21 December 2025
News
Is the TikTok deal finally happening?
This deal is Groundhog Day, but it looks like it’s finally happening, with the company actually signing the deal to spin off the US assets into a consortium led by Larry Ellison (whose family also owns Paramount / CBS and is trying to buy Warner Bros.) and Silverlake. Bytedance will keep 19.9%, though the Chinese government still has to agree. Note that MGX, the Abu Dhabi state investment firm, is in the deal - there are also petrodollars in the Warner bid. As before, it is striking how cheap this deal is - $14bn where analysts were estimating up to $50bn, given that it has US revenue of at least $10bn. LINK
Agentic interfaces
Following my notes on agentic interfaces last week, Google released A2UI, a framework for LLM agents to generate GUIs in a standardised way. Worth comparing to OpenAI’s developer doc of UX principles for agents. This all feels very 1997, when the web was clearly going to be huge but no one really knew how it should work and there were no standards for all the basic building blocks that developers and designers wanted to use. Remember ‘This site best viewed in Netscape’ buttons? So, you make your own tech, and try to use your market share to make them standards, or use your standards to capture market share. GOOGLE, OPENAI
Amusingly, OpenAI also has a set of app submission guidelines, which read exactly like Apple’s early app store guidelines. Everyone wants to be a gatekeeper. LINK
China cloning ASML?
Reuters has a big splash saying that China has a huge ‘Manhattan project’ to clone ASML’s semiconductor fabrication equipment so that China can make cutting-edge chips, based on a team of former ASML engineers. Of course they do, this is always the playbook, and in the long term this is probably inevitable, but China has been trying to catch up for a generation without getting anywhere close, and TSMC is ahead for a lot more reasons than just buying ASML equipment. LINK
The week in AI
Amazon is in talks to invest $10bn in OpenAI, closing another loop: it also has investments in Anthropic, while Microsoft also has investments in both. LINK
Amazon re-organised its AI group and changed the leadership. (It is fascinating that Microsoft and Amazon have spent three years trying to get onto the leaderboards and can’t crack the top 10, while Xiaomi made it there this week.) LINK
Databricks is raising $4bn at a $134bn valuation. Remember those things called ‘IPOs’? Also, apparently this is a Series *L*, which I’ve never seen before. LINK
OpenAI hired a senior product head from Shopify. More agentic commerce aspirations coming. LINK
UPS is buying hundreds of robots for unloading trucks. Not humanoids - very traditional-looking arms with a lot of very non-traditional AI, from ‘Pickle Robotics’ (great name). LINK
Meta is shifting a lot of budget to AI from the VR-focussed ‘Reality Labs’, which may be why this week it paused the third-party headset programme that it announced last year in an attempt to catalyse an ecosystem. But then, this isn’t a mass-market consumer product yet anyway, so maybe it’s switching emphasis to making the product itself more viable? LINK
Age verification
Australia’s ban on under-16s using ‘social media’ has taken effect, with predictable definitional challenges (WhatsApp, Roblox, and Discord aren’t covered). The law requires either AI age estimation, which is pretty unreliable, or uploading a government ID, if you have one, which comes with its own challenges - see the next story. LINK
Viewing data for Pornhub premium subscribers was affected by a breach last month at MixPanel, an analytics platform, and the hackers are now trying to extort payment. This prompted a lot of prurient schadenfreude, but it also hints at the problem with demanding such companies collect IDs. LINK
New money
With the US loosening financial regulation, PayPal has applied for a banking licence, so that it can offer small business loans and savings accounts directly rather than going through licensed intermediaries. This may be the only company that can deliver a worse experience than a traditional bank, which I suppose is one kind of differentiation… LINK
Meanwhile, Visa is doing more stablecoin stuff. LINK
New TV
I’m not following the Warner/Netflix/Paramount take-over battle, because I’m interested in tech, and this is a matter for M&A bankers and TV analysts. But this week YouTube bought the rights to show the Oscars live, winning them from ABC, which has had the deal since 1976. Tell me again how YouTube isn’t TV. LINK
The great car reset
I am equally hesitant about how far to dig into the EV transition, which seems to me almost entirely a conversation for automotive analysts. Tesla pointed to the opportunity, but has fumbled model releases while the CEO’s antics have wrecked the brand, and pumps retail demand for the stock with promises of autonomy that never arrive. Legacy manufacturers struggle with the cost and execution of the transition, and demand is behind their hopesL This week Ford is taking a $19.5bn charge as it slows down the shift and cancels the electric version of its best-selling F150 pickup, while the EU has pushed back its planned 2035 ban on internal combustion engines. But meanwhile, China is running the same playbook we saw with smartphones: hundreds of manufacturers winnowed down to a few champions, with domestic overcapacity and subsidies, and then a global wave. BYD is hiccupping, but Marques Brownlee (the current king of YouTube gadget reviews) loves Xiaomi’s car, and Ford’s CEO keeps saying the same. FORD, EU, XIAOMI, CEO
