16 November 2025
News
Cursor’s two years to a billion
Cursor, the red-hot AI coding assistant, raised its seed round less than two years ago - this week it raised $2.3bn on a 29.3bn valuation. More significantly, it passed $1bn in annualised revenue. Valuations may be frothy, but people are buying products. LINK
Agentic shopping from Google
Google launched a range of new AI and agentic shopping tools this week. It’s a lot easier to say that people will use agents to shop than to work out how (or even if) that would work, and how it would reshape existing funnels, but obviously enough Google isn’t waiting for someone else to work it out first. So, it’s building new tools but also, mostly, making this a feature and leveraging its existing distribution (one reason Gemini has grown so strongly). LINK
Agentic ad assistants
Again making the new thing a feature, Amazon and Google both launched new agent assistants for their ad and analytics platforms. This really is what Clippy was trying to help with: the more features you add, the harder it is to work out how to do anything. This applies to all SaaS, but especially to self-serve products aimed at SMEs - how can you help more people to access the power of the tools without making them go on courses? GOOGLE ADS, ANALYTICS, AMAZON
Cheap power for Chinese AI?
The FT reports that local governments in some Chinese provinces are offering subsidies of up to half of energy costs for AI data centres that use Chinese chips. The catch, of course, is that those chips use 30-50% more energy than their Nvidia equivalents (if you can smuggle them in). Still, apparently China has significantly less constraint on new supply and connectivity than the USA, where energy suppliers are now backed up for years; some hyperscalers have said access to electricity is now a bigger problem than access to chips. China is pulling levers for industrial policy. LINK
LeCun leaves Meta?
The FT reports that Yann LeCun, Meta’s chief AI scientist and OG of machine learning, is planning to leave to do his own startup. This would not be very surprising. On one hand, he has argued forcefully that the LLMs that have taken over the tech industry are limited and won’t lead to AGI (as OpenAI and many other labs think). INstead, he argues for a new approach of more embodied systems that can learn from observation instead of being ‘retrained’ on vast amounts of static data. On the other, Zuck layered him, and has been spending billions to hire lots of new researchers for a whole new lab. LINK
The week in AI
So far, Anthropic has rented cloud AI infrastructure from Amazon and Google (both investors), but now it’s decided to join the rush itself and will spend $50bn building its own data centres, all in the USA. I’m old enough to remember when a $50bn infra plan would be big news. See the next story. LINK
Meanwhile, ex-OpenAI CTO Mira Murati’s ‘Thinking Machines’ is apparently raising again at a $50bn valuation, still pre-revenue and pre-product. Only 50? LINK
OpenAI’s product rush continues, with group chats in ChatGPT. It’s almost as though Sam hired a bunch of growth people from Meta. LINK
Google quietly announced its own version of Apple’s Private Cloud Compute, wherein a model running on the big, heavy compute in the cloud processes your data in a secure enclave with Google never seeing it. Google calls it ‘Private AI Compute. LINK
Rightmove, the UK property aggregator, made the market happy by announcing a new AI investment plan. Wait, not like that - the stock fell almost a third. LINK
Amazon discounts
Amazon is expanding its ‘Haul’ discount bundled shopping concept with a new app called Bazaar, targeting a range of middle and low-income countries, with an interesting overlap with the Chinese expansion data linked below. LINK
Apple’s WeChat deal
Today in headlines from 2015, Apple has done a deal with Tencent over payment for ‘mini-apps’ within the WeChat ‘super-app’. There was a time when everyone in tech was very interested in every part of that sentence, but now only Tim Sweeney cares. LINK
Anthropic as a hacking tool?
Anthropic claims it discovered that Chinese hackers used Claude to build and run a cyberattack on ‘roughly 30’ targets, and a ‘small number’ succeeded. This might be a big deal, but it also fits a pattern of Anthropic making scary-sounding claims that don’t withstand scrutiny. Yann Lecun thinks they are saying this just to try to get regulation (and shut out competition), but it might also be self-selection - this is the company that says ’come work here if you’re scared of AI but also want to build it’. Either way, security professionals are pretty sceptical. ANTROPIC, REACTION
Ideas
Two interesting papers from Meta on the ways it’s using LLMs to optimise the existing business: ad recommendation and search. ADS, SEARCH
Remember when streaming was the end of piracy? Rights windows, fragmented subscriptions, and, well, the love of getting something for free means there’s a thriving market in illegal streaming sites and HDMI sticks to access them (mostly hacked Amazon Fire sticks). 10% of UK adults in the last six months, apparently. LINK
Unpicking the UK’s proposed ITV/Sky merger: the internet has changed all of the old arguments about media consolidation. LINK
India runs on WhatsApp (for anyone who didn’t know that) - lots of interesting details. LINK
Ofcom study on use of AI to replace search. LINK
“Self-driving” means a car that can drive itself (memo to Elon Musk). LINK
Outside interests
Mapping London’s unbuilt ring-ways. LINK
Apple trolled its critics with a $230 iPhone sock. LINK
Audiophile snake-oil, from cable elevators to quantum. Lots of quantum. Do your HDMI cables have silver connectors, for better ones and zeros? Maybe that’s the path to AGI? LINK
Data
A database of cases where lawyers used LLMs to create documents and didn’t check the output. These things are very useful, but they are not databases. LINK
Bain survey data on consumer use of LLMs for search and shopping. LINK
Similar from McKinsey. LINK
Bain also has some interesting data on Chinese e-commerce going global (it’s not just Shein and Temu). LINK
Where does LLM web search like to look? The most cited domains in AI. LINK
Tiktok Shop is now roughly the same size as eBay. LINK
Column
I've been on planes this weekend, on my way to present my new macro trends presention at Slush in Helsinki on Wednesday. That means no column this week, so here’s one from earlier this year.
