23 November 2025
News
Google goes back to the top of the leaderboards
I’m old enough to remember when Google was doomed, search ads were doomed, and Sundar Pichai was the new Steve Balmer. Not any more. Gemini 3 is SOTA and the new version of the ‘Nano Banana’ image generator has gone viral. I struggle to get particularly excited beyond that: there will be another SOTA model in a month or two, and one of these is really changing the trajectory. What does interest me is that the images going viral aren’t the usual cheesy stock images - instead, people are feeding in videos, audio or slides, or just asking questions, and getting detailed diagrams and infographics. How long will it be before you ask for a recipe or instructions and the chatbot will make you an interactive animation or video explainer? GEMINI, NANO BANANA
OpenAI Ads coming soon
OpenAI is getting a bit more forthcoming about ads. It has intent, and wants to have commerce flows (see the app platform announced recently, and now there’s a deal with Target): once it has that, ads are inevitable. Note that while it claims 800m weekly active users, apparently only about 5% are paying. LINK
Nvidia keeps doing it
Nvidia’s latest quarterly results show demand isn’t slowing down yet: revenue was $57bn versus $35bn a year ago. The stock is still down: the markets are getting more worried about bubbles and circular trades. See this week’s column. LINK
The FTC loses again
The FTC has lost yet another big tech breakup case, this time against Meta, failing to undo the acquisitions of Instagram and WhatsApp.
This is the cast that the judge originally threw out because the FTC just forgot to say what market it was claiming Meta has a monopoly of or what share it claimed it had of that: he allowed it to refile, and the FTC then claimed that Meta does not compete with TikTok, YouTube, or iMessage, but does compete with something called MeWe, which no one has ever heard of. Market definition is the basis of any competition case, and this was laughable - gerrymandering, not analysis. Now the judge has agreed, saying that events in the market since the WhatsApp a decade ago show how dynamic and competitive it remains.
It would be easy but also true to say that the antitrust crusade against Big Tech, spearheaded by Lina Khan, now looks like a failure. I have a lot of sympathy with the ‘neo-Brandeis’ idea that US competition law, unlike the UK or EU, was too narrowly focused on consumer prices as the only measure of harm: a free monopoly could also be bad. But this wasn’t accompanied by rigorous analysis of the actual market structures they wanted to change, nor a clear sense of what a good remedy would be (the problem in the Google search case). So, the ‘landmark’ 2020 House Antitrust Committee report was riddled with basic errors: for example, it claims there had been a ‘sharp decline’ in startup creation and funding, citing as source a report that uses data ending in 2011 - after which the US experienced the hottest market in startup creation in its history. Meanwhile, Lina Khan’s famous and career-making paper on Amazon did not contain the word ‘cash’ once. The more interesting problem, though, might be that revising price as the only test is not enough - current models of market definition and ‘dominance’ might also be too narrow. COVERAGE, RULING
The week in AI
Filling in the boxes: last week Anthropic announced a plan to build $50bn of its own infrastructure; this week, it’s done a deal to buy $30bn of compute plus one GW from Azure, while Nvidia and Microsoft will invest up to $10bn and $5bn. LINK
As reported last week, Yann LeCun confirmed that he will leave Meta and do a startup focused on his projects to move AI research beyond LLMs. LINK
The White House is still looking at trying to ban US states from passing their own AI acts, on the basis that any U AI company would have to spend a lot of time and money complying with dozens of slightly different rules, instead of one set of national rules (which, however, do not yet exist). LINK
Adobe is acquiring SEMrush, a hot company in the world of trying to work out where your brand shows up in ChatGPT, for $1.9bn. LINK
Google hired a former CTO of Boston Robotics, the famous robot dog company. LINK
Google’s head of AI infrastructure said that the company needs to think in terms of doubling compute capacity every six months, and 100x over the next five years, without increasing cost or power budgets. LINK
Death to cookie banners?
The EU has a new bill aiming to rationalise and simplify all the digital regulation it’s introduced recently, resolving overlaps, trying to add more certainty to outcomes, and (more controversially) relaxing rules around things like cookies (by expanding the allowable use cases and allowing automated consent at the browser level). You can tell this is a good thing by seeing who gets angry about it. LINK
WhatsApp’s oops
A security researcher discovered that WhatsApp wasn’t rate-limiting the feature that matches phone numbers in your address book against accounts, allowing anyone to scrape a couple of billion profile names. This is public data, but still not ideal. LINK
Foreign actors
Twitter turned on a feature letting you see where an account is based (derived from the IP address used to post), and unsurprisingly a lot of high-profile pro-Trump influencer accounts are not based in the USA - they’re in Pakistan or Nigeria, and probably only in it for the revenue shares. (This is a general problem, but Twitter fired most of its platform integrity team). LINK
RIP Ocado?
