5 October 2025

News

OpenAI unbundles

OpenAI launched a new and much-improved version of its Sora video generation model, and rather more importantly launched it as a new stand-alone iPhone app with lots of social mechanics. It went straight to the top of the App Store chart. See this week’s column. LINK

OpenAI tries shopping

ChatGPT launched an integration of Shopify into ChatGPT, so that you could ask the model to recommend a product then buy it within the app. This sounds like a very big deal (say ‘agentic!’ a lot), but we should also step back and ask why Google shopping and Meta’s shopping integrations never really got traction. On one side, as soon as you want to explore anything about the product, you really need to see the retailer’s website. And on the other, retailers and brand owners are very reluctant to give up that touch point. Will either of those change now? Meanwhile, of course, this will be tightly linked to OpenAI’s advertising strategy. LINK

Capex

The FT says Blackrock is doing a $40bn data centre deal and looking at a $38bn one as well. Brookfield, an asset manager with about $1tr AUM, said it thinks the AI data centre build-out could need $7tr over the next decade. Other people have been suggesting $5tr (McKinsey), $3tr (Morgan Stanley) or $2tr (Bain) by 2030. Obviously, these are really just impressionistic suggestions of the rough scale assuming current trends, not hard estimates. While these sound like absurdly large numbers, on an annual basis this is (say) $500bn, which is comparable to what capital-intensive industries like telcos ($300bn) or oil & gas (600bn) spend.  Of course, those are established, mature industries with revenue and cash flows… BLACKROCK, BROOKFIELD, MCKINSEY, BAIN

On this theme, Meta signed a $14bn infra deal with CoreWeave, which also did a $6.5bn deal with OpenAI. (CoreWeave looks more like a financial instrument than a company to me, but we shall see.) LINK

The week in AI

Following Meta, and prompted by a set of horror stories, OpenAI has launched parental controls. LINK

Meta, obviously, will start using your interctions wiht its chatbot (which has 1bn MAU) as training data. LINK

DoorDash launched a delivery robot (I remember these in Palo Alto 10 years ago…) LINK

An AI ‘actress’ (‘Tilly Norwood’) went viral, annoying lots of people in Hollywood. LINK

The Information reports that OpenAI had $4.3bn revenue and $2.5bn burn in H1. LINK

California passed a limited AI regulation law, requiring some transparency on safety. LINK

Spotify cracked down on AI-generated music spam. (Meanwhile, Daniel Ek will step upstairs to chairman - a good run for the Swedish kid who saved the recored music industry). LINK, LEADERSHIP

Apple back doors

The UK is trying again to make Apple put a back door in iCloud. First it tried to get global access, and had had to back down after the US said this was unacceptable (correct). Now it’s trying for UK residents only, which kind of misses the point: if you break the security to create a backdoor, it’s broken everywhere. LINK

Apple punting Vision Pro?

Bloomberg claims that Apple is shifting development resource form the Vision Pro to a response to the Meta glasses - practical limited product people can use now, rather than cool but over-engineered proof of concept? LINK

Ideas

Zebra (a big frontline automation solution provider - think barcodes) surveyed its customer base in retail, logistics and manufacturing on how they automate now and what they're trying with AI. Interesting both on its own and as a way to think about how tech deployment works in the real world. LINK

Interview on GEO (Generative Engine Optimisation - what does an LLM recommend?) with Google’s head of search. LINK

Automated content marketing at Spotify. LINK

The problem of LLM chatbots’ tendency to sycophancy and ‘delusional spirals’. LINK

AI music versus ‘electronics music slop’. LINK

Many cheap lithium batteries (found both in battery packs and random gadgets) are badly made and dangerous. LINK

Outside interests

Why the US over-pays for buses. LINK

Christie’s is selling one of the mechanical calculators that Blaise Pascal made in the 17th century. It all began here. Only eight examples are known to exist. LINK

Data

Google Cloud research on the state of AI-assisted software development. LINK

Another attempt at analysing the impact of AI on the US labour market. This one says that there no apparent impacts so far. LINK

Column

Unbundling ChatGPT 

OpenAI is a case study in how you might try to build a giant platform company through force of will. 

On one side, it’s trading hype, enthusiasm, and bull market (private) stock for hard assets, alliances, and market position. There are $100bn deals with Nvidia on one side and 10-figure deals with Oracle on the other (and of course, both Jensen Huang and Larry Ellison are playing their own games), and there’s a deal with Broadcom to try to join the hyperscaler ‘we have our own AI chips!’ club. 

On the other hand, there’s a very aggressive push to turn science and raw technology into products and businesses. The Information reported that this summer OpenAI was at a $12bn run-rate, of which almost half is the consumer app, which has 700m weekly active users. Not bad after three years, yes. But Gemini has come up fast from behind, and given that the underlying models remain commodities, how can you differentiate? Name recognition isn’t a great moat, and OpenAI doesn’t have Google’s power of distribution. A network effect might emerge from the science (continuous learning?) but right now there’s a rush to stuff features into the app that might make it feel different - Memory! Shopify! Another button with an inscrutable icon! But how can you make it viral and social? 

We’ve seen a couple of spurts of virality in generative AI already, with MidJourney, and then the ‘Studio Ghibli’ meme earlier this year, and now Google’s ‘Nano Banana’ image model, that helped pull Gemini to the top of the App Store. But now OpenAI is trying to pull a whole capability out into a new social experience, with lots of obvious signs that people from the Meta Growth team have worked on it. Mark Zuckerberg says ‘we have a playbook’ and some of the people who know that playbook are at OpenAI now. 

That doesn’t mean it will work. The history of social is full of ‘fireworks’ - you need to create something eye-catching and different, but that can’t just be something that everyone else can add as a feature (e.g. Stories), and you need to create a new habit that sticks (remember Clubhouse?). Midjourney now reminds me a little of drones: there are lots of compelling enterprise use-cases, but as a consumer, once you’ve seen the roof of your house and flown around the park once or twice, there’s nothing else to do with it, and a week after Christmas it went in the cupboard. A lot of people made a lot of cute images in MidJourney in the summer of 2023 and never went back. Will Sora be the same? I suspect that success would mean your feed is about creators, not friends - more TikTok than Instagram - but then isn’t that TikTok? Can it get a model? Instagram? Can Meta? (Is anyone seriously doubting that Meta needs to spend whatever it takes?) Will there be a whole new mechanic of trading and remixing prompts the way we remix memes now?

Stepping back, though, Sora might be more interesting for existing at all. It demonstrates OpenAI’s ambition. It also demonstrates the foolishness of the idea that a chatbot can ever be a universal interface that does ‘everything’ - there is more to software than that. In a couple of years, you’ll probably have a bunch of new apps on your phone that are all powered by generative AI, and possibly very similar models, but that all do completely different things. That’s always how this works - Tinder, Uber, and LinkedIn are all ‘just databases’, after all, but we don’t call them ‘thin SQL wrappers’, because the value of the company is in working out a new thing that old technology could do for the user.