26 October 2025
News
OpenAI tries a browser
Some things are so predictable that you could ask ChatGPT what will happen. Two weeks ago it tried an app platform, and this week it’s trying a browser. See this week’s column. LINK
Anthropic does TPUs
Anthropic joined the AI infra shuffle, doing a deal with Google for ‘tens of billions of dollars’ for access to up to a million of Google’s bespoke TPU AI chips and ‘well over a gigawatt’ of capacity in 2026. Google is an investor in Anthropic (at least 14%), but so is Amazon, and to date Anthropic has been using TPUs, Amazon’s Trainium and Nvidia GPUs, which sounds like a headache. LINK
Meta has layoffs in AI?!
There have been two stories for Meta AI in the last six months: Llama 4 was a fumble and it’s fallen badly behind; and Mark Zuckerberg is paying seven and eight figure salaries to hire more AI researchers. That makes it seem strange that this week the new head of new AI is cutting 600 people. The story seems to be that it’s the old AI unit being cut - there had been some suggestions that this was bloated with managers and politics, but some big names have apparently been cut even so. LINK
OpenAI pays ex-bankers to create proprietary training data
Bloomberg reports that OpenAI has a project paying over 100 former investment bankers $150/hour to create new training data that reflects the grunt work done in PowerPoint and Excel by junior bankers, so that ChatGPT can automate it. This is the tip of an iceberg - much (most?) professional work of the kind that people talk about LLMs automating isn’t on the Internet and isn’t in the training data that any LLM can use (the same goes for the data flows in enterprise software), so if you can get it for yourself you have an edge, and paying to create this de novo is a competitive advantage if you can afford it. Remember when people were surprised that Google paid people to drive down every street in the developed world to feed Google Maps? Note also, though, that Excel didn’t mean junior bankers got to work shorter hours. LINK
Conversely, Reddit is suing Perplexity (them again) and a set of little-known data scrapers for using its content as training data without permission. LINK
Oracle’s funding puzzles
A story from last week: Oracle held a big investor day to talk about its AI datacentre plans. Last year it forecast $104bn in its fiscal 2029 (meaning the 12 months to May 2029, since the company has an annoying year-end), but now it says that will be $185bn and $225bn in 2030 - up from $57bn in the 12 months to May 2025. In other words, it’s forecasting revenue will quadruple in five years, almost all driven by the new business building and renting out cloud infrastructure. This is backed by a $300bn contract with OpenAI plus a hope for at least $200bn more to come.
This came without much clarity on capex, which is admittedly difficult given how much uncertainty there is for so many moving parts, but it seems pretty clear that this implies capex comparable to the hyperscalers, which in turn would imply levels quite close to Oracle’s current revenue. The credit analysts are getting busy, but stepping back, there is a classic story to tell of a very cash-generative legacy business that was on the wrong side of first cloud and now AI, leveraging that cashflow to jump to the new thing. Silicon Valley is paying attention to Larry Ellison for the first time since the 90s. ORACLE, CAPEX?, DEBT?
Meta locks chatbots out of WhatsApp
Distribution seems like the new frontier for LLMs (browsers!), and WhatsApp is the global platform outside the USA and East Asia: this week Meta quietly changed its TOS to block third-party chatbots, including ChatGPT, from using the API to serve users inside the app. LINK
The week in AI
Google’s AI infrastructure is becoming more explicitly global, with the announcement of a $15bn investment in a 1GW datacentre and associated hub in India. LINK
Amazon is working hard on automating more warehouse tasks with robots, targeting 600k jobs and 30 cents cost saving per item shipped. (In other news, Pope is Catholic, bears…) LINK
Meta did an SPV with investor Blue Owl Capital whereby Meta will own and fund 20% of its planned 5GW, $27bn ‘Hyperion’ data centre and Blue Owl’s funds will take the rest. Meta is very cash-generative, but not as much as Microsoft or Alphabet, and they may well do deals like this too. LINK
Youtube is launching its tool to let you use AI to search for content using your face so that you can take copyright action, much as for a long time it’s let music and video right-holders search for use of their content. Yesterday’s AI is today’s automation. LINK
Remember VR?
