Benedict's Newsletter: No. 666

Preview from the Premium edition

What’s the TAM for AI?

This year the hyperscalers will collectively spend around a trillion dollars building AI infrastructure, and estimates for the total build-out, if the music doesn’t stop, are multiples of that. Meanwhile, OpenAI and Anthropic are looking at a combined valuation of something over $3tr, and Anthropic is pushing towards an IPO even bigger than SpaceX.

What is the total addressable market size for all of this – what’s the TAM? It would be nice to know, with that many trillions of dollars at stake, but I don’t think there’s a useful answer. It is useful, though, to think about why.

In principle, you could say that there are three kinds of TAM. The first looks like a good old-fashioned interview question – “what’s the size of the domestic refrigerator market in the USA?” Well, you know how many households there are. You know that refrigerators are replaced every couple of decades or when people move, and you can take a view on the spread of pricing and the ASP. You can look at the housing cycle and the remodelling market, and if you go out and validate all of those you’ll get to a pretty good number.

PREMIUM SUBSCRIBERS GET THE FULL COLUMN EVERY WEEK. TO READ THE REST, UPGRADE HERE.

News

Black October

The warnings about AI and cyber that kicked off with Mythos earlier this year all came home in the last week. PHP, a scripting language used by the majority of sites on the Internet, turned out to have a remote code execution vulnerability that’s been there since at least 2012 – this was found independently by two different researchers using ChatGT and Gemini. Several governments told people to shut down their websites until they could fix it. Meanwhile, an Indian security certificate provider turned out to have been using the password ‘4’… and a whole bunch of other pieces of infrastructure scrambled to fix all sorts of different vulnerabilities.

I’m not a cyber expert, but I can make some observations from the outside. First, there is a general sense that this is going to carry on escalating before it gets better. So far the pace has been set by software companies looking at their own code and by researchers releasing somewhat responsibly (though there is some debate about the PHP release), rather than from people outside using these tools to find the vulnerabilities and exploit them. That might change. Cloudflare reported seeing millions of probing attacks looking for sites with the PHP vulnerability once it was disclosed. Second, people in the long tail who’ve been running their own infrastructure and often haven’t really bothered to keep up with every single security development should be moving to managed services where it’s somebody’s job to keep on top of this stuff (ironically, that’s an argument against coding your own web tools with AI). There will be some spectacular cases where that doesn’t happen something gets badly breached. But third, does this resolve to a new normal of many more attacks on a much wider range of targets by many more people, or does it resolve to a new baseline of a much higher level of presumed security because people have used AI to catch “all” the vulnerabilities that would otherwise have been found bit by bit over the next decade?

Finally, we should presume that the NSA and the Chinese knew about some of these vulnerabilities already and had been quietly using them for a decade. Equally, governments were often getting around encryption by buying zero-day exploits and may now lose that ability. Companies like NSO may be really, really unable to get into the iPhone, and that might paradoxically lead to a policy swing against encryption. BLACK OCTOBER, LINK, CLOUDFLARE

As one outlier, Anthropic proposes that it might now be possible to do formal verification of code security at scale. This seems like wishful thinking. LINK

Anthropic’s S1 bubbles away

Anthropic’s IPO filing continues to be expected any day now, and bits of the draft continue to leak to the press. The Wall Street Journal reported the H1 numbers: the company is breaking even at an operating level before a massive one-off accounting charge, and more relevantly before SBC. I would obviously like to know the unit economics of this company today, but it appears to be very widely established that inference is very profitable and all the cash goes on training the next model, and as I keep saying, I don’t think today’s economics tell us very much about what the economics will look like in two or three years’ time. Today we have shortage-based pricing, and we don’t know how much value of broader use cases in deployment the model companies themselves will be able to capture. There are also some suggestions that the revenue growth has slowed down in recent months - again, that tells us little about the future. LINK

Ken Griffin from Citadel says we’re almost certainly in an AI bubble. I’m inclined to agree (but good luck timing that), and apparently Eric Schmidt owns 1% of Anthropic, which is worth a lot of money, however you cut it. GRIFFIN, SCHMIDT

