Upside 115 – EU Inc’s 28th Regime, the US Midterms, and Anthropic’s IPO

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.

I was away this week, so Lomax Ward of Outsized Ventures took the chair for Upside episode 115. With him were my SuperSeed colleagues Mads Jensen and Priyanka (Pri) Savjani, plus Andrew J Scott of Seven Percent.

Context: the week was a mess. Democrats won both the House and the Senate on 3 November. Markets rallied, then fell hard the next day. Lomax had the Nasdaq Composite down more than 5 percent, with credit default swaps on CoreWeave and SpaceX at record highs. The Bank of England raised rates to 4 percent. Anthropic filed an amended S-1 with the price range still blank. And Sequoia closed its London office.

Why moratoria made Amazon richer and CoreWeave poorer

Lomax set it up. Texas governor-elect Gina Hinojosa ran on a data-centre moratorium. Voters sided against data centres in 40 of 41 local ballot measures. Ohio’s new governor, Amy Acton, backs a conditional moratorium.

So why did the market buy Amazon and Microsoft and sell CoreWeave on the same news? Pri’s answer: if Texas and Georgia pause new builds, existing capacity gets scarcer and more valuable, and Amazon and Microsoft own huge amounts of it, paid for out of their own cash. CoreWeave has to keep building to grow, funds that with expensive debt, and its bonds already trade at distressed levels. Then the chipmakers got sold too, because fewer new data centres means fewer chips bought.

Odd outcome. A backlash against Big Tech’s data centres could end up making Big Tech stronger.

Wall or speed bump?

Mads called it “one hundred percent” a knee-jerk reaction, though possibly one with some real basis. His bigger point was a comparison with dot-com. Back then the worry was overbuild and capacity nobody would use. Now “there are no dark GPUs”. The big labs say they could grow faster with more capacity. The nervousness is about politics, not economics.

Lomax added that the hyperscalers mostly fund this out of free cash flow, with Google Cloud growing at 100 percent a year off a massive base. Mads added the clincher: margins at the big cloud vendors are rising, “the telltale that there is real strong economics here underneath.”

Andrew was blunter. Demand for compute was growing exponentially before AI, and AI just compounds it. If you’re putting billions into a data centre, “you can probably afford to build a few playgrounds”. He thinks Democrats are just playing politics with AI, and expects European politicians to do the same. Lomax wasn’t so sure about the UK.

My view: the anti-AI vote looks less like a verdict on AI than a verdict on electricity bills. If that’s right, the fix is local and solvable.

Sequoia leaving London

Sequoia’s line: “We simply don’t see the need for a local presence.” It follows Coatue and a16z.

Andrew called it déjà vu. US firms either ship in West Coast partners who struggle to build local networks, or hire locally and lose the connection to HQ. “If a European firm left the Bay, nobody would bat an eyelid.” Lomax thinks Sequoia hasn’t had much European success beyond Klarna, and has plenty to do at home.

Mads quoted Synthesia’s Riparbelli: the N in Europe is small enough to cover from SF. But US investors were in more than 22 percent of European rounds in the first half, the highest this decade, and HV and Vsquared have opened in London. Pri’s point: if US funds only fly in at Series B and later, they arrive after the hard early work is done locally. That’s room for us.

EU Inc: pass it or kill it?

The 28th regime was meant to be a European Delaware. Pri summed up the damage: member states cut the single company registry and the common insolvency regime. Lomax blamed the notary lobby, as the pod had predicted.

Andrew’s view: “kill it” and come back when it can be done properly. Mads took the other side: pass it, bypass the notaries, and improve the options programme over time. The options regime covers companies under 750 staff and up to 150 million euros of turnover, with tax due on sale, not grant or vesting. Andrew objected that big payouts come above that size. Mads pointed out the threshold applies at grant. Andrew: “Okay. I take it all back.”

Pri’s wider point was the funding gap: US venture runs at 0.5 to 1% of GDP, the big EU economies a tenth of that, the UK around 0.2%.

UK principles, the covenant and copyright

Andy Burnham’s government approved seven AI principles, including “control and sovereignty matter”, plus a data-centre covenant: priority grid access and fast-track planning in return for environmental and community conditions. Mads read sovereignty as procurement: no foreign kill switch, no dependence on one vendor. Pri liked the covenant and argued for faster adoption with fairer licensing on copyright. Andrew still wants a real statement of strategy.

