advertising and other’ is now 28% of revenue and probably all of the net income. LINK
Bloomberg says Bytedance is Microsoft’s single biggest customer for AI, spending $1bn a year and mostly using OpenAI’s LLMs, where Microsoft’s deal means it can set its own terms. Lots of other Chinese customers too, apparently - no-one tell Anthropic. LINK
Why China dominates the robotics supply chain. LINK
Analyzing the global scale of GPS jamming. LINK
Outside interests
An American diplomat’s home-movies of Stalin’s Russia. LINK
Data
The Information has OpenAI Q1 numbers: $5.7bn revenue, $3.5bn cost of revenue (i.e. inference) for 39% gross margin, $8.6bn R&D (including costs of training new models), and $73bn cash on hand. LINK
The annual Reuters digital news report. LINK
Pew’s latest data on consumer AI use in the USA. LINK
Deloitte’s UK digital consumer survey data for 2026. LINK
And SensorTower has estimates of app and web use. LINK
Column
Voice as universal input
A bit over a decade ago, machine learning started working with image recognition. Slowly, we worked out that, really, this was pattern recognition, and then we spent a decade turning things into pattern recognition, and, indeed, turning things into image recognition. Imaging became a universal input, and companies like Sony thought a lot about image sensors whose output would never be seen by people, that might look for different things in different wavelengths, and could turn a new kind of unstructured data into structured data.
I wonder how much the same thing now happens to voice.
All of the attention in voice for generative AI is as UI - to talk to your computer and have your computer talk back to you. I think this is generally a pretty bad interface (though at this stage, who knows?). But the more interesting thing to me is unlocking the vast amount of material in all the conversations that happen inside organisations. In particular, a huge amount of the challenge of automating things with AI is how much is implicit and isn't captured in the training data, let alone anything structured. But if every call is now part of that, how much of that implicit knowledge becomes data? Every customer call and every internal call and discussion is transparent to the machine.
So, with the last wave of machine learning, you could say "find me customers who seem angry" or "are there call service agents who sound rude?". But now you can ask "have there been shifts in the kind of reasons people give for cancelling their account in the last six months?" You're no longer dependent on all of your staff to notice this on the call and type it into Salesforce in structured and consistent ways, and you can go back and ask questions and look for data that wasn't being thought about, and you can push to quite different levels of abstraction. What sort of concerns are coming up in one-on-ones, and why is our retention slipping? You have that data now, and you have systems that can manage those kinds of questions.
All of that might be a little bit far off, but there are much more tangible things happening right now. On one side, anyone doing consumer research used to be gated by the logistics of making the calls: you might be willing to pay for 200 interviews, but how many people do you have to make those calls, how long will it take, and how long will it take you to transcribe all of that and synthesise the result? Now all of those calls can be conducted by AI, consistently and at scale, and you can do it in an afternoon - the Jevons paradox in action. Last week I also mentioned an interesting example at the high end of this: automating expert network calls, where the AI might do a better job than whoever you'd roped in at the last minute to do 20 more calls.
Back in maybe 2015, I had a meeting with a big sports broadcaster, where they said that they'd been playing with machine learning, video recognition and speech recognition, and realised that they could index all their archives of their talent interviewing athletes at events, and so now they could search and scan and analyse it. And then they'd scratched their heads and said "Yes, but what would we do with that? Why would that be useful?" I don't know if it ever was useful (let me know if you know) - just because you can index everything, that doesn't mean it's useful, and maybe some of the use cases I've outlined above aren't the right ones. But I think the key thing is that we now have these enormous corpuses of information and understanding at scale, that were never really accessible to any kind of analysis before, and now all of that will be possible.