Benedict's Newsletter: No. 665
How agent-ready is your sales team?
Attio’s State of Agentic Sales report tells you exactly where you stand. We built it from interviews with GTM operators already running agents in their pipeline, and it ends with a readiness audit that turns your score into concrete next steps for your revenue system.
Give every rep the leverage of a hundred, with agents working every account in your market.
Find out where you stand now.
Preview from the Premium edition
Leverage
As we all try to work out what AI will do to different industries, it's obviously useful to look back at the internet, and I think one framing that's useful is to suggest you can split the impact into three.
There were some industries where we all know that the Internet changed everything (recorded music, or retail, say), and there were also industries where the internet really was just a useful chain or tool, and changed very little. If you were in construction or mining, email was great, but you didn't have new internet-enabled mining companies. The interesting part, I think, is in the middle. If you were an airline or a bank, the Internet changed a great deal about how you went to market and presented yourself to your customers, ut ultimately, the business of an airline is renting the planes, buying the fuel, earning the landing slots, and selling the seats, and the business of a bank is making a spread between deposits and loans, and the internet did not fundamentally change what those were or how they were done. It was fundamentally incremental.
PREMIUM SUBSCRIBERS GET THE FULL COLUMN EVERY WEEK. TO READ THE REST, UPGRADE HERE.
News
Wall Street questions
The two fundamental changes in AI this year so far have been firstly the emergence of agentic coding with explosive product-market fit, and secondly that capex has outgrown the cash flows of the hyperscalers, pushing them to the capital markets to raise funds (and giving the capital markets the decision on how far this goes). Indeed, the scale of spending is now a conversation for macroeconomists. That’s a particular issue for the neoclouds, which claim (following Jensen Huang) that they are completely safe investments because the chips underneath have massive demand and act as gold-plated collateral. But if you’re worried that the whole segment could have a price collapse, then that argument doesn’t work. Spreads widen. However, it does strike me that there’s a circular argument here: if Wall Street stops funding the capex because it’s worried there’ll be a glut, then the capex stops and there won’t be a glut 🙃. LINK
Anthropic numbers
Anthropic’s upcoming IPO is starting to look like the big test - is it too big to fail? Revenue (gross) was apparently $16.8bn in Q3, up from $11.5bn in Q2, $4.8bn in Q1 and $4.85bn in all of 2025, and the company claims to have a positive operating profit before SBC, but that slowing growth means the target market cap might pull back from $2tr to $1.5tr. Stepping back, as I’ve said repeatedly, the current unit economics of these companies really aren’t a great guide to what will happen next: we don’t know where supply or demand or model training will go. LINK
The week in AI
Meta released an open source variant of its new Muse model family; however, while the flagship Muse model is now right at the top of the frontier benchmarks, this one is not. Last year Meta’s AI strategy was to use open source to drive commoditisation and turn models into commodity infrastructure where it could build value on the top. The new strategy seems to be more conventional, keeping the model closed and selling API access. I don’t think anyone knows what this will look like in two years though (see the pattern?). LINK
OpenAI claims that it found a bug planted in one of its meeting rooms, and the FBI arrested a Saudi employee. All of these companies are now very, very high value targets. LINK
Amazon has a consumer AI device line called Bee, doing meeting notes. I don’t understand why that would come from Amazon, nor why that isn’t a smartphone feature. LINK
Microsoft pulled out of a UK data centre scheme because it wasn’t able to get power. There will be a lot more of this. LINK
Everything is hacked
Someone hacked the refrigerators in hundreds of supermarkets in France, using a known, unpatched vulnerability, and cranked up the temperature, ruining all the food. Meanwhile, Bank of America says that it’s hired 1,000 contractors to help it patch all of the security issues it’s found this year with AI, and Microsoft’s patch releases are now running at 10x the number of vulnerabilities that it was finding last year. FRIDGES, EXPLOIT, BoA, MICROSOFT
Somebody subverted the supply chain for Ledger’s hardware crypto wallets, adding a very small chip with a key logger and a SIM card and putting them back into the reseller channel, allowing it to steal $100 million or more from people’s balances. LINK
Tesla 2017
Way back in 2017, Tesla showed a ‘Roadster’ supercar, launched the empty shell of one on a SpaceX rocket as a stunt, and took a bunch of $50,000 deposits. The car never actually shipped. This week, Tesla held another event to launch another completely different Roadster supercar. Again, there was a stunt, but a much more impressive one: Tesla rigged a mock-up of the car with a bunch of nitrogen gas thrusters and put on a stage show with this mock-up doing backflips and somersaults. This time it’s taking deposits of $250,000 - again with no promise of when or even whether you’ll ever actually get a car. The company claims that it will eventually sell a version of the car with those thrusters, by invitation only, and obviously not road-legal.