Revolut and the false economy of self-hosting

Revolut cut AI coding costs 10x by self-hosting GLM, a Chinese open model, on its own Nvidia chips. Pri’s point: the interesting part isn’t the 90 % saving. When tokens became effectively unlimited, engineers started using AI for the boring work (debugging, Jira tickets, documentation), and that’s where the biggest gains came from. Frontier models still get the image and design work.

Mads was sceptical. Across tens of thousands of tests, “GLM is not on the frontier for anything”. Self-hosting only changes that if you run your own data centre, which reminded him of the early cloud arguments. Then there’s what’s in the sausage: some models write less secure code depending on who they think the user is. His verdict: great PR for Revolut, “a bit of a false economy” for most firms.

Hacked freezers

Hackers raised freezer temperatures in supermarkets across Europe, and French blood banks were hit too. The controllers are made by Danfoss, which says it told customers to upgrade. Mads pushed back on blaming the vendor. If you want kit supported year after year, you pay for it. A fridge with a network connection is “a little computer”. Pri sees a real company to be built here, helping supermarkets, factories and hospitals lock down connected kit without replacing it. Lomax also flagged the Binance hack, where 6.3 billion of crypto went after an employee downloaded malware during a fake job interview.

Physical AI: warehouses first

Walmart committed about 6 billion to Symbotic automation across 25 US distribution centres. Meta built a humanoid prototype, Max, with China’s Fourier. Mads’s line: a warehouse “has no attachment to being walkable”, so build the machine for the job. Humanoids are coming, just later than people think. Tesla is running both bets at once with Optimus and Robotaxi.

Andrew calls robotics a software problem in a saturated sector, and backs companies solving an immediate painful need with today’s stack, like robots that replace online shop pickers, or Lumen Labs using biomimicry for GPS-free navigation. His warning: the LLM sledgehammer approach may not work in the physical world. And “distribution beats product.”

Anthropic’s IPO

Anthropic’s amended S-1 says it “may imminently meet” its own threshold for dramatically accelerated progress attributable to AI. Pri listed what an IPO buyer has to believe: revenue keeps compounding (still 47 % quarter on quarter, but slowing) to justify $895 billion of compute commitments. Margins are real, despite a swing from $974 million operating profit to a $567 million loss. And Anthropic stays at the frontier.

Mads thinks it could become a hyperscaler in its own right, and may be overcommitting on purpose, expecting to land 400 billion rather than $900 billion of new capacity, because the contracts let it walk away if partners run late.

Predictions

• Mads Jensen on Anthropic: “there’s a case for for underwriting this at a two trillion dollar valuation now”. He added that people are now talking about a valuation as low as one trillion.
• Lomax Ward: “absent like a c absolute continued catastrophe in the market, I think um they’re gonna go. And even if the price is lower, they’re gonna get it away.”
• Andrew J Scott on European politics: anti-AI campaigning is coming because “they’re the same cloth ultimately.”
• Andrew J Scott: robotaxis are “about to go from innovators and early adopters to early majority”.
• Lomax Ward on cybersecurity M&A after Google bought Wiz: “We’re gonna see more and more of that, I think, over time.”

Deals of the Week

Fractile (Mads): A London AI inference chip company founded by Walter Goodwin, an Oxford robotics PhD. Fractile puts memory right next to processing and is pitching 1200 tokens a second against today’s 40 to 100. It raised a Series C of more than half a billion pounds and has a deal with Anthropic. Lomax thinks it could be one of the biggest companies to come out of Europe.

Lumaril (Pri): A $66 million Series A at a $400 million valuation, led by Elad Gil and also backed by General Catalyst, from the team behind Palantir’s Foundry. It runs AI on a factory’s existing control systems and gets paid on the extra output it creates.

Hypersonica (Andrew): A European hypersonics startup with what Andrew thinks was a 100 million Series B led by Lightspeed and General Catalyst. His frustration: big deep tech and defence rounds keep being led by Americans.

Notable Quotes

Why This Matters

For early-stage European VCs, Sequoia leaving London and US funds arriving only at Series B leaves the seed and Series A ground to us. For founders building AI infrastructure, the midterms show that public consent now carries as much weight as capital, and the UK’s covenant could become a selling point. For anyone pricing AI companies, Anthropic’s IPO is the test case. And for operators tempted by “10x cheaper” models, the question is whether the savings survive the hidden costs.

Personally, I’m a bull. Speed bumps sort the strong from the weak. Just don’t be the one with the loan.