Ocado pioneered online grocery delivery, creating and building cutting-edge robotic fulfilment centres, then pivoted to selling the technology. But this week Kroger said it will put down three of its eight facilities that use the tech (it shut another three last year), taking a $2.6bn charge. The question was always to work out where the economics and density supported picking in-store versus ‘dark stores’ versus robot warehouses, and the robots are hardest. Ocado has a market cap of $2bn today versus Instacart’s $10bn - another tough model. LINK
Ideas
US AI companies look for users in India and SE Asia. LINK
S&P on how the global electricity generation industry does not have capacity to meet demand from AI. People are booked up through 2030. LINK
Almost six years ago I saw a presentation on Gaia-X, a European project to replace hyperscalers by publishing a lot of PDFs describing new standards for how hyperscale data centres should work. Yes, this was almost clinically delusional. They’ve only just given up. LINK
More on that LA AI podcast farm churning out 3000 episodes a week with eight employees. LINK
Pinterest posted a long piece on their experience building ML systems over the last decade. LINK
China is looking at limiting default performance for EVs, since electric torque means that basically any car can have the acceleration of a supercar, and that’s dangerous. LINK
Outside interests
A Polignac wardrobe. LINK
The Ohtake Residence. LINK
And the Stahl House (you know this one). LINK
Everyone in LA wants a giant door. LINK
Data
Interesting Edelman global survey on consumer attitudes to AI. Too many points to summarise - well worth reading. LINK
Google and Bain’s annual report on tech in Southeast Asia. LINK
Atomico’s annual report on the state of European tech. LINK
There is now a ‘Lumascape’ market map for AI. LINK
Another decent survey on LLM use in the use, this time split by generation and use case - about 20% of US teenagers have used this to write a report so far. LINK
Column
In a bubble…
My new presentation has two slides explicitly about bubbles. The first points out patterns. In a bubble, people draw smooth lines on log-scale charts and convince themselves that this has predictive value. They say things like ‘you don’t understand exponential growth.’ I’ve heard this one far too often on podcasts with researchers from AI labs in the last few months, none of whom seem to remember that exponential growth generally begins and ends with a curve. A couple of years ago, a bunch of crypto people made both of these mistakes and made very compelling charts that ‘proved’ hundreds of billions of people would be using Bitcoin by 2025.
Most of all, of course, in a bubble, people say ‘this time is different’. The trouble is, it always is different. The Dotcom bubble was different too. And it can be both different and a bubble. Indeed, everything changing, right now, in really dramatic ways, is what creates bubbles. People forget that a good company and the right price are not the same thing: a company or a technology can be absolutely perfect and still too expensive. The fact that AI is working really, really well and has no signs of hitting a wall and will change the world… does NOT mean that there cannot also be a bubble in AI.
Of course, that doesn’t mean we’re in 1999. We might be in 1997. We’ll find out.
The second slide talks about the three companies at the centre of the discussion about ‘circular revenue’ - Nvidia, OpenAI, and Oracle. I think it’s useful to ask what you would do as a rational actor if you were the CEO.
So: Nvidia had $77bn of free cash flow in the last 12 months. It’s still behind demand, but it can’t get TSMC to ramp up any faster (TSMC, presumably, is telling itself that semiconductors are a cyclical industry). It has a big lead in AI accelerator design, and in the CUDA software ecosystem it’s built on top, but as Andy Grove told us, only the paranoid survive. So what should it do with that money? Well, use it to build moremarket dominance, fund more growth, and drive more lock-in, more market share, and more FOMO. That might inflate the bubble further, but what else should they do?
OpenAI has massive mindshare and stock that almost every investor wants to buy. But foundation models remain commodities (expensive, difficult commodities but still commodities) and don’t yet have network effects or any other winner-takes-all effect. Meanwhile, OpenAI does not yet have meaningful infrastructure of its own; on one side, and on the other it does not have any distribution besides the models themselves (although those have 800m WAUs), whereas Google and Meta (and Microsoft) have used their existing surfaces to drive very rapid consumer adoption of their AI products. And, it remains very unclear what the product itself should really be. The whole thing could just slip away from you. So, you use that mindshare and inflated stock to buy market position, lock-in, BD deals, infra deals, platform partnerships, and anything else you can think of (or that ChatGPT suggests). Everything, everywhere yesterday, before the music stops.
Oracle is the anti-OpenAI - the joke in the Valley was that no-one even knows what it does, except build AI infra. It has a legacy business that hasn’t been exciting since the 90s, and it’s been losing markets share to could and unbundlers of every kind for a long time. But it's also very cash generative, and Larry Ellison wants o be a player again. So, what do you do? Gear up and burn your way into the new thing! I’ve seen research suggesting capex needs might be 100% of revenue, and apparently there’s a lot of hedge-fund activity treating the stock as a proxy for an AI crash.
Back to 1997: no one knows how this is all going to work. There are capital markets stories (even more for Coreweave and the other neoclouds), and chip stories, but the core I keep coming back to is the level of uncertainty around the actual applications - remember Yahoo, and Netscape, and Pointcast? i-mode and Blackberry? And meanwhile, Google didn’t exist yet and Mark Zuckerberg was 13. Bubble or no bubble, no one knows anything.