I once described Elon Musk as a bullshitter who delivers, sometimes, and this is the perfect example - a bunch of very cool stunts to promote the relaunch of a product that was promised and never delivered, and this one might never be delivered either - but meanwhile there is the car company and the self-landing rockets. He reminds me a lot of Henry Ford, who was also both a brilliant industrialist and a racist. LINK
Apple’s kitchen computer
I sometimes put things at the end of the news section that I think are both interesting and irrelevant, and Apple’s HomeView, a $300 HomePod mini with a screen fits both. This is Apple’s entry into the kitchen computer segment, which has been a tar pit for 30 years: tech companies keep making things because it fits their strategy and it looks like there’s a space in the market map, but none of them ever get used as much more than a kitchen timer, because the gap on the market map does not line up with a gap on any actual consumers sense of what they need. This one looks the same.
Apple’s pattern is to take spaces that haven’t worked and do them right, but I can’t see that here. The launch video is one of the worst pieces of product marketing I’ve seen in years. This is supposed to tell us why the product is great and how it solves our problems, but here the HomeView makes things worse every time it appears, and every use-case would be done better with a phone. The crowning absurdity is when the teenage girl messages her secret boyfriend that the family doesn’t know about… from the shared screen in the kitchen, surrounded by her family.
Meanwhile, the device itself needs a power cable, so in this video there’s a hole drilled in the middle of the kitchen counter to plug it in. Nothing about this product makes much sense, except as the output from internal strategy and politics. LINK
What matters in tech? What’s going on, what might it mean, and what will happen next?
I’ve spent 25 years analysing mobile, media and technology, and worked in equity research, strategy, consulting and venture capital. I’m now an independent analyst, and I speak and consult on strategy and technology for companies around the world.
Ideas
Local delivery robot vans in China are doing real scale. Meituan did 5.5m deliveries last year and Neolix has 27k units, growing to 50k this year and 100k next year. LINK
Today in silly AI studies, Anthropic published “AI models may be capable of self-replicating without access to a datacenter.” Translation: it managed to get software to run on devices that run software, if you turn off the security (if there is any) - very small AI models, running very slowly. Theoretically, could an evil LLM turn itself into a global distributed system, hacking millions of smart light-switches to spread itself onto them? Sure. Is that enough compute to run such a thing? No. Should we worry about it? No. Should you step away from your computer and chill? Yes. LINK
Mathematician Terence Tao on AI and maths - ‘the end of the heroic era’. LINK
Lots of websites are blocking agents. As I wrote last week, this is partly operational and partly strategic. There are lots of security, fraud and liability reasons why companies don’t want to allow this until they’ve worked out what it might mean and how to manage it, but strategically the question has always been whether you want to allow somebody else to own the customer relationship in exchange for distribution, and often the answer is no. LINK, ANALYSIS
Russia built its own, highly centralised and state-controlled internet ecosystem and its own datacentres, and now that’s a target for Ukraine. LINK
Outside interests
Fascinating NBER economics paper about the role of meetings. LINK
An apartment in a renaissance palace in Blois for sale. Just €470k. LINK
Data
Smartphone unit sales, just like PCs, are down 20% year on year as the AI surge pushes up memory pricing. LINK
ASML says demand for its equipment is ahead of supply for 2027 and 2028, even with 30% annual capacity growth. (Of course, photonics companies said that too.) LINK
Agentic flooding of government services (and also the end of one flavour of middle-class privilege). LINK