23 March 2025
News
Nvidia day
Nvidia’s Jenson Huang gave a 2 hour keynote at the company’s GTC developer conference. See this week’s column. LINK
Apple’s Siri problem
After announcing last week that the new LLM-based Siri it’s been promoting for nine months is delayed indefinitely, Apple has shuffled execs, putting the head of the Vision Pro in charge of shipping.
That might seem odd given that the Vision Pro has arguably flopped, but IMO while the decision to launch the Vision Pro was questionable at best, the actual execution of what they decided to ship is first-rate: it’s a bad idea well-executed. Conversely, Siri 2 is a great idea, but either badly-executed or beyond the reach of current execution. We will see. LINK
Apple’s vanity TV project
Meanwhile, the Information reports that Apple has been cutting heads at its subscription TV service, which is apparently losing $1bn a year.
This project has always been a puzzle. Yes, Apple builds services to drive iPhone sales, but those services always produce revenue and support themselves as well. Meanwhile, Apple services are either essential for Apple to have (iCloud, Music, Maps), or try to solve problems for users in some uniquely Apple way (Fitness, the credit card).
Apple TV does none of this. It doesn’t solve any problem, it does nothing different and Apple adds nothing. All that Apple contributes is money: Eddie Cue flies down to LA and buys LA stuff from LA people. Some of the shows are good, yes, but that has nothing to do with Apple. Worse, it doesn’t even drive device sales - almost no-one watches the shows and you can watch them on competitors’ devices anyway. As I and many others have said - why does this exist? LINK
Wiz opens the M&A market?
Google bought Wiz, a private Israeli cloud security startup, for $32bn, its biggest acquisition ever. This comes after a failed purchase attempt last year for $23bn. This is not my industry, but Google is clearly betting that big tech acquisitions will be allowed again - there’s a 10% break clause. LINK
Pirate training
LibGen is a cache of ebooks, many/most of them pirated, that’s floated around AI research circles for years as a source of human-written text without many people noticing. But now LLMs are a trillion-dollar industry, some writers are very upset, and it’s emerged in a lawsuit that Meta used LibGen to train Llama.
There are narrow legal points to this, there’s the general legal question of whether using material (however obtained) for training data counts as fair use and doesn’t need permission, the moral question of training on people’s work… and also the inconvenient question of how much a writer thinks their work is worth if they worry it can be replaced by a statistical averaging machine. LINK
Google chips
Google’s TPU server AI chips are designed in partnership with Broadcom (and made by TSMC), but now Google may be adding MediaTek as a cheaper partner. LINK
Anthropic web search
Anthropic added ‘web search’ to Claude - as for competing systems, this means that it can answer questions by looking at the web today rather than relying on whatever was in the (static) training data for the underlying model. In practice, that means running a standard, old-fashioned web search and then using the LLM to analyse the results: in Anthropic’s case, it appears to be using Brave’s index. LINK
BYD’s fast charging
The Chinese EV industry is already clearly pulling ahead of the legacy incumbents, much as Chinese Android OEMs pulled ahead of the incumbent phone makers in the 2010s. But now BYD is also claiming to have a solution to ‘range anxiety’ with a fast-charging tech that can recharge a car in 5 minutes. Actually deploying this will be hard; such charging stations will need higher-power grid connections and be more expensive besides. But any Americans still claiming that all China does is copy and pay lower wages really need to wake up and look at the last 20 years. LINK
EU versus Apple
The EU produced a long list of new changes it wants in the ways Apple devices talk to each other, opening those features up to competitors. Some of these demands come with significant security and ease-of-use challenges (do you really want your iPhone to accept incoming Airdrop requests from any untrusted device?), but there’s also a basic principle: part of Apple’s competitive proposition is that its devices all work better with each other, and the EU is declaring that’s not allowed. LINK
Ideas
A good profile of Skydio, which was making small autonomous drones for the last decade without finding a good use-case, until first Russia’s invasion of Ukraine and then the rise of defence-tech sovereignty (driven first by China and now Trump). LINK
Ben Thompson interviewed Sam Altman. Interesting quote: “I am more excited to figure out how we can charge people a lot of money for a really great automated software engineer or other kind of agent than I am making some number of dimes with an advertising-based model.” LINK
