12 October 2025

News

OpenAI wants to join the club

OpenAI held its annual developer day, with two big announcements: a framework to run third party web apps inside the ChatGPT chat feed, and a tool to build ‘agents’. See this week’s column. APPS, AGENTS

Bubble talk

A couple of weeks ago, OpenAI did a $100bn, 10Gw highly ‘structured’ vendor-finance deal with Nvidia. This week, it’s doing a parallel deal with AMD: it will buy ‘tens of billions’ of dollars of current and future chips, for up to 6Gw of capacity, and in exchange, also get warrants for up to 10% of AMD (maybe $75bn). Like the Nvidia deal, this is heavily structured with lots of targets and tranches. Meanwhile, Elon Musk’s xAI is raising a $20bn round with Nvidia, again, participating. ‘Bubble’ is the word of the week, not least for he Bank of England, Jeff Bezos (’it’s a good bubble!”) and Jamie Dimon. See this week’s column. AMD, XAI, BANK, DIMON, BEZOS

The week in AI

Softbank bought a robotics unit from ABB (Swiss conglomerate) for a bit over $5bn, as everyone gets excited about how AI might make robots more generalisable. I do hope Softbank is not becoming a counter-indicator, though. LINK

Perplexity’s ad chef quit a few weeks ago, and now it says it’s not accepting new advertisers. LINK

One of the five OpenAI researchers who founded Thinking Machines with Mira Murati (former OpenAI CTO) has bailed to join Meta. They raised a $2bn ‘seed’ in July. I don’t know the story here, and people who can build and run SOTA models are in very short supply, though it’s not clear if Meta is still throwing around double and triple-digit million (or more!) hiring packages. But objectively, it is very striking how many of these people seem to feel free to jump from job to job after a matter of months, or even not turn up to an offer they’ve accepted. LINK

ITV is launching generative AI streaming TV ads for SMEs. I wonder how Meta and Google feel about broadcasters globally entering this space. LINK

Google (finally) has a unified enterprise version of Gemini, with governance and ecosystem all connected up (it claims). Distribution is important, and incumbents always try to make the new thing a feature. However, see the next story. LINK

Deloitte’s AI moves

Deloitte did a deal to roll out Claude to its close to 500k employees worldwide. Meanwhile, in Australia, it had to give a refund to the government after a report it produced was full of AI hallucinations. At this stage, this is a training problem, not an AI issue: everyone at Deloitte should know that these things are not databases. ANTHROPIC, AUSTRALIA

Influence M&A

CBS bought the ‘Free Press’, a right-leaning news website founded by Bari Weiss in 2021 after quitting the NY Times in protest at its groupthink, and appointed her editor-in-chief of its new org. She has no broadcast news experience, nor of running anything before the Free Press, and the price was apparently $150m, none of which makes much sense except as a piece of political positioning as the Ellisons (who own Oracle, Viacom, and hence CBS, and apparently soon TikTok) try to cosy up to Trump. They may also be bidding for Warner (CNN). LINK

Everything is hacked

Jaguar Land Rover lost more than a month of production after a cyberattack, while Salesforce had a breach, with customer data for more than 40 companies exposed. Meanwhile, the UK police arrested two teenagers who hacked a nursery school chain and posted the children’s and personal details online in an extortion attempt. Every enterprise technology survey shows security software at the top of the priority list, with growing budgets, but… JAGUAR, ARREST, SALESFORCE

Ideas

Bollywood stars are fighting for ‘personality rights’. LINK

OpenAI launched Sora with minimal copyright restrictions in classic ‘ask for forgiveness, not permission’ mode, but has had to tighten up after a Hollywood backlash. LINK, COVERAGE

Machine learning systems everywhere are mirrors of training data, and so, unsurprisingly, they contain caste bias. LINK

OpenAI published a prompting guide for Sora. LINK

Outside interests

Skydomes. LINK

ThredUp’s automated warehouse. Fascinating. LINK

Data

I’m starting to make a collection of malpractice in AI stats (there’s an irony here). This week, a study of 9th-12th grade US students (roughly 14-18) says that 20% have ‘had a romantic relationship with a chatbot’ - or knows someone who has. At that age, most people would say that they ‘know’ everyone in their year and most people in the year above and below - so this headline is really ‘20% of kids say one of the 2-300 people they know might have done this (or they heard a rumour about it)’. Do better, please. LINK

A detailed Tencent survey on Chinese consumer attitudes to generative AI (hit the translate button in your browser - if you’re using Safari, it will also translate the charts). LINK

OpenAI says its Sora social video app has reached 5m downloads (it’s still invite-only). LINK

Microsoft used anonymised telemetry data from Windows to extrapolate monthly active use of a wide range of LLM chatbots. Useful (though unsurprising numbers) in developed markets, but doesn’t really capture mobile. LINK

