14 September 2025

News

Oracle drops a bomb

We all knew that OpenAI wants to build its own infrastructure separate from Microsoft (remember the $100bn ‘Project Stargate?’), and we also knew that Oracle wants to do AI infra. But this week OpenAI signed a deal worth $300bn over five years, and Oracle’s market cap jumped by $230bn to $930bn (partly because there aren’t many leveraged AI stocks for the public markets to buy). Last week we saw some reports that OpenAI plans to burn $115bn by 2029 ($8bn this year), and deals like this are how. Sam Altman clearly wants to make OpenAI the next (or only) trillion-dollar platform company, and it’s growing fast, but this feels very bubbly. The question, though: is this 1997 bubbly or 1999 bubbly? LINK, BURN

The slow-motion Microsoft/OpenAI break-up

The other half of this story is the slow separation of OpenAI from Microsoft. I always wonder how much context to give when I write up this story, given that versions of it repeat every couple of weeks: OpenAI has a weird ownership structure, Microsoft has a weird investment that’s supposed to give it access to tech but kind of doesn’t, and they’re both looking over each other’s shoulders for a better option. This week: Microsoft will buy AI tools from Anthropic as well as from OpenAI. LINK

Meanwhile, the new capital structure is slowly emerging. LINK

App rankings

Google’s Gemini app is at the top of the US iPhone App Store chart, after ChatGPT had been at number one since some time in 2023. Perplexity, meanwhile, has climbed to 28 (Grok is at 62). Technology versus product versus distribution. LINK

Penske sues over AI Overviews

Penske Media (Rolling Stone, Billboard) is suing Google over AI overviews. This is a live topic, we’ll hear a lot more about this, and there are lots of different numbers floating around, so I downloaded the complaint to see what’s happened to Penske’s traffic. Oddly, the only specific number that Penske gives is that their search traffic to pages with affiliate links fell by a third in 2024. I wonder if the real issue is that Google cracked down on spammy affiliate pages last year? LINK

ASML does Mistral

ASML, the Dutch company with a monopoly on cutting-edge semiconductor manufacturing equipment, took a €1.3bn, 11% stake in Mistral, the French second-tier model lab. That seems very down-stream for a straightforward corporate investment: I wonder how much this is about industrial policy and politics. There are rumours that Apple is looking at buying Mistral, which prompts the question of whether France would allow that: is it more important to hold onto a European model (geopolitics), or build the European tech ecosystem with a great exit? (Remember how it blocked the DailyMotion deal.) LINK

The week in AI

Microsoft did a six-year $19bn deal with Nebius, a cloud GPUs-for-AI company that was spun out of Yandex. LINK

Anthropic added memory and file creation. LINK

After a couple of horror stories about some people getting much too attached to AI chatbot ‘companions’, and those conversations becoming mutually reinforcing spirals into delusion (including several suicides), Meta is changing its rules and the FTC is opening an investigation. LINK

Glasses are still on the way

Meta and Anduril have a joint project to work on AR glasses for the US military. Lots of history here. LINK

Amazon is also working on glasses. Most people in tech think that glasses might be the next device after phones, so you want at least to have the option to be ready to dive in. Plus, Amazon has a use case right now in its delivery drivers. LINK

Adland value-capture 

Amazon sold $60bn of ads in the last 12 months and is becoming a full-spectrum ad platform; Netflix wants some of that scale, so now you’ll be able to buy Netflix ads through Amazon’s broader platform. And Amazon already has Disney, Roku, NBCU, and a bunch of others. LINK

Starlink buys more spectrum

Starlink paid $17bn in cash and stock for a block of 1900MHz US spectrum from the struggling US satellite operator Echostar, to boost its direct-to-cell project. People get a bit over-excited by this, and talk about replacing cellular operators - Starlink’s service struggles to work indoors, and is many orders of magnitude away from having enough capacity to compete directly in dense urban areas (for reference, Starlink today has about 8k satellites, roughly half as many as there are cellular base stations in New York City alone, which is not a direct comparison but a relevant one). However, this is a great story for lower density rural, remote areas, boats, airlines and a bunch of other niches, including of course the military. LINK

iPhone day

Apple did its annual iPhone updates, iterating a very mature product category with the usual better cameras and faster chips. There’s also a new, much thinner ‘iPhone Air’, compressing all the chips into a thin bar at the top, which seems very obviously a precursor to Apple entering the folding phone segment next year (as usual, Apple has waited until it thinks the tech is ready). Perhaps more notably, AirPods will now be able to do live simultaneous translation - Apple can do AI too, sometimes. 