Google ran a test removing news for a sample of users, and found minimal impact on use or revenue. The only people who do not already know this are determined not to hear it. LINK
The French and German governments have a ‘sovereign’ Notion alternative. No longer as silly as it would have seemed three months ago. LINK
How good was the TikTok ban for its competitors’ ad businesses? LINK
The incentives behind streaming TV lead to a lot of expensive, generic, predictable and forgettable filler. I mostly agree. LINK
The industry behind Dubai’s drone shows. LINK
Zapier wrote a story on the best app builders of 2025 (lots of repositioning going on right now). LINK
Yahoo is still here, and has plans. LINK
Outside interests
Graydon Carter’s memoir is a window into a time when the ad industry paid for magazines to put writers on retainer at half a million dollars a year for three articles. LINK
Sotheby’s has a sale of sci-fi movie props and memorabilia, including an ET model. LINK
UBS inherited a lot of art from Credit Suisse, and isn’t quite sure what to do with it all. LINK
Data
Meta’s Llama has had a billion downloads (kind of a vanity metric). LINK
Meta says that its VR headset use is increasing ‘by 30% a month’ - with no denominator, though. LINK
Lenny Rachitsky, who runs a popular newsletter aimed at tech product and growth, surveyed his user base on the software they and their companies use. Heavily self-selected to startups and early adopters, obviously. LINK
A survey saying that no-one can tell different streaming TV services apart. LINK
The IFPI says that global recorded music revenue grew 4.8% in 2024, driven (obviously) by subscription streaming. LINK
Column
Nvidia as the new Sun?
Jenson Huang gives lots of keynotes, but it’s worth watching the latest if only to get a sense of the scale of his ambition. Nvidia doesn’t sell GPUs - it sells computers, and those computers are designed to be bought in batches of hundreds of thousands and plugged together into an integrated system the size of a small city, using as much power as a small country. Nvidia wants to be the new Sun Microsystems, and the new Microsoft: it will sell its own custom-defined boxes, but also its own software stack on top to build your applications to sell to the world. This was, after all, a developer conference, and a lot of the keynote was about development - all the things that people are building on top of what Huang calls ‘AI factories’.
Factories is an interesting word. A lot of the keynote was about apps, but a lot also was about the economics of running LLMs at scale on Nvidia hardware, ad really, when he says ‘AI factories’ we might also say ‘token factories’ - Nvidia says it’s selling the most efficient ways to make tokens - the atomic unit of LLM computation.
That makes me a little nervous, though. Huang said ‘trillion’ a lot, especially when talking about data centre spending, where he showed an analyst’s forecast that data centre capex would rise from a quarter trillion dollars last year to over a trillion by 2028 or 2030. Yes, the big four platforms spend $220bn last year and are on track to spend over $300bn this year, but Microsoft, spending $90-100bn this year, will be at over 30% capex/sales. These kinds of numbers presume an enormous net increase in corporate and consumer spending in just a couple of years - not replacement of spending on Oracle, Salesforce and Office, but additional.
Equally, Nvidia has two headline charts: one showing how much cheaper it is to make tokens with each of its new and upcoming architectures, and one showing how much more people will spend buying those architectures. This is theoretically possible, of course (the Jevons paradox, yes), but there are plenty of examples where a cost collapse produces massively greater use without massively greater revenue. Price elasticity can work in more than one way. And the entire direction of travel of tech for the last 50 years has been away from paying per use, per second, per bit or per character, and towards bundled flat rates. Sam Altman (see above) might want to sell agents for tens of thousands of dollars, but he would say that when Deepseek, Llama and AWS are pushing prices orders of magnitude lower than his. There are no apparent winner-takes-all or network effects to LLMs, and in that case the natural state is for this to be commodity infra sold at marginal cost, with marginal cost in free-fall.
That said, Nvidia’s product itself is not a commodity, with everyone in the field agreeing that it has a long lead and high moat. Unlike, say, Cisco (an over-used comparison), its customs aren’t going to go bust - the hyperscalers have far more cash than the fibre carriers, and hundreds and thousands of startups will run LLMs (whether those are commodities or not) on Nvidia infrastructure. But there is a conflict between costs falling 10x and capex increasing 10x, and a conflict between monopolistic vertical integration and disaggregated commodity tokens. And remember what happened to Sun.