4 May 2025

News

The LLM push for retail and advertising

The global ad industry is about a trillion dollars, and roughly half of that goes to Google, Meta, and Amazon. That’s very tempting if you’re a model-builder with Nvidia bills to pay and a technology that looks like it could tell people what to buy in an entirely new way.

Hence, this week OpenAI added some basic shopping search features, so some queries will get a carousel and links. However, there’s no indication of this in the UI (a systemic problem with the command line as a product), so you’ll just have to guess and see what works. Perplexity said it will build its own browser to collect user data and build an interest graph, and use that amongst other things for ads (step one: get people to switch browsers).

And, Mark Zuckerberg ruffled feathers in ad-land by saying (not for the first time) that one aspiration is that an advertiser can give Meta a budget and a business objective, and Meta’s AI systems can work out the rest - not just the targeting, but generate the best actual ads as well (especially given that it will be able to show every single user a different ads and compare results).

Meanwhile, Visa and Mastercard, both trying to become a consumer choice other than a B2B product, had corp-devy announcements about autonomous payment systems integrated with AI agents (so in theory you could tell ChatGPT to suggest dinner and then buy you the ingredients - good luck getting that actually working though). CHATGPT, PERPLEXITY, META, VISA, MASTERCARD

Apple has another App Store problem

Last week the EU fined Apple €500m because it doesn’t like how Apple implemented the new rule that says it must allow iOS developers to use third-party payment, but didn’t say what Apple should do to comply.

This week a US judge was much more explicit. In 2021 Apple beat Epic Games in court, with a judge ruling that Apple didn’t have to open up the platform entirely, as Epic wanted. However, she did say that Apple has to let developers use third-party payments and has to let them tell users about this. This week she decided that Apple put so much friction and restrictions in the way of this that Apple has effectively violated the 2021 order and is in contempt. For a bonus, she found that an Apple finance exec perjured himself. Oops. 

I’ve always thought that Apple was basically right on the principle that it controlled what developers could do with your phone, and increasingly wrong on how it handled the payment question - the original idea might have been good for users and developers (frictionless trusted payment!), but Apple has treated every exception and corner-case like Paulie in Goodfellas - “f*ck you, pay me.” And it’s reacted to judges and regulators pushing back on this by slow-walking. That might not last much longer. 

However, as I’ve written about other regulatory issues, some people feel very strongly about this, but it’s hard to see major structural change even if Apple keeps losing. The payment rule is theoretically a big deal for Spotify but it’s doing just fine regardless. It’s margin for games developers, but if it goes away half of them will keep using Apple payment anyway (better conversion) and much of the rest of the money will go straight back into acquisition ads on Meta and, um, Apple. And Meta would like to run a game store inside Facebook, but do developers want that? So, much sound and fury, but… what big, real things would change if this rule is finally gone? Does OpenAI care? RULING, COMMENT

Stripe was straight in with an alternative payment option. LINK

Meta goes direct with Llama

Meta wants more distribution for its generative AI model. It’s already added it to the search boxes in its apps, and then added a floating blue circle you could tap on (did anyone know what that meant?), but now it’s gone all the way and added its own website and smartphone app, ‘Meta AI’. This is a pretty standard move, but it does point to the way ChatGPT dominates consumer consciousness - it’s been at the top of the app store for a year when the apps for other big model-builders (Gemini, Copilot, Claude, and stretching a point Perplexity) struggle even to break the top 100, and you can see the same in Google Trends. There’s a huge difference between Silicon Valley insider mind-share and consumer mind-share, and bolting Llama into Instagram might not have worked as well for this as it did for Reels. This new app will have the usual Meta growth playbook behind it (Threads now has over 300m users) and it has some interesting social ideas, but is that enough?

In parallel, Meta now offers developers API access to its own cloud-hosted instance of Llama, Meta’s underlying model family, and reportedly also asked Microsoft and Amazon to help fund development - Meta would like some money for this, please. But that didn’t stop it increasing the 2025 capex plan from “$60-65bn” (December 2024) to “$64-72bn.” APP, API, FUNDING

An interesting side-note: WhatsApp is working on a way for you to use the cloud LLM with anonymous encryption, so that Meta can’t see what you’re asking. LINK

The week in AI

Deepseek got all the buzz, but Alibaba’s Qwen keeps delivering good model with good benchmark scores (though not topping LMArena). LINK

OpenAI rolled back an update to its 4o model that people complained was too ’sycophantic’. Even if you think it’s a good idea to anthropomorphise these models (I hate it), they are deep in the uncanny valley. LINK

Apparently Fox has done a deal with Runway AI for video generation. Of note - the output is now rights-cleared for deliverables. LINK

On the same theme, Clova is a new AI video editing assistant - find the right footage and put together a rough cut more quickly. LINK

Sam Altman’s Orb, which wants to scan your retina to generate a digital ID (on a blockchain) you can prove you’re a human online, has finally launched in the USA. LINK

Elon Musk’s xAI built an LLM data center in Memphis in record time - partly, it has become clear, by deploying gas turbines for power without bothering to get pollution permits or filters. There is a strong case that the US in general makes it far too hard, slow and expensive to build anything, but… LINK

Bloomberg reports that Apple is working with Anthropic on an AI coding assistant. Apple talked about doing this in Xcode at WWDC last summer, but that hasn’t shipped yet. LINK

Amazon’s satellites

Amazon launched 27 LEO satellites into orbit for its ‘Project Kuiper’ competitor to Starlink for niche broadband service (not using Jeff Bezos’s Blue Origin rockets, though). I confess I don’t really understand the business logic behind Amazon investing in this. LINK

A US revenge-porn bill?

