7 December 2025

News

Netflix buys Warner

Netflix, ‘the Albanian army’, is buying Warner Brothers, for $72bn. See this week’s column. LINK

OpenAI’s ‘Code Red’

Apparently, Sam Altman reacted to Google’s Gemini getting to the top of the leaderboards with a ‘code red’ memo telling everyone to pause work on ads and on new user-facing products (agents, Pulse) to focus on getting the models back ahead. Pausing product and doubling down on models, and not the reverse, is interesting given that models seem to be commodities and OpenAI’s core strategic problem is that it has no unique product and no distribution. LINK

Meanwhile, the WSJ claims that Altman has explored buying or partnering with a competitor to SpaceX. People have talked about datacentres in space (a good sign of a bubble), but this seems unfocused, unless of course it’s just a way to annoy Elon Musk. LINK

Amazon’s silicon

Last week we heard that Google is looking at selling its TPU AI accelerators to other big tech companies; this week Amazon announced a new version of its “Trainium” chip, claiming efficiency gains over Nvidia. Jensen Huang’s lead is solid for now, but merchant silicon is chipping away at the edges. LINK, ANALYSIS

Capex wobbles

The big four platform companies will spend about $400bn of capex on data centres for AI this year, and as I pointed out in my latest presentation (linked above), so far most of that has come from cashflow, but that’s changing. Meta has done two $30bn deals, one straight debt and the other an off-balance-sheet SPV, while Oracle has gone on a borrowing spree, leveraging its cash-generative but declining legacy business for debt to buy its way into the new thing. Morgan Stanley, which is in a lot of these deals, is hedging, and hedges on Oracle’s debt have also spiked. DEBT, ORACLE

The week in AI

The Information reports that Microsoft had to lower some sales quotas for Azure as customers pushed back on deployment and ROI challenges. Welcome to the Trough of Disillusionment. LINK

OpenAI will have to hand over 20m anonymised chat logs in its copyright dispute with the New York Times and others, which want to use the logs to prove people are using ChatGPT to bypass paywalls. OpenAI has been fighting this on user privacy grounds. LINK

Amazon says its ‘Rufus’ chatbot is driving real sales, with 250m users in the last year (how many on purpose?), those users 60% more likely to complete a purchase, and with the model on track to deliver $10bn extra annualised revenue (which seems hard to reconcile with that 60% number, though?). LINK

Google, obviously, now has an ‘agent builder’. LINK

Meta won’t invest in Yann LeCun’s new ‘world model’ startup, which will be based in Paris. LINK

Bytedance is building a $38bn data centre in Brazil. LINK

Anthropic is hiring advisors to look at an IPO, perhaps as soon as next year. Apparently, it’s currently in talks for a new private round at a $300bn valuation. LINK

Apple turnover

Apple’s AI chief, John Giannadrea, will retire from Apple, being replaced by another Google veteran (poached from Microsoft after spending only six months there). This is not a surprise: Apple isn’t shipping the AI stuff it wanted to, and Giannandrea has apparently struggled to mesh with the inner circle of Apple execs (which I’ve seen compared to a family business). As for Meta, this seems more of a political and organisational problem than a technical one, though recall the story a few weeks ago that Apple was working on using a version of Google’s Gemini model (on its own servers) as a stand-in until it can get back on track. LINK

Meta cuts Metaverse?

Meta’s accounts have shown a $4-5bn quarterly operating loss for the ‘Reality’ Labs’ unit for close to a decade now, with any scaled consumer VR or AR products still years away, let alone anything that could really take on smartphones. One has to be a little careful here, given that quite a lot of general R&D is also in that number, but either way, Bloomberg reports that Mark Zuckerberg is cutting headcount by a third and squeezing spending, as he scrambles to get Meta’s LLM labs back on track. After all, with the $18bn operating loss that Reality Labs booked in the last 12 months, they could have hired three or four more researchers. LINK

Apple’s design head goes to Meta

On the other hand, Alan Dye, Apple’s head of interface design, left for Meta to work on VR and AR stuff. One could use the old joke that this improves the talent at both places, except that that might be unfair to Meta: Dye leaves a problematic track record, and there are a lot of places where Apple software UX is kind of a mess. The recent ‘Liquid Glass’ redesign is a misconceived fiasco, bringing inconsistency and a lot of basic and design problems: Apple always says “design is about how it works, not how it looks” but this looked cool in demos and works badly. Meanwhile, the iPad now has three different windowing systems, each successively more confused, and the Watch has dozens of faces that ape physical watches and almost none that care about information design, while most of them even support the once-per-second screen update in the newer models. Yes, this is partly a customer’s complaint about details, but design and details are Apple’s whole proposition. LINK, and read John Gruber’s COMMENTARY

Banking stable-coins  

People still happily speculate in Bitcoin, but NFTs seem like a pandemic fever dream now and ‘Web3’ - the idea you could build actual web-scale software on top of blockchains - is five years away and may always be. But meanwhile, people are busy building useful plumbing with this stuff, deep inside the finance industry, solving problems you need to work at a bank to understand. Here ten EU banks are working on a stablecoin, including BNP Paribas. LINK