Stepping back, these new phones are a reminder that even if a large chunk of our computing really does move to AI, we’ll still need a device to use that AI, and AI doesn’t necessarily change very many of the reasons why so many people chose Apple devices. This is also the problem for any new form factor (c.f. the OpenAI/Jonny Ive tie-up). A chatbot might be an essential new kind of computing, but does it have all the other reasons you use your phone? No - ‘agentic’ won’t do all of them. LINK

Ideas

A startup called Inception Point is using AI to generate 3,000 podcast episodes a week. LINK

The Arab Spring ran on social media - Nepal’s protests run on Discord? LINK

WPP’s GroupM asked a bunch of people about where AI and advertising might go in the next few years. (I am quoted, amongst people who know a lot more about this.) LINK

Analysing a leak of internal documents from China’s ‘Great Firewall’. LINK

Apple has a new local wireless tech for audio. LINK

Outside interests

David Lynch’s compound in LA is for sale. A Frank Lloyd Wright house, and a lot else. LINK

Data

Atlantic’s annual report on Latam tech. LINK

Column

Is Apple getting Microsofted?

30 years ago, the web meant that Microsoft lost the development environment: the tech industry moved to making websites and web services instead of Win32 applications. Microsoft tried making services of its own (MSN, Expedia) but those didn’t really need to be part of Microsoft, and didn’t give the same kind of dominance. But the web also gave far more people a reason to want a PC, and for 95% of them that meant a Windows PC. So Microsoft lost its dominance, and could no longer set the agenda - no one cared what Windows did. But, it also became a much bigger, richer company: there were less than 100 million PCs on Earth, and well over a billion today. (Arguably, Apple also benefited from this - if the purpose of the computer was to use the web, it mattered less that the Mac had so much less software than Windows.)

Apple doesn’t have a competitor to YouTube, Amazon, Instagram or Uber. It sells thousand-dollar devices at a 40%+ gross margin and those devices are a platform that people use to watch YouTube, and that’s a great business. So if people use those devices to use ChatGPT (or whatever comes after it), why is that different? 

There are maybe two or three possible answers. The extreme view is that LLM chatbots will become so powerful that they could make it possible to replace the smartphone entirely, instead pulling all services into one unified UX, and interacting with third-party services through agents. This was the pitch from Humane and the Rabbit phone, and it sounds superficially clever, until you realise that in this model a company (Tinder, TikTok, JetBlue) can’t actually control its own user experience, and indeed don’t really know what the user is seeing, only what this intermediate LLM might decide to show them in some kind of ‘generative interface’. That might work for relatively generic experiences like e-commerce, but it seems much harder for anything innovative. How would you tell ChatGPT ‘OK, now the user is going to use something we call ‘Snapchat’ and this is how we want you to present it to them…”

More interesting, right now, is how far LLMs can be used to change the experience of the smartphone itself, integrating across many aspects of the device. This, of course, is what Apple showed last year in a reimagined ‘Siri’ and then tuned out not to be able to ship: we’ll see. Meanwhile, Google outside China and many Chinese companies inside China are already integrating all sorts of new LLM-powered features across Android. Apple has some of these too, up to a point, but it’s behind. If it never caught up, we’d all switch to Android, but that’s a big if. We had this conversation about machine learning a decade ago (when ML was called AI), and it turned out that Apple could build those features well enough. Indeed, many of Google’s own LLM-powered features are embedded inside its iPhone apps, or even inside Apple’siPhone apps - for example, Google’s AI visual looks are embedded inside iOS 26. That reflects a broader complexity: for Google and now OpenAI, the purpose of the device is distribution, not selling devices, and Apple can sell distribution too - Google is paying $20bn for search distribution, after all. This isn’t a simple split. 

Meanwhile, people buy iPhones because of a software experience that might change with AI, but they also buy them because of the hardware design, the quality of the components, and the experience that comes from Apple’s still industry-leading chips. The iPhone Air reminds us of that - if you’re going to hold this device in your hand all day, many people prefer a nicer device and think Apple’s device is nicer.  Phones are not commodities even before you turn them on. 

This also presumes, of course, that Apple can’t get it together. Things don’t look great right now… and then Apple casually mentions that the AirPods can do live translation. It might never be leading, but, like Apple Maps, it might become good enough. 

The real discontinuity, I suspect, might be if and when we change both the hardware and the software, which means glasses. This was also the second step for Microsoft: it lost the developer environment with the web and then lost the device a decade later with smartphones. If and when we get to optics that enable practical mass-market star glasses displays, then that might mean a real replacement for the smartphone, and one where AI drives far more of the experience. We still don’t really know how close that is, and we don’t really know if it would be a replacement for the smartphones, or just an accessory like the watch. Also like the watch, though, this is both new and context-aware software and elegant and miniaturised hardware - which Apple remains the best in the industry at combining.