Upside is a weekly podcast designed to look behind the headlines that will affect European venture, startups and investing.

Below are the notes from this week’s episode. Episode links above to tune in and stream wherever you pod.

Winning is losing

Start with the scoreboard, because it barely makes sense. The FTSE 100 touched a record intraday high, the index with none of the big chip makers hitting its peak more or less because it owns none of them. Across the water the Nasdaq 100 closed in correction on Wednesday, down more than 11% from its June top. The SOX, the semiconductor index, was in a bear market, 25% off its high. The Dow rallied 500 points on a rotation out of the very thing that had been carrying everything.

Lomax said the quiet part. The UK gets to celebrate its main index as a good hedge against AI precisely because it is oil, gas, banks and pharma, less than 1% tech, up around 9.7% on the year. A win you take, if you insist on seeing the bright side, but a strange one to be proud of.

Then the earnings. Five big AI spenders reported: Alphabet, Microsoft, Meta, Amazon and Apple. All of them, Apple aside, spend like emperors on AI hardware, all of them said the spend is going up, and the market treated them nothing alike. Microsoft jumped on a phenomenal backlog. Meta is down on the year. Apple is quietly the best performer, up around 15%, with Amazon right behind after Andy Jassy said on the call, not in the printed remarks, that AWS could become a trillion-dollar business. Sit with that number.

Mads’ sharpest point was on the two the market is arguably mispricing. Meta is brilliant at using AI on itself, optimising ads and its own platform, and has no real model for selling it to anyone else, so the spend reads as pure cost. Microsoft gets credit for its backlog and an enterprise sales force that is possibly the best in the world. But look at what it sells: other people’s models and Nvidia’s GPUs, and that eats the margin. Google makes its own models and its own TPUs. As Mads put it, distribution is destiny, and Microsoft’s is formidable, but over time he is more excited about Google’s business.

The purest winning-is-losing story was the hedge fund. Situational Awareness, run by a 24-year-old German who went from FTX to OpenAI to a 167-page AI-supercycle essay, was up over 400% by the end of June on leveraged long-AI, short-old-SaaS bets. Then the memory stocks wobbled and the wheels came off. Ken Griffin’s Citadel swooped in and bought the book, probably at a tasty price. Lomax reached for Charlie Munger on the three routes to ruin: liquor, ladies and leverage. When you are that levered, the amplification runs both ways.

The memory IPO that makes no sense

The one that made everyone’s eyes go wide was CXMT. China’s memory champion, founded in 2016, listed in Shanghai on Monday and closed up 466%, briefly worth more than Intel. It raised $8.5bn, the biggest mainland chip listing since Agricultural Bank of China in 2010, valued at roughly 1,600 times earnings, with the retail tranche 212 times oversubscribed.

Here is the catch. Everybody is excited about memory because of AI, and this is not that memory. CXMT makes good old-school DRAM, the stuff in your laptop, not the high-bandwidth memory the AI build-out is starving for. HBM is barely in the business plan, and they are several generations behind SK Hynix and Samsung.

So why the moonshot? Only a tenth of the stock floated. You have capital controls, a domestic market full of people who want an AI trade and cannot buy Hynix or Samsung, and then a local memory name lists. The retail money floods in and bids it to something wacky. Lomax added the smarter-money version: some traders are betting CXMT eventually cracks HBM and becomes China’s sovereign memory player. A bet on the future, priced at 1,600x earnings. Good luck.

Trump bans the robots (sort of)

The headline said Trump banned Chinese humanoids. The reality: the US moved to ban imports of foreign-made humanoids and power inverters on national security grounds, clearly aimed at China, where about 85% of the market sits, with a carve-out for units already in the country. Dan used to file this as protectionism dressed up as security, until he read the reports naming specific back doors in specific tooling. The Huawei playbook, except the detail is hard to wave away.

The venture angle is where it bites. The fastest way to build a hardware startup is to fly to Shenzhen and buy your supply chain, or buy a dead humanoid for £40k-£50k plus VAT and program it yourself. The ban shuts that door early. Sell to the UK government, the MOD or the Ministry of Justice and Chinese parts are already a non-starter, and some SuperSeed portfolio companies have been pushed to an entirely UK supply chain. Andrew’s warning cut through the sovereignty cheerleading: if a European company threatened a US champion, he does not think Washington would blink before switching off a critical part.

A crack in Claude’s cryptography

First, the number that should stop you. Nvidia is backstopping a roughly $250bn slice of an OpenAI campus in Ohio, part of a project Mads put north of $500bn. He could not recall a corporate guarantee at that scale. And Nvidia wins either way: open or closed models, the chips still get bought, and the top-end kit is replaced every two to three years, so the demand recurs.

