19 October 2025
News
OpenAI’s full-stack assault continues
I wrote last week that OpenAI is trying to become a full-stack platform, swapping stock and its position in the hype cycle for markers in everything from data centres to ads to apps, and now, as had been rumoured for a while, it’s also planning to create its own AI accelerator chips, partnering with Broadcom. This is a small and very expensive club, in which Alphabet has a long lead with TPUs and Amazon, Meta, and Microsoft have been trying to catch up, partly to offset Nvidia and partly just to optimise for their own workflows. (Apple might have a story here with its own market-leading chip team and ‘Private Cloud Compute’). Naturally, given where we are in the cycle, this comes with huge numbers: OpenAI will deploy ‘10GW’ of accelerators, in addition to the tens of GWs of deals already announced with Nvidia and AMD. The funny thing is that, theoretically, OpenAI is getting cash from Nvidia, which comes from OpenAI’s rivals, and using it in part to build up Nvidia’s rivals. LINK
OpenAI expands shopping to Walmart
Where Amazon decided (so far) not to support OpenAI’s in-feed shopping (and has blocked the OpenAI crawler), Walmart has decided to join. Good for headlines, but not very surprising: some kinds of companies are ready to partner with platforms (or at least, ready to do an experimental deal to see how it works), and others want to be the platform. As I’ve pointed out before, for retailers and brands this will lean on the same questions that applied to Google shopping and Instagram shopping - does this give you new customers, and is that worth giving up control of the relationship and the experience? How much do you want to trade control for distribution? It depends. LINK, COVERAGE
AI sidebars
Someone should make a taxonomy of all the ways that people are trying to add LLMs onto the top, bottom, or side of their product, and in particular, adding it as a sidebar. This week, Microsoft is adding ‘Copilot’ as a sidebar to Windows 11. With models as commodities, how much does brand and distribution, or product, take over? Of course, Microsoft doesn’t have a SOTA model of its own yet. Indeed, it’s funny to think, with all the big tech antitrust cases going on, that no one really cares what Microsoft does here. (I should test this, but I haven’t used a PC since 2008.) LINK
Apple does Formula 1
After months of rumours, Apple signed an exclusive deal for Formula One rights in the USA (for Americans, Formula One is a car racing thing that’s quite popular elsewhere). The price isn’t public, but apparently, it’s in the region of $150m a year, which is far more than anyone else thought was justified. Meanwhile, Apple also did a bundling deal with NBCU’s Peacock, and Eddie Cue, the exec responsible, did a podcast with Matt Belloni in which he managed to sound frank and unguarded while saying almost nothing, except to repeat that Apple makes TV shows as a business in its own right rather than to promote the hardware. FORMULA ONE, PEACOCK, INTERVIEW
ChatGPT does ‘erotica’
Porn is a pretty obvious use case for generative content, both for conversation and stories/images/video, and now Sam Altman says that ChatGPT won’t block it for verified adult users. Google doesn’t block searches for this (it has a ‘safe search’ filter), which isn’t the same thing but is relevant - if your purpose is to contain ‘everything’, then should you exclude things that are legal even if you personally wouldn’t do them? Sam Altman says he doesn’t want to be a global moral arbiter (and this also follows the current Silicon Valley reaction against content moderation). Of course, people always think content moderation is easy until they have hundreds of millions of users and see how they really behave. LINK
AI content at Spotify and Pinterest
The reflexive use of the word 'slop' makes me wonder what words people used to dismiss electronic music in the 80s, or remixes. Jive Bunny versus Fatboy Slim, perhaps? This week Spotify announced plans with record labels on AI creation tools, while Pinterest announced tools to let you remove AI-generated images from your feed, split by category. Part of the challenge is that different people want content for different reasons. Do you care about the integrity of the artist’s vision, or want five hours of background music while you drive Uber? Do you care whether that table or that room really exists, or just want it for a mood board? It depends. SPOTIFY, PINTEREST
The week in AI
Donald Trump posted an AI-generated video in which he flies in a jet over US cities and bombs anti-Trump protests with excrement. You get what you voted for - a shitposter-in-chief. LINK
Under US pressure, the Dutch government took control of Nexperia, a Dutch chipmaker now owned by a Chinese company. LINK
OpenAI has expanded its ‘Go’ discount-for-emerging-markets product to more countries (note below that India is Anthropic’s second-biggest market). LINK
Waymo will launch in London next year. LINK
Apparently, the head of Apple’s non-public ChatGPT-like search product has left for Meta only a couple of weeks after being appointed. This fits a narrative of Apple struggling, but note also that last week a co-founder of Thinking Machines went to Meta - a lot of people are playing musical chairs in this industry at the moment. LINK
The USA does a $15bn pig butchering’ bust
‘Pug butchering’ is Chinese slang for online romance scams (because you fatten up the pig), and as covered in stories I’ve linked to a few times this year, this is a big, labour-intensive business that seems to be concentrated in large compounds in Myanmar and Cambodia, each with thousands of workers, most of them brought there under false promises of legitimate jobs and then forced to work at online scams or be beaten (or worse), and held at gunpoint. Now the DoJ has seized $15bn of bitcoin held by a Chinese scammer it claims oversaw just one of these operations, ‘Prince Group’.