Just to emphasise the point, this week California, a decade late, finally told Tesla that it cannot use the terms ‘autopilot’ and ‘full self-driving’, since Tesla cars do not, as a matter of fact, have full self-driving capabilities and no-one knows when (or if) they will. LINK
Naval drones
The war in Ukraine continues to be a testing ground for new concepts, although the uniqueness of the situation should prompt caution (the Spanish Civil War had mixed lessons for WW2, and the Balkan wars little about WW1). But this week the Ukrainians used an underwater drone to enter a Russian harbour and cripple a submarine. LINK
Chinese tech regulation
Staff at PDD got into at least two fist fights with staff from a Chinese regulator investigating fraudulent deliveries. I wonder if this lobbying model will spread - UFC versus FTC could get record viewing. LINK
Ideas
Andrey Karpathy wrote a good non-technical write-up of technical developments in generative AI this year. LINK
And Matt Turck wrote up the venture/startup/SaaS perspective. LINK
Mapping the SIM farms that enable fake accounts, spammers and manipulation. LINK
A one-hour interview with Sam Altman. Worth listening to the whole thing, but one quote stuck out for me quite tangentially - pointing out that the term ‘AGI’ means different things to different people: “There’s a lot of people that would say we are at AGI with our current models.” LINK
The Information says Google has set up an internal committee to allocate finite compute resources between different AI projects. LINK
Gossip from the Information that OpenAI optimised model development for benchmarks that focused on complex and obscure maths and scientific puzzles, and was then surprised that those don’t map well to normal consumer use cases. This puzzles me. I have always thought of benchmarks as trying to say something about underlying conceptual capabilities (and of course in how much they’re being gamed). I didn’t realise that the people who work on them would think users really wanted a model that can do the actual tasks in the evals. LINK
Back in the real world, while we’re all told to get excited about agentic shopping Right Now, here’s an adversarial analysis of how well the latest models respond to questions about products that don’t exist. Mostly, they drop straight to hallucination. LINK
Conversely, Nieman has a good overview of use cases in journalism. LINK
And Nokia wrote about telco use cases. LINK
Meanwhile, one strand of private investment in AI (especially for the 99% of funds that can’t get an allocation in the latest OpenAI round) is to back sector roll-ups: a startup raises cash to buy a bunch of companies in a legacy service industry to consolidate and optimise them with LLMs - healthcare brokerage, customer services, insurance, logistics, etc. I can see the logic, but these kinds of structures are always popular in bull markets. LINK
With Trump trying to block state-by-state AI regulations in favour of one federal rule book, A16Z, which is deeply connected to the technocratic side of his administration, published a proposal for what that rule book might want to address, looking both at protecting against tangible harms (children, consumer abuse, cyber) and at driving investment. (NB: I worked at A16Z from 2014 to 2019.) LINK
Bloomberg found that around 90% of H1-B hires at Tata Consulting, Infosys, and Cognizant would be covered by the new $100k fee. LINK
Outside interests
If you need a last-minute Christmas gift, there’s always the Cloud Appreciation Society. LINK
Apparently, the new most obnoxious thing to do at an airport is booking a wheelchair to take you to your gate when you don’t need it. When these people suddenly get up and walk to their seat, airline staff talk about ‘Jetway Jesus’. This would make a great Seinfeld episode. LINK
RIP John Olson, who was fired after refusing to make Enron a Buy. LINK
Data
Cloudflare’s year in review - lots of data on LLM robots scanning web pages. LINK
Gallup says generative AI DAU at work in the USA is about 10%, versus 23% WAU. LINK
People watched 700m hours of podcasts on YouTube on their TVs in October, almost double the number a year ago. (For context, Netflix does about 15bn hours a month, and its top show got about 75m hours last week.) LINK, NETFLIX
The value of music copyright in 2025. LINK
Last but definitely not least, someone claims to have scraped the vast majority of Spotify, downloading 300TB of music with 256m tracks and all the associated metadata and making it available as a torrent (remember those?). This will be very tempting to anyone building an AI model for music (remember the Anthropic books case!), but the popularity data might be most valuable - a giant set of human feedback? LINK
Column
AI, 2025 and 1997
I took 65 flights this year, and gave a lot of presentations to a lot of people, because lots of things are changing. Indeed, things are changing so much that I moved from making one macro tech trends deck to two. This is the last newsletter issue of the year, and so I’m supposed to write a recap of everything that changed, and there’s a lot you can say, but looking back, it seems to me that the important thing is how much the conversations have become separated and specialised. ChatGPT blew up three years ago, and for at least the first two years there were only a pretty narrow set of questions: how much better would the models get, how many would there be (would Google/China/open source catch up?), and how far could the models just do ‘the whole thing’ versus needing lots of product on top. Now all the conversations have broken apart, from the top to the bottom of the stack, and there are three-hour podcasts for all of them.