CNN does ‘prediction’

Somewhat related, CNN did a news and data deal with Kalshi, the hot new ‘predictor market’ company. You can call this the wisdom of the crowds, but you could say that this is decentralised, unregulated gambling? LINK

Ideas

No-one in Hollywood seems to be happy about the Netflix Warner deal. Vanity Fair sums up the mood. LINK

A profile of the Syrian start-up scene. LINK

AWS says they’re OK not having the best LLMs. Well, they would say that, wouldn’t they? LINK

The UK’s budget watchdog, the Office for Budget Responsibility, accidentally put its review of this year’s budget at a publicly accessible URL too early, effectively leaking the budget. The post-mortem, says, in effect, that 1: it’s their fault for thinking they were a unique special flower that should run their own website instead of using government infrastructure and 2: it’s their fault for using Wordpress. Key quote: “Technical commentary has, for many years, noted that WordPress can be onerous to configure and that mistakes are easily made in so doing.” It’s ironic that the thing that was supposed to democratise publishing became a headache no-one without on-call tech should ever touch. LINK

Outside interests

A first edition of William Harvey’s book explaining his discovery of the circulation of the blood. LINK

Data

Anthropic released another of its attempts to analyse LLM impact on productivity. I am off-the-scale sceptical about the methodology in these. LINK

The OECD’s growth outlook points out that US GDP growth is now flat if you exclude data centre construction. Generative AI is effectively a private stimulus programme, funded mostly by advertising (and now debt). LINK

A new McKinsey report on enterprise LLM deployment. LINK

BCG survey data on how the marketing industry is thinking about AI. LINK

OpenRouter released a report on which LLMs its user base favours. Treat with caution, though, given that 50% of the open source use it reports is for ‘role play’ (yes, this is a euphemism), and for that and other reasons this might not be the most representative sample. LINK

Column

Content isn’t king

In 2019, I wrote an essay arguing that all of the questions that mattered for Netflix are TV questions, not tech questions. Netflix has used technology as a crowbar to build a new TV business. Everything about how it executed that technology has to be good. The apps are good, the streaming and compression are good, the UI is good, the recommendation engine is good, and the customer service and experience are good. Unlike American cable subscribers, Netflix subscribers are generally pretty happy with the tech. The tech has to be good - but it’s still all about the TV. 

If Netflix were only showing reruns of Frasier and Ally McBeal, no one would have signed up. It used tech as a crowbar, and the crowbar had to be good, but it’s actually a TV company. You can see this pretty clearly if you contrast Netflix with Hulu. The reasons that Hulu doesn’t have 150m paying customers have nothing to do with its technology, but to do with the TV. 

You can see this split in the reaction to the Netflix Warner deal this weekend: all of the questions are LA and New York questions, not San Francisco questions. No one in tech cares. 

The other side of this, I think, is another essay I wrote, back in 2027: “Content isn’t king”. My thesis there was that TV, movies, and music no longer have any strategic leverage for tech companies. The shift to subscription and to streaming instead of buying and downloading means that content is no longer a way to get people to buy a device or join an ecosystem: if you cancel Spotify and move to Apple Music, or replace your iPhone with an Android, you don’t lose anything: you don’t lose content you’ve bought. Since then, Apple’s decision to offer the TV shows it buys in LA on non-Apple devices just proves the point: this has no strategic leverage. 

There’s an old joke that consultants are like seagulls - they fly in, make a lot of noise, crap everywhere, and then leave. Tech does this to media sectors: it flies in, makes a mess, changes everything, and then leaves. People in tech used to care a lot about digital music; now it’s a rounding error. No one cares about ebooks, and no one pays attention to streaming. Those are music, books, and TV questions: like Netflix, all the questions for Spotify are music questions, not tech questions. 

YouTube is the exception in this framework. TV people spend a lot of time right now arguing whether YouTube is ‘television’, which of course depends entirely on what you mean when you say ‘television’ - are you really talking about whether YouTube ads achieve the same objectives for an advertiser, say, or about how professionally produced content works on YouTube, or something else? Definitions, like analogies, can be traps, though: you can spend more time arguing about the definition than the question. Netflix shows essentially traditional TV in a non-traditional way; conversely, MTV showed things that didn’t look at all like TV, in a completely traditional linear channel. YouTube is something else, but who understands the questions - MrBeast or someone in LA? Which question? 

And so, on to the next question: what will generative AI do to this industry? The answer to that will be very different if you’re Netflix or YouTube, but it will also be very different if you’re Netflix or ESPN. YouTube competes for time with Netflix, yes, but they also both compete with a PlayStation, and all the questions for what AI means for games are different again. Back when we were all interested in autonomous cars, I made a related point about retail. The car industry made car ownership ubiquitous, but Sam Walton wasn’t an automotive engineer: he was a retailer who realised what mass car ownership would mean. ”What happens when everyone has a car?” wasn’t a car industry question, just as “what does broadband mean for TV?” wasn’t a telecoms or tech question. And what does generative AI mean for XYZ industry?” will be a question for those industries, not for AI scientists.