China on the frontier

I am not very interested in the daily benchmark scores, but I am interested in the strategic shifts within them. Meta and SpaceX dropped off the leaderboards entirely, rebuilt their teams, and are now at the frontier (Meta) or close to it (SpaceX), and now, as was inevitable, China has a frontier model as well, and it’s open source. The latest Kimi model is marginally behind Anthropic and Meta and marginally ahead of OpenAI and Google. As things stand, there is no fundamental barrier to entry or winner-takes-all effect. (This is one of the reasons why discussion of recursive self-improvement has bubbled up: if one lab gets there first, and long enough ahead of anyone else, does that give it a sustainable lead?) LINK

Meanwhile, China issued new draft technology export restrictions, which include a suggestion that Chinese open-weight models might be subject to export restrictions. LINK

Agent economics

Mark Zuckerberg pointed to what he calls ‘clear product-market fit’ for Muse in Meta’s Q3 call: one month after launch, 10m people have set up the app, of which 25% use it on any given day (which is not the same as 25% DAU!), and a quarter of those early users are making e-commerce purchases at an average price of $200. Meta plans to monetise through commissions on this. However, Zuckerberg also said that Muse is so compute-intensive that he can’t justify scaling it until compute costs fall by at least 10x, which he expects to happen ‘in the next year’ (that would match the general market environment for efficiency gains). This is also why the product isn’t available outside of the USA and Canada.

Indeed, one of the unique characteristics of generative AI today is that marginal cost precludes the standard model of the consumer Internet as we’ve known it for the last 30 years. You can’t make the product free, get to scale, and then work out a revenue model. Hence, OpenAI’s Dots needs the $100/month plan (and is being rolled out slowly), xAI’s Grok Bot needs the $30/month plan, and Anthropic’s Duet below (also waitlisted) needs the $17/month plan, and all of these have usage limits.

The other side of these economics, though, is that Meta (and Alphabet) continue to see significant upticks in their monetisation, as they use AI to optimise the ads and the ad targets: Meta also has high hopes for unlocking better use of advertising by SMEs without dedicated teams. LINK

Anthropic joins in

Anthropic bought its own entry to the consumer agent party, ‘Duet’ – unlike the entries from Meta, OpenAI and SpaceX, this one doesn’t have a cute little animated character and there’s a bit more clarity of thinking in the product marketing. I think the idea of passive observation is the most interesting opportunity for consumer agents, and that’s what they lead with. LINK

The week in AI

Sometimes tech strategy is very predictable. Amazon is blocking Muse and other consumer agents, as are a wide number of other companies, in the short term for security, liability and manageability questions, and in the long term because Amazon in particular doesn’t want to give up customer ownership and become a dumb API call for Meta. Walmart takes the other side of the trade-off: it will take the distribution from Meta over the customer ownership, and has a big splashy announcement that it’s supporting agents. The reality, of course, is that as Mark Zuckerberg said, hardly anybody’s actually doing this yet, and it isn’t going to scale soon: all of this stuff is mostly slide-ware. LINK

Anthropic and OpenAI have done a bunch of deals with consultancies and with PE as routes to market, and this week Anthropic extended that to Europe in a deal with HG Capital. LINK

Anduril’s founder, Palmer Lucky, has started a new company called WarHound to create genetically-enhanced dogs for the military, that might be equipped with Anduril technology. This is very Gibsonian. Still, at least it’s not named after something in the Lord of the Rings. LINK

Apple’s shadow regulation

Five years ago, Apple changed the rules on app install ad tracking on the iPhone, making some very pious statements about privacy (as defined by Apple) while coincidentally diverting tens of billions of dollars from Meta and other ad platforms to its own P&L. Now it looks like Apple is doing this again, this time for companies that let advertisers build profiles across different websites based on the email address and phone number you use to sign up for a service. Safari will now block the domains that websites use to do that, if they base it on code on the web page itself.