So, we now have detailed analysis about ‘infra’, which splits into chips, data centres, electricity generation and connections. Faster chips haven’t really meant much to anyone further up the stack since the 90s, and data centres haven’t been interesting to non-specialists since the early days of Google, but now these are the stack and they’re the constraint for everyone else: the big four platform companies will have spent around $400bn this year and are far behind demand. So suddenly everyone needs to know what EUV means, and where Oracle’s CDS trades, and look up what ‘gigawatt/hours’ are in Wikipedia, and we discovered that no, we can’t double the production of gas turbines next year. I don’t think Nvidia is Cisco (Sun Microsystems is a much more interesting comparison), but this does look like fibre, particularly in the way people start to confuse ‘there is more demand than supply’ with ‘any investment will pay off’, and complicate the balance sheets (remember IRUs?). But for now, Jensen Huang keynotes are the new Steve Jobs keynotes.
The next level up in the stack, in the models themselves, there’s a basic split. On one hand, there is an enormous number of new acronyms, reflecting proliferating technical sophistication and complexity, as people iterate around a handful of core concepts, and, meanwhile, it’s very clear that there is no defensibility to a SOTA model - with the right engineers and a billion dollars, anyone can make one. But on the other hand, the gurus of the field (and I choose this work deliberately) make Buddha-like prognostications on how the field might evolve, all of them wrestling with the fact that we don’t really know how or why this works so well, nor what it can’t do, nor what separates it from people. We can say that progress has continued, and lots of people say ‘exponential’, but we have neither an empirical nor a theoretical way to know whether we are now building out a technology that will get much better and more mature but not change in character - like the internet in 1997, say - or whether we’re at the beginning of a vastly bigger change. The vibes seem to be shifting towards the former view, that we need other unknown breakthroughs (Demis Hassabis, Ilya Sutskever, Yann LeCun are on this side), but this is only vibes - we don’t know.
Going up the stack again, into the product layer, even this summer it really felt like there was no product strategy. All the labs made a chatbot, and the chatbots all looked the same. The models might be better or worse (though not very), but the products were all the same - the biggest difference between Claude and ChatGPT was the colour of the icon. That’s changed a lot since then, or at least become a lot more clear. Google’s reaction has come on stream: the models caught up, but much more importantly so did the product and distribution, with a lot of survey data now showing it with half or 2/3 the use of ChatGPT already. The bear case for OpenAI, indeed, is that it’s MySpace but without a network effect, or Netscape. On the other hand, Sam Altman doesn’t sit still, and the company is ‘flooding the zone’ with product ideas and press releases, trying everything, all at once, yesterday, at every level of the stack. Really, it seems to me that OpenAI is moving as fast as it can to swap mindshare and expensive stock for tangible infrastructure, market position, and consumer lock-in, because today, all it really has is a lab that’s as good as anyone else’s, a brand, and 800m users who mostly only visit one a week and could easily switch. It’s a mile wide and an inch thick.
I think about the web in 1997 a lot, metaphorically, and about mobile in 2005 and 2010. We knew this was huge and there was a lot going on, but most of what was going on turned out not to matter, and we didn’t really know how this would work. I have an i-mode phone, I worked on a DVB-H project, and I used Pointcast. It’s not just that we didn’t know who would win - we didn’t know where the competition would be. Microsoft won browsers, and I’m old enough to remember that Internet Explorer was actually a better web browser than Netscape (sorry Marc) but it turned out that winning browsers didn't get Microsoft anything. That doesn’t mean we can’t make predictions, though. Looking at that stack - commodity infrastructure doesn’t tend to have super-normal margins in the long term; if models remain commodities too then the same will apply; and if the right use-cases and interaction models haven’t been invented yet, the companies that work it out probably haven’t been founded yet.