The US passed the ‘Take It Down’ act, which now goes to Trump for signature or veto. In theory it’s hard to argue with the intent, though there are plenty of arguments for unintended consequences (partly the scope for selective enforcement, partly argument that the terms are too aggressive for practical implementation). LINK

Everything is hacked, UK retailer edition

Several large UK retailers have had major hacking incidents this week: Marks & Spencer had multiple systems down and couldn’t manage inventory or take card payment for most of the week, the Co-op has the same and also had user data stolen, and Harrods, a large shop in London that sells souvenirs to tourists, reported a breach but gave no details. MARKS, CO-OP, HARRODS

EU fines TikTok

The EU fined TikTok €530m for breaking data privacy by transferring user data to China. Waiting for people in Silicon Valley who hate both TikTok and the EU to square this one… LINK

Tech tariff news

Trump attacked Amazon after a (untrue) rumour that it would show the cost of his tariffs next to product prices. Why not, though? It shows sales taxes, after all. LINK

Apple said that for US consumers almost everything except iPhones would now be coming from plants in Vietnam, and confirmed that about half of US iPhone sales now come from India. (This won’t bring manufacturing back to the USA.) It also said that given today’s tariffs, the cost impact would be about $900m next quarter - about 1% of revenue and 4% of net income. LINK

Meta reported strong growth and pointed out that given the auction model for ad sales, if one set of advertisers withdraw (i.e. Temu and Shein) that doesn’t have 1:1 impact on revenue, since the next lowest bidder takes up the inventory (presuming no broader recession, of course). LINK

Ideas

As the Google TAC remedy trial hears arguments, Biden’s (now ex) DoJ head, Jonathan Kanter, argues for breaking up Google. If the only tool you have is a hammer… LINK

How to fend off fake remote IT workers from North Korea - ask them if Kim Jong Un is fat. LINK

Ukraine’s drone industry. LINK

The impact of generative AI on stock photos. This sounds a lot like what photography did to professional illustration in the 1960s. LINK

More notes on food delivery consolidation. LINK

Outside interests

Why archers didn’t use volleys (and didn’t ‘fire’ either, or course). LINK

Data

The IAB’s 2025 video ad report. LINK

Waymo released more safety statistics. LINK

Column

Monopolies past and future

There's a funny sort of unreality in looking at headlines and arguments about regulation and prosecution of Apple, Google or Meta today: these are discussions about things that happened 10 and 15 years ago when large language models were very clearly setting a completely new platform and a completely new set of market dynamics for the next 10 or 15 years. 

That applies even more if, like me, you tend to think that most of the changes on the table, even really eye-catching ones like splitting off Chrome, aren't really going to change very much anyway. It all seems desperately irrelevant.  Maybe an independent Instagram would have become a global giant, or maybe it would have ended up like Snap - we can’t know, and what do we gain from arguing about it? Generals are always fighting the last war, and regulators are always fighting the last platform battle: the rest of us should move on. 

But if we step back and look at some of the stories in this week’s newsletter, they're very clearly setting up the antitrust cases and EU regulations of 2035 and 2040. Meta, Google, Microsoft and Apple are bundling and embedding their LLMs on every available surface in their products, and they are thinking desperately how to leverage whatever distribution they can get their hands on to build critical mass for whatever this is (and no-one really knows). And that’s exactly how it worked when Google and Meta were looking over their shoulders at YouTube, Instagram and WhatsApp, and a generation earlier how Microsoft looked at web browsers. “Of course we should do this!” Of course Windows should bundle TCP/IP! (Yes, that became an argument.) Of course a PC should come with a web browser! Of course Google should add AI Overviews to search, and not leave that to Perplexity! 

It seems to me that the impetus to go for distribution is even stronger because it’s not clear what else there is to compete on. Right now, the foundation models themselves seem pretty much like commodities - there are differences in emphasis, but are you confident you could tell the difference in the result for a generic query between Claude or Gemini? The same applies to the products themselves - there are only so many ways you can wrap a product around a chatbot. Indeed, I sometimes think that the real ’thin wrappers’ are ChatGPT itself, plus Claude, Gemini and so on - they’re a few buttons and a lot of graphic design on top of a command line. So, if you’re selling a commodity, you go for bundling,  leverage, branding and distribution. Stick it on the home page and make the button bigger, with more sparkles.  

On the other hand, Google was never able to bundle and leverage into search (and Google Video failed). More interesting - while Microsoft succeeded in using bundling to make Internet Explorer the dominant browser for a while, that didn’t stop it losing the dev environment to the web and web servers to LAMP. Browsers turned out not to be the key point of leverage for the internet. The value capture was in search advertising and social media, which didn’t exist until 5 and 10 years after Microsoft decided to bundle IE in Windows. 

We're at the same point now. We know LLMs are The Thing but we don’t know how that works. ChatGPT has pretty dominant consumer mind share, but then, so did AOL and Yahoo. I am very skeptical that the ‘blank screen’ LLM as a chatbot is the right UI, but even if it is, is it even the right question? When I was a baby analyst covering mobile in 2000, investors kept asking me ‘what’s the killer app for 3G?’ Maybe it was video calls? Games? It turned out that was the wrong question - the killer app for ‘having the internet in your pocket’ was, well, having the internet in your pocket. Some of the more interesting parts of the Meta disclosure in the last couple of weeks have been the stuff that never happened. Mark Zuckerberg was worried about Instagram, but he was also worried about Path (I miss Path). The same this time - it will all seem obvious in hindsight. 

Benedict Evans