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.

Record numbers, falling shares: the AI capex wobble

Mads started with the numbers. Anthropic’s investor day showed nearly $17bn of Q3 revenue, up 45% on the quarter. ASML reported nearly €12bn, up 55% on a year ago, so Anthropic now out-earns ASML. TSMC had another record quarter, up 41%.

And yet. Samsung fell 9.5%, SK Hynix 11% and SoftBank over 9%. Pri pinned it on two triggers: Broadcom’s CEO saying many AI investors are behaving like they’re in a casino, and hotter-than-expected US inflation. Gartner added that 80% of neoclouds will be gone by 2030. At these valuations, Pri said, great results get punished if they aren’t perfect.

Mads questioned the profit. Anthropic’s adjusted profit leaves out stock-based comp, and labs are handing out stock packages “like there’s no tomorrow” to hire AI engineers. Count the dilution and the profit disappears. Growth is slowing too: 45% this quarter, down from 140% the one before. Token use keeps rising, but prices may fall faster. If lab revenue can compound at around 50% a year for three or four more years, semiconductors still have plenty of money to make. If not, the neoclouds are the soft underbelly.

Pri’s split: the picks and shovels at the top (TSMC, ASML) and the labs with real revenue look fine. The risk sits in the debt-funded middle, data centres renting out GPUs.

The hot take: I agree. The market is pricing AI by who actually makes money from it. That favours founders saving real costs in real industries, not businesses whose whole model is renting out compute.

Nvidia on your cap table

Jensen Huang was in London this week. Nvidia has put around $8bn into Europe, and in the UK it picked five companies including Boltz, Foresight, Hiverge and Prima Mente.

Pri flagged the tension: Nvidia can be your investor, your supplier and your competitor’s investor at once. It backs Wayve, Synthesia, these five and four French companies.

Mads wasn’t worried. The cheques are rounding errors for Nvidia, a cash version of the free credits AWS and Google Cloud used to hand out. And it can’t get too predatory without the industry ganging up on it. His advice: never let a strategic investor control your roadmap.

My view: the sheer volume of deals protects founders. Nobody in particular is captive.

Everyone is building the whole AI stack

Chipmakers are moving into models, cloud firms into chips, model labs into data centres. Meta says it has reached the frontier with Wonder. Google shipped Gemini 4 Argon. Microsoft announced its Maia 300 chip. Nvidia is in talks to buy Reflection AI.

We ran Argon on SuperSeed’s own work. Mads’s verdict: it competes but doesn’t necessarily beat what we use, and it didn’t replace a single model.

Pri said Meta is picking a different lane. OpenAI and Anthropic go after work, Meta goes after your personal life, with agents that order groceries and book travel. But Amazon has already blocked Muse agents from shopping on its site. And trust is a problem when Meta has gone closed, cut 8,000 jobs and wants to track employees’ keystrokes to train AI.

On the Artificial Analysis index, Wonder is roughly level with the latest OpenAI and Google models. Four or five frontier models now sit within a few points of each other, so the difference is price and fit.

Physical AI, and China in the supply chain

LLMs could hoover up the internet. World models need real-world data. (My theory: that’s why everyone is releasing smart glasses.)

Atomic Machines came out of six years of stealth with $250m, led by Jeff Holden (Amazon Prime, Uber’s first chief product officer). Its Matter Compiler builds tiny working machines straight from code, from design to device in hours. The first product is a tiny power switch for AI data centres.

Agility Robotics is listing at $2.5bn via a SPAC. Pri noted no disclosed revenue, about $100m burned last year, and $300m of orders from one unnamed customer.

On China, Mads argued it makes trade impossible because it doesn’t really want to buy from the West, and weak domestic consumption pushes its surplus overseas. Founders need a plan B: what if you can’t buy a part next month, or a 40-50% tariff lands? France and Germany are discussing a kill switch for the single market, “one step further than Trump”.

I put China’s cost edge down to state subsidy. Mads called that “a cheap excuse”: it’s just much cheaper to make a car in China. Pri’s answer is for European founders to own what China can’t easily copy: the customer relationship, the data and trust with European industry.

The EU-China hybrid deal points the same way. The cap is 350,000 a year from December, against around 800,000 sold this year, halving China’s share from 6% to 3%. Cars built in Hungary or Spain don’t count, so Pri expects Chinese factories in Europe rather than fewer Chinese cars.

EU Inc, Project Europe and September’s funding numbers

Mads ran through the EU Inc cuts: the central register is now just an online front door to national registers, the insolvency fast track looks gone, the start date slips by a few years, and stock options remain in limbo. He still thinks it’s worth passing. I worry we’re near the point where it stops mattering for founders.

Project Europe’s founders, Kitty and Jade, are handing over to a new team. The $10m Thiel Fellowship-style fund backed 25 companies with 40 founders across 22 countries. Harry Stebbings framed it around a “doom loop” of talent leaving for the US.

September funding hit nearly €16bn, 5x a dead August. Mistral’s €3bn and cloud and compute (about 35%) did much of the work, but 25 companies raised over €100m. And SAP buying TechWolf after investing and partnering shows the realistic European exit.

Can anything IPO right now?

Firmus pulled its IPO last week, after Oura, Holtec, Bamboo and Clear Street. Only giants like SpaceX and SK Hynix got out.

Mads thinks listings priced to perfection have burnt investors, who want to be up about 10% after six months. Which puts a lot on Anthropic.

Predictions

• Me, on Anthropic’s IPO: Strong out of the gate, then down 50% over the next three months. “I think it’s gonna be difficult.”
• Mads, on Anthropic’s IPO: He still expects it before Thanksgiving, at “that sort of two trillion dollar price tag or, you know, give or take”. If it’s a difficult one, “that’s gonna sour markets for sure.”
• Mads, on Tesla’s Q3 results (Wednesday 21st): He thinks CyberCab probably has more legs than Optimus, so watch for robotaxi disclosures.

Deals of the Week

• Extend Robotics (Pri’s pick): A UK startup that raised a $3.3m seed. Instead of selling factories an expensive machine, it sells them the finished work. Customers pay per result and Extend takes the operational risk, which removes the upfront capex that stops most UK manufacturers from automating. It already has 35 paying customers across automotive, aerospace, nuclear and agriculture, from Leyland Trucks to a UK apple grower using a humanoid for quality control.
• Rabot Energy (my pick): A German company that raised a $50m Series B to turn homes into virtual power plants. Before I became a VC, I looked at starting a micro-generation company. I like the idea of your car and your home being part of the grid, and I don’t think we’ve done nearly enough there.

Notable Quotes

“I think what’s happening is the market has stopped asking, like, is AI real? And started asking who actually gets paid and when.” – Priyanka Savjani

“They put twenty, thirty, forty million into a company. It doesn’t even register.” – Mads Jensen

“I think the risk there is you might be waiting forever. And sometimes it’s better to go incremental and start with something and then try to build on top of that.” – Mads Jensen

“I don’t think it’s the job of Startups to play geopolitics. I mean, that’s not that’s not why they’re here.” – Mads Jensen

“And if your agent needs someone else’s platform to be able to work and that platform can switch you off, then you may not have a business.” – Priyanka Savjani

Why This Matters

If you’re raising for an AI company in Europe, the bar has moved. Investors will care less about having AI in the deck and more about who pays you, how much, and whether your margins hold. Neoclouds and anything else financed with debt in the middle of the stack will feel it first. Hardware founders need a supply-chain map and a plan B before the tariff arrives, not after. Anyone hoping for an IPO exit should watch Anthropic’s debut closely, because it will set the tone for everyone else.

Build the thing someone pays for. Everything else is a lease.