How does AI change e-commerce?
Anyone trying to sell something to consumers - brands, marketers, ad agencies and retailers - used to be defined by limits. There was limited retail space, limited inventory, limited selection, limited media to tell you about it, a limited number of ad slots, and a limited number of people who could afford to buy them. Mark Read, the CEO of WPP, said a while ago that until about 2005, if a big CPG company wanted to work out its ad strategy for the year, the CMO would have lunch with two or three TV execs and they were done for the year. And those TV execs had the same hard limitation - there were only so many prime-time Saturday night programming slots they could fill.
All of that’s now reversed: we have infinite product, infinite inventory, infinite media and infinite ad space, and every ad now has a hundred variants. That produced the D2C wave a decade ago, and Shopify (now powering $270bn of GMV), and now Shein and Temu, plus of course TikTok.
We also have a lot more concentration. Global advertising is now roughly $1tr, Alphabet has a quarter of that, Meta and Amazon combined have another quarter, and (outside China) everyone else is scrambling for scraps. (Note that this means about half of the $300bn that the big four platforms will spend on data centres for AI in 2025 is funded by advertising.) E-commerce is less concentrated globally, though - it’s about $4tr, and maybe $2.25tr outside China, of which Amazon is just over a third, but of course that $4tr is only 25-30% of total ‘core’ retail.
Meanwhile, it used to be that if you wanted to tell customers about your products, you would pay for a store or pay margin to a store, and you bought ads. Now you ask if you should open a store in that country or just advertise there - now you can ask that. Should you put your budget into Instagram or paying for placement at that retailer, or a better returns policy? You wonder if BOPIS lowers your margins and how it compares to a different shipping budget. Marketers and retailers are all native to this stuff now - they call it ‘omnichannel.’ All the old bundles and value chains break up and new ones are created.
And, now there’s this AI thing. Two years ago it was new and exciting, but now everyone’s had a bunch of presentations from WPP, Microsoft, or Bain, depending on what department they work in, and some of the implications are already pretty clear. A lot of the grunt work of advertising and marketing will get automated, from creating dozens of variations of the same ad, to managing media buying, to updating thousands of product description pages. Within the platforms, targeting and placement itself has long been heavily optimised with machine learning, and now they’re selling new LLM-powered optimisation.
What else, though? One of the patterns of new technologies is that we begin by trying to absorb them and to make them fit into the things we already do. That’s bottom-line innovation, but what about top-line innovation? For retailers in particular, chatbots by themselves don’t seem to be a good UX (yet), but unstructured questions (‘what would be good to take on a picnic?’) seem promising, as do review summaries, but that’s still just optimising the current thing - what are the new things that are only possible because of this? We might still be a while before being able to describe the dress you want and see Shein make and ship it to you on-demand, but how far?
The same kind of structural question applies to advertising. There’s been a joke for a while that half of LLMs will be turning three bullet points into an email, and the other half will be summarising emails into three bullet points. What if half the LLMs are turning an ad into a thousand ads, and the other half are looking at all the ads and telling me what shoes to cool right now? What’s the SEO for LLM agents?
More fundamentally, the challenge of infinite products and infinite media is really one of scaling, automation, or time. You can’t possibly look at all of that anymore, so how do you know what you want? But this new wave of AI works by looking at everything, so it seems that AI should change that problem. It will suggest new things, and perhaps suggest new tastes and preferences, or create them. Instagram and then TikTok changed how discovery, taste, and suggestion work in much more complex ways than just the advertising and e-commerce, so if you have machines that see everything and generate new ideas and suggestions in response, is that search and summarisation, or advertising, or retail, or something else?