Of course, this is also exactly the kind of work (boring, repetitive, reliant on human conversation, gated by translation) that LLMs will be very effective at automating. LINK, COVERAGE
Ideas
Your upmarket food delivery might be made by your local kebab joint. Fascinating picture of logistics and cross-licensing. LINK
Unilever case studies on how it’s using AI for scale. LINK
AdCP: an attempt at a standard for selling ads into Generative AI response feeds. LINK
The Indian call centres being replaced by AI. LINK
Ian Whitaker on ‘broadcasters’ using AI and streaming to go after SMEs for the first time, and the question that poses to Meta and Google. LINK
Summarising everything that Sam Altman has said on podcasts this week (this would be a good ChatGPT use-case). LINK
The WSJ points out a wave of power generation startups based on the data centre construction surge that are often speculative science projects. LINK
Walmart is doing a large-scale deployment of ambient (unpowered) IoT sensors in its pallets. LINK
Outside interests
When Donald Judd wanted to do architecture. LINK
The ‘pink it and shrink it’ approach to women’s running shoes leads to injuries? LINK
Maharaja Ranjit Singh in procession. LINK
Data
Memo to the White House - a LOT of America’s AI startups have founders who arrived on H1B visas. CEST has numbers. LINK
Anthropic says that India is its second largest market by use. Unsurprising. LINK
Apple’s manufacturing partners in India exported $10bn of iPhones in March-September. LINK
Wikipedia says (human) traffic is down 8% year-on-year and blames LLMs giving people the answers directly. LINK
Column
Infinite product
For thirty years or so, the internet has meant that we have infinite product, infinite advertising, infinite media, and infinite retail space. All of the old filters, constraints, and gatekeepers fell away: we are not limited by who can afford to buy a nationwide TV campaign or an ad in Vogue, nor by who can fill a distribution deal with a big retailer or build their own retail.
That created huge opportunities, many of which turned out to be a lot harder to fit than they looked (remember the wave of D2C startups? Celebrity brands?) but Instagram ads on one side and Amazon Marketplace on the other powered thousands of new brands, or almost brands, or just people scraping out a few basis points of margin. And if you need a 2700k 50-watt GU10 LED bulb for your kitchen, Amazon will show you 20 ‘brands’ and deliver your choice tomorrow - which is one reason Amazon sold $60bn of ads in the last 12 months.
But all of this also created a very basic problem - when Amazon has 750m or a billion SKUs, how do you know what you want? How do you know about the thing that you didn’t know existed until someone showed it to you? Amazon recommendations are apparently very effective, but they’re also very crude, which gets us all the jokes on the lines of “Amazon, I bought a toilet seat - I’m not collecting toilet seats!”. Amazon knows that ‘if you bought this, you might like that’ only in the sense that your dog knows that if it hears the sound of your door keys, you’re probably going out. It has learnt a correlation, but it doesn’t know why that correlation exists: it doesn’t know what keys are or why they mean a walk. Google and Meta’s ad targeting is better, but not much - they know, but they don’t understand. How do LLMs change that? It’s still correlation, yes, but a very different kind that can move a lot more things from knowing to understanding.
Hence, all three companies have already been talking a lot about how LLMs can change relevance, both for recommendations and advertising, and Google and Meta have been rebuilding their ad stacks. Some of this is ‘just’ short-term optimisation, but the real story has to be how much you can map better understanding of what the thing itself (a video, a post, a SKU) really is on one side with a better understanding of what someone might want on the other. Today, Amazon knows that if you buy packing tape, you might want bubble-wrap, but what if it also knew that you’re probably moving home, and might be a good candidate for a home insurance ad?
The obvious next step here, of course, is ChatGPT, which has hired a bunch of Meta people and dropped hints all over the place. Free-ad-supported has always been the playbook for consumer internet services at scale and although ChatGPT has much greater marginal cost, it still has 800m weekly active users with only about 5% apparently paying - everyone else gets the free, basic version. The routing system introduced in ChatGPT65 has systematic control of which question gets which kind of model (and how much cost), but could clearly use profitability as a parameter as well: is this a paying user, and does this look like a high-value query?
But setting aside revenue, how much could ChatGPT be capturing a different kind of insight into your needs and hence your targeting? How far could it make truly different kinds of recommendation and unprompted suggestion to Meta or Amazon? How different is the query stream to Google? All of these companies are trying to create a picture of you, but each of them only has one part, like the story of the blind men feeling an elephant (a device, like commerce integration, would get OpenAI more signal again here).
On the other hand, speculating about LLM advertising today will probably, in hindsight, end up looking like banner ads for 1995 (one of many ways that today feels like 1995). There are multiple further changes in user experience and business model to come. Indeed, this applies to any speculation about LLMs: the web wasn’t settled in 1995 or even 2000, and mobile wasn’t settled in 2005 or even 2010.