The underlying question is whether these kinds of policy decisions should be taken by a private company and imposed top-down without debate, recourse, or any discussions of trade-offs: of course, the counter-example is GDPR, where equally simplistic decisions are imposed top-down without debate, recourse, or any discussions of trade-offs. LINK, APPLE

Musk’s foreign policy

The war in Ukraine is now in large part a drone war: tactical drones are controlled over trails of fibre kilometres long, spun out across the landscape like spider webs, but those with longer range, and those that intercept Russian attacks on Ukrainian cities, are mostly connected by Starlink. This week, Starlink suddenly stopped working for anything moving at over a certain speed in Ukraine, and no-one seems to know why: the US government and Donald Trump denied responsibility, and it turned out that (apparently) Elon Musk had made the decision by himself, before reversing it after Trump (apparently) yelled at him.

What is the overlap between people who think it was absolutely right for Anthropic to decide on its own whim what the US government could do with its products, and people who think it is absolutely wrong for Elon Musk to decide on his own whim what Ukraine’s government can do with his products? Why should it be their decision? There is lots of cognitive dissonance on this one: what are your principles, and do you apply them equally to people you like and to white supremacists? Personally, I think it is a bad idea for private individuals to try to control what governments can and can’t do. But whatever your own view, the net result of all of this is most likely to be the accelerated deployment of fully autonomous systems, that don’t have these dependencies. LINK

Ideas

Apollo rounds up a bunch of economic data that reflects what all of the corporate survey data in the first half of this year showed: companies have given everyone chatbots and run a bunch of pilots, but don’t really see any impact on the bottom line from AI yet. The trough of disillusionment from the classic Gartner hype cycle seems to be building just at the point that the financial markets are feeling that this has got out of hand. LINK

Meanwhile, Bloomberg points out that Apollo has a huge AI chip credit business. LINK

Do people make a purchase based on an AI recommendation, or is that just one step at the top of the funnel? It’s still very early, but here’s some survey data. LINK

Dennis Crowley on the end of Foursquare, which is also a requiem for the Internet before it had billions of users and trillion-dollar platforms. The interesting thing to me is that the problem Foursquare aimed at remains unsolved: where do you go for opinions about places? Any product that gives authoritative reviews with a point of view struggles to get enough reviewers of enough places in enough cities to get enough of an audience for the economics to work, and so you open up to user-submitted reviews. Then you can scale, but now all the reviews converge on the mean and to self-selection: everything has four stars. So, then someone starts a great new product with unique, opinionated reviews, but then it needs to scale, and the cycle begins again. I don’t think Dennis really wanted to scale Foursquare, but scaling didn’t scale for anyone. LINK

Bellingcat on how far AI can geolocate a random photograph by using clues in the image – almost as good as human experts, but not quite there yet. LINK

The big agency networks are running frantically to turn themselves into technology consultancies. WPP now has its own agent creative platform and is talking about ‘forward deployed creatives’, channelling Palantir. LINK

With Apple apparently about to launch a touch-screen Mac, here’s a nice round-up of all the times that Apple explained why this was a terrible idea. LINK

Outside interests

Marking the 25th anniversary of the iPod, the 404 page for Apple Music has a fully working mockup of the original model. (Note: this link won’t work on iOS.) LINK

A set of playing cards from 1720, showing 52 of the crazy startup ideas that failed in the South Sea Bubble. Horse insurance, a whale fishery and ‘Puckle’s Machine Gun’. Sold at Sotheby’s this week for $102,400. LINK

Data

The FTC released estimates on how much money is stolen by online scammers from people over 60 in the USA: the reported number for 2025 (much lower than the actual losses) was $3bn, up from $600m in 2020. Rather like call centres, this is a skills arbitrage between low-income and high-income countries, and like call centres, this is going to be completely changed by AI, in this case certainly for the worse. But, perhaps next year we will all set up agents to monitor our parents’ communications and block anything that looks like a scam? LINK