Now the one to remember. Everyone colloquially knows quantum computers will eventually break today’s encryption, maybe a decade out, which would open up banking, messaging and government traffic. So NIST runs an open competition to pick the replacements. Cryptographers submit new schemes, the world tries to break them, and the survivors end up inside every browser and every bank.

HAWK was a survivor. Two years in, nobody had broken it. Then someone who is not even a cryptography specialist sat down with Anthropic’s newest model and, in about 60 hours and $100k of compute, halved its effective key size. Halve the key and you halve the wall an attacker has to climb. The scheme has been withdrawn.

Why it matters beyond the security crowd: these schemes need to survive for decades, because of harvest-now-decrypt-later. Someone can hoover up your encrypted traffic today and read it the moment the codes fall. A scheme that dies two years in is worse than useless. The uncomfortable subtext is that the same models sold as productivity tools are getting good enough to knock over defences that took experts years to build.

That fed a wider AI conversation, and Mads’ verdict on the new pricing is worth having. Opus 5 shipped at half the price of Fable and matches or beats it on some benchmarks. Counterintuitive, until you realise it is very good on narrower tasks but needs handling differently, so you may have to rewrite your scaffolding and skills to use it well. Call it a learning tax. Opus is the workhorse. Fable is what you reach for when the problem is genuinely hard. Andrew is less convinced the big leaps are still coming, seeing gains now in how models use existing smarts (longer-horizon coding, bigger codebases, fewer hallucinations) rather than raw intelligence. Mads pushed back: Fable, barely a month old, was a real step change, and the release cycle has compressed to one or two months.

The engineering wall, and the paperclip in the code

Mads’ best riff was on why all this capability is not yet showing up as automated work. He pointed at a remote-labour index tracking how many roles AI can actually do end to end, and it is far lower than the hype implies, with Fable topping it at about 13% of completed roles. The gap between what the models can do and what they do in the wild is not research. It is engineering: data in the wrong format, systems not configured, pipelines not wired up.

Then the story that stuck. SuperSeed has been building software factories, autonomous graphs that write code, check it, security-test it and deploy it. In one run the checkers found bugs, raised tickets, spawned more checks that found more bugs, and it cascaded. The system got so good at finding things to fix that new code outpaced the pipeline’s ability to ship it, and the whole thing clogged into hundreds of competing issues. Mads reached for the paperclip parable, the AI that consumes the universe pursuing the goal you set it. This was a tiny, real version. The capability is here. The plumbing is not.

Predictions

Andrew Scott: RSA-2048 will be cracked this decade, and he has said 2028. The old consensus was you would need one to two million qubits given current error correction. Andrew thinks the better architectures bring that down to the hundreds of thousands, so we are edging closer than the timelines assume.

Mads Jensen: the Anthropic IPO news could land soon. No date yet, but with revenue he believes is still growing exponentially, he would not be surprised to see something drop in the coming weeks.

Deals of the Week

Mads: a nuclear turbine spin-out, £15m seed. A BAE Systems spin-out from Manchester, on-brand for the Burnham era, built via BAE’s launchpad. Founder Jeremy Oston spent his career on maritime submarine machinery. The twist: they delete steam entirely, heating compressed air directly into a gas turbine, the kind you find in a jet engine. Small for nuclear because there is real R&D still to prove, but it could be a genuine route to small reactors that work.

Dan: Humanoid, a $152m Series A. The UK robotics startup hit unicorn status at a $1.35bn valuation. A very large early-stage round for a European hardware company, which is rather the theme of the week.

Lomax: Aeon, a $30m seed led by Northzone. A UK defence startup building synthetic environments for autonomous systems to train in. Lomax did the pre-seed.

Andrew: ZeroQ, a $25m seed. A quantum startup claiming a new architecture that scales chips faster. Andrew’s real point: Europe keeps piling money into quantum, and he wants European governments writing big cheques as customers, not just watching a global winner emerge and buying it later.

Notable Quotes

“Distribution is destiny. If you have the right sales force, that can sometimes trump having the better technology.” – Mads Jensen, on why Microsoft outshines Google in the market’s eyes.

“Everybody’s excited by memory because AI. That’s not the stuff these guys make.” – Mads Jensen, on the CXMT listing.

“Some chap who isn’t even a cryptography specialist sat down with Anthropic’s new model, and in 60 hours and about a hundred thousand dollars of compute, he halved the effective key size.” – Mads Jensen, on the HAWK break.

“The factory became so good at identifying issues it wanted to solve that it completely clogged up and fanned out.” – Mads Jensen, on the paperclip moment in a real software pipeline.