A Reuters Institute survey on generative AI adoption this year in a range of countries. LINK

Air Street released their annual ‘State of AI’ report. Probably the best compilation of metrics and survey of all the individual developments in the space. LINK

Column

1/2: Capex bubbles

A couple of weeks ago, OpenAI did a $100bn, 10Gw highly ‘structured’ vendor-finance deal with Nvidia. This week, it’s doing a parallel deal with AMD: it will buy ‘tens of billions’ of dollars of current and future chips, for up to 6Gw of capacity, and also get warrants for up to 10% of AMD. (Like the Nvidia deal, this is heavily structured with lots of targets and tranches.) Meanwhile, Elon Musk’s xAI is raising a $20bn round with Nvidia, again, participating and, again with lots of structure (apparently, the purchase is through an SPV that will lease out the compute).

There are lots of interesting exchanges here. Google and Microsoft, and to a lesser extent, Meta and Amazon, are hugely cash-generative, and they just buy the chips, for now (though their CFOs are now exploring raising debt, and Meta borrowed $29bn for data centres recently). But OpenAI (and xAI) doesn’t have vast existing businesses generating huge cash flows to fund this. So on one side, there’s a scramble to raise petrodollars from the Gulf, and on the other side, they’re doing what in the telecoms bubble was called ‘vendor-financing’ to buy from Nvidia (except that this is also equity). On the other side of this, Nvidia is using those cash flows from the hyperscalers to drive broad deployment of AI, to drive more chip sales, and further entrench its market position. Hence, you could half-seriously suggest that it’s really Google and Microsoft’s money, passed through Nvidia, that’s buying OpenAI’s chips.  

Meanwhile, AMD wanted to be Intel but was no more than a second source: now it wants to be Nvidia, and OpenAI wants it as a second source (and also probably wants to weaken the hold of Nvidia’s CUDA software layer?). If this works, and OpenAI can help boost AMD into an Nvidia alternative (the agreement says they will ‘share technical expertise to optimise their product roadmaps’), then AMD will be more valuable, and OpenAI will get some of that. So if Google’s money is buying Nvidia chips for OpenAI, Nvidia’s money is buying AMD chips, and paying for OpenAI’s ASIC deal with Broadcom...

It’s not surprising that people have started talking about ‘circular revenue’. Nvidia is using its cash flow to drive more cash flow. OpenAI is using its hype to drive more hype: it’s trying to bootstrap its way into the hyperscaler platform club, by swapping its highly valued paper for petrodollars, and leveraging its market position to juice Nvidia and AMD in exchange for more chips. Elon Musk, no novice in the art of raising capital on promises, is doing something similar. The ‘structure’ involved gives more leverage (warrants, tranches and SPVs - this kind of thing is a banker’s dream).

Of course, as everyone has been pointing out in the last couple of weeks, leverage like this works both ways - it accelerates both the upside and the downside. There are lots of ideas for how the music might stop: the demand might slow, all of this capacity might change things when it starts all coming online, the models might stop improving and needing ever-more compute, or we might have a breakthrough that allows the same results for far less compute. There’s an old observation that the most dangerous words in investing are “this time is different,” and the trouble is that every bubble is different, but it can still be a bubble. We don’t and really can’t know where we are in that cycle, and so people ask whether ths is like 1997 or like 1999, and we can’t know that either, but we do know that when something can’t go on for ever, it doesn’t.

2/2: platforms 

Meanwhile, the other side of OpenAI’s bootstrapping is its push to become a platform. Models remain largely commodities, with little differentiation at the general level (though specialised use cases like coding or image generation have more variation) and no sign of a ‘winner takes all’ effect. That might change at the model level, but it can also change at the product level, or at least that’s what OpenAI hopes. So, we have the Sora experiment, which might be the new TikTok and might be the new Clubhouse (it seems very likely that this kind of generative content will be a big deal even if it’s not this), and now we have an app platform. Third-party websites that support MCP can choose to make a widget-style version of their app available to appear right inside your ChatGPT window, responding to your query. 

It’s easy to say ‘this changes everything!’ and it’s also easy to list all the times that this kind of thing has failed before. And OpenAI has some nice launch logos, but then everything like this always does - Windows Phone had good partner logos too. One can also talk about building blocks. If it works, this would help bootstrap a new device - if companies support the app with 800m WAUs, that runs on whatever Jonny Ive is making as well for free. If you give companies reasons to support MCP for this, then it’s in place for other bigger things. More prosaically, this is a data and conversion feed for the imminent ad business. And stepping beyond that, this is an attempt at making an LLM into a distributon layer, of the kind that we had in previous platform shifts from Google, Facebook and then the App Store. Pose your problem, and ChatGPT will either answer it or invoke someone else who can answer it. And yet… do I really want to see half the Zillow website? Why have widgets always failed before, and why is this different?