Samsung launched the ‘Galaxy XR’, a mixed reality headset on Google’s platform with a spec somewhere between Meta’s Quest and Apple’s Vision Pro (it’s priced at $1800) and an experience that owes a lot to Apple and a lot to Gemini. It’s hard to say anything about this space that we didn’t say two and five years ago, but every now and then I get my Vision Pro out of the cupboard and power it up, and then put it away again. We are still a long way from the optics, price, and form factor that could make this a real mass-market consumer product, but what concerns me is that even if the Vision Pro was light and cheap, I’m still not sure that it would be useful. LINK
Ideas
The WSJ test-drives a ‘flying car’ (really, an electric helicopter). Much less mechanical complexity and much more software = cheaper and easy to fly. Use-cases still seem niche, though? Let’s see. LINK
The UK’s Channel 4 used an AI-generated presenter to anchor a programme about AI, not revealing this until the end of the show. LINK
Dwarkesh Patel’s 2 1/5-hour interview with Andrej Karpathy has gone around a lot lately. Dwarkesh himself is interesting as a newish player in the tech video podcast scene; Karpathy is a prominent researcher who’s now a free agent and hence free to comment (he coined the term ‘vibe coding’). The conversation is a mix of very technical engineering and very philosophical musing about the nature of intelligence, but the real takeaway is a good articulation of the view that this is all great and isn’t slowing down, but doesn’t always work that well yet outside of demos and won’t take us to AGI by itself. LINK
For comparison, here’s an Anthropic researcher talking about how everything is exponential and we’ll have human expert-level breadth by 2027. LINK
Outside interests
The Aga Khan collection of Indian, Persian, and Ottoman paintings. LINK
Oscar Wilde’s personal copy of Salomé. LINK
The European Food Safety Authority looked at the wave of studies raising concerns about ‘nano-plastics’ and found that most of them don’t really stand up to scrutiny. LINK
A study that claims you can see which US states use COBOL for their benefits systems in the economic statistics. LINK
Data
Marketplace Pulse’s report on Amazon Marketplace. LINK
There are now over 6,000 MCP servers. LINK
Outcome of a big UK NHS LLM trial. LINK
Emarketer points out that time spent on generative AI search appears to be about 3-4% of total search so far in the USA. This is all still very early. LINK
What’s the conversion rate for referral traffic from ChatGPT? This is a big question - some early data in this study suggests that so far it underperforms. LINK
Column
Browsers, bubbles and bundles
OpenAI is clearly trying everything and wants it yesterday. It needs its own infrastructure, without the legacy cash flow to pay for that itself, which gets us all the ‘circular revenue’ deals. It needs to build moats and features further up the stack, given that the models remain commodities for most general users. Some of that also means distribution, with Google now using the power of the default just as Microsoft did 25 years ago. That product and distribution work needs to deepen adoption for the hundreds of millions of people who today use it only once or twice a week. And it might be wondering if it has to do this before the bubble pops, if it is a bubble.
Deterministically, most important new technologies do seem to produce bubbles (smartphones did not, though 3G did a decade earlier), and if we’re not in a bubble now, we probably will be soon. No-one can call market timing reliably, though; we may be in 1999, or we may be in 1997. One feature of a bubble is that prudent investors swap their paper for hard assets (AOL), or at least lock in their gains (Mark Cuban), and Nvidia and OpenAI both seem to be doing that now. It’s not Jensen Huang’s first bubble, after all.
But another feature of 1997 was that we really didn’t know how this was going to work. It still wasn’t entirely clear that this would be the web browser for the next decade, but though Microsoft crowbarred its way into browsers, that turned out not to get it anything much. Meanwhile, Amazon was only a bookshop, no-one had heard of Larry Page or Sergey Brin, and Mark Zuckerberg was 13. All the value capture was ahead. But in January of that year, Rupert Murdoch bid $450m for Pointcast, the hottest company of the moment (ask your parents).
So, I think a lot of what we see from OpenAI now is just trying everything, and sometimes that seems so predictable that you could ask ChatGPT for a list. The old quote from Jim Barksdale is that there are two ways to make money: bundling and unbundling, and OpenAI is trying both, with Sora on one side and its app platform on the other. A browser fits both. It’s distribution, data collection and inventory for ads, and a way to embed itself into daily use rather than fighting to make itself into a new habit.
Then the problems start. Chrome was a huge upgrade from Internet Explorer, but Chrome today is already good, so what is Atlas offering? It doesn’t actually change web browsing itself at all - it adds some ChatGPT stuff around the sides, and a few obvious features that Chrome should (LLM-search of your history, say), but the web is still the web. That might be fine, or just too hard a problem - in the last 20 years there have been plenty of attempts at reinventing the browser and they all failed. That leaves a pretty limited reason to use this, though - indeed, it looks more like a browser toolbar (which in a sense is exactly what it is - a new shell on Chrome). Of course, Google used browser toolbars as distribution too, adding a Google search out to Internet Explorer back in the day.
Indeed, some of this looks like people doing what they know how to do. Sora is the Meta growth playbook, and there’s a device and a browser because devices and browsers are what you do. But how can you pull the whole thing inside out, and break through the elemental LLM problem that you have to think of what to ask? There’s very little sign of that kind of thinking in the Atlas launch - for heaven’s sake, the first two demos they offer are searching Google Docs and (sigh) writing code. This feels like people who said they needed a browser, so did a browser, not people challenging their comfort zones and pushing to new thinking. Maybe that’s why they paid all that money for Jonny Ive.