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.

What secondaries actually are, and what everyone gets wrong

Matt Russell from Vencap started in what used to be the quiet corner of private markets. Not anymore. Companies are staying private far longer, and the secondary market has become where people go for cash. The whole secondary market across private equity and venture is about $200bn at the fund level, with around $3 trillion of unrealised value sitting in venture funds globally. Venture is far younger than PE, so the runway is long: no reason the venture slice isn’t a $100bn market in short order.

Here’s the bit most people get wrong. Say “secondaries” and people picture distressed sellers and big discounts. VenCap’s experience is the opposite. Some of their best-performing deals were bought at a premium, because the prize was exposure to a great company with a lot of growth ahead, not a cheap stake in something mediocre. Secondaries follow the same power law as everything else in venture: quality and access beat a bargain. With 500 funds and 17,000 portfolio companies tracked over four decades, the book lands roughly two-thirds US, then India and China. No geographic targets, because the moment you force an allocation you invite adverse selection.

Europe writes a rulebook for the kill-switch

Brussels is rolling out a sweeping tech sovereignty package, and we’ve heard the drumbeat before. This time the interesting piece is CATA, the Cloud and AI Development Act. As Mads put it, this isn’t about building our own Google. It’s about making sure critical workloads can’t be switched off remotely or subpoenaed under the US CLOUD Act.

The mechanism is a tiering system. Sort the workloads governments buy by criticality, levels one to four. Level one means data on EU soil. Level four means a full EU supply chain with no foreign interference and no remote kill-switch. A proposal, but the direction is clear.

The backdrop is uncomfortable. Europe leans on non-European tech for 80%-plus of its services, and households park roughly €300bn a year outside the EU, mostly in the US. And while Macron does his startup-nation last hurrah at Versailles, with €110bn of AI pledges and €75bn from SoftBank, French debt is climbing toward €3.5 trillion, about 115% of GDP. The pitch is sovereignty. The balance sheet says something else.

The cloud majors are already improvising. Microsoft has licensed the whole Azure stack to European operators, Bleu in France and Delos in Germany, and Google has a Thales arrangement, 3NS. AWS hasn’t. So the open question is which of these clears level three, “owned and controlled by,” and whether a European provider could even run a US frontier model locally while licensing the IP. Nobody knows yet, and Mistral is the obvious beneficiary. Mads, predictably, dragged it back to the boring fundamentals: complete the single market, free up pension capital. That’s what moves the needle, not tinkering with legislation.

The Palantir problem is a lock-in problem

This runs straight on from sovereignty. MPs want the government to tear up Palantir’s £330m NHS data contract and hand it to a British supplier. There is no British supplier. There’s also a £240m MOD deal that went through with no competitive tender, and Sadiq Khan blocked a £50m Met deal on the side. The Dutch, Swiss and Germans are doing the same maths on their own sensitive systems.

Chi Onwurah is leading the charge in the Commons, and she’s no lightweight, possibly the only chartered engineer in Parliament, 20 years in telecoms. The lock-in worry is legitimate, because once an enterprise platform is embedded you can’t easily get it out. And the federated data platform she’s worried about delivers: it sits on top of every trust’s data, builds digital twins of how hospitals run, and by every account cuts stays and improves outcomes.

The cautionary tale is the National Programme for IT from the Blair years, scrapped in 2011 after £10bn spent and nothing to show. So the fear is real: scrap Palantir for a British thing that doesn’t exist, and you’ve commissioned another £10bn hole. Matt’s read, when I asked whether he’d rip it out as an MP, was that he wouldn’t, but he’d panic about how to build the ability to swap providers later. Mads landed it: two wrongs don’t make a right. We regulated ourselves into a thin tech sector, and now want to legislate our way out of a hole we dug fifteen years ago. Better to ask what conditions let us win the next one. It might just be Palantir for now.

Enterprise AI spend, and the ROI nobody can forecast

Companies are starting to complain about the bill. Uber capped staff at £1,500 a month each after torching its annual AI budget in four months. Pizza Hut is being sued over a system that served cold pizzas, and Starbucks built a barista vision tool that couldn’t count milk. Consumption-based token pricing means nobody can forecast spend, which is the real problem under the PR.

Matt’s framing on ROI is the right one. Long term it’s a margin thing, more productive people on higher-value work, but you can’t run that analysis cleanly while you’re still iterating, so for now it stays anecdotal. VenCap, like most firms, is in the experimentation phase. I keep coming back to a BVCA panel I did earlier this year, maybe 250 VCs in the room, where I asked who’s using AI where it’s genuinely moving the needle, embedded, not experiments. A small smattering of hands, which is a fair proxy for most businesses right now. The ROI shows up in cash flows eventually. Business just doesn’t move as fast as the people expecting a profitability bump next quarter would like.

The $80bn club

Everyone seems to be raising $80bn. Anthropic is going public and jumping OpenAI’s queue. SpaceX goes live on 12 June at near $1.85 trillion. And Alphabet is raising $85bn on top of $100bn of debt, with Berkshire in for $10bn.

Can the market absorb three of the biggest IPOs in history in one year? Goldman’s study says yes, and the reasoning is counterintuitive. IPO activity is historically low: about 100 IPOs this year against 250 in 2021 and 400 in 1999, and issuance is near 1% of market cap against a 1.5% norm. The floats are tiny, around 10% to start, and even as lockups roll off the demand holds, because corporates keep buying back their own stock. Worth noting Goldman is also running the SpaceX IPO. On their own numbers, the path to $1.78 trillion needs SpaceX’s AI revenue to grow 100x by 2030, from $3bn to $300bn. It comes down to space AI and whether people believe it lands in that window.

Matt won’t stock-pick, VenCap never has, but pushed for his last $100 he leaned where the data leans: the top 0.1% keep surprising to the upside. On Alphabet, Mads waved off the Buffett-signal reading. Buffett isn’t making this call, Greg Abel is, and the raise is capex anyway, demand needing more data centres and more hardware.

America sues its own

The usual choreography is Europe legislates, America sues. This week the table turned and America sued its own. Florida became the first state to take AI to court, suing OpenAI and Sam Altman personally for an unsafe product, off the back of a criminal probe tied to a campus shooting. Same week, Trump watered down his AI order: a 90-day mandatory review became a voluntary 30-day one.

Mads drew the useful line. Suing works when the harm is bounded and local, the kind a trial and a remedy can handle after the fact. It breaks when the harm is unbounded and has already diffused by the time you reach court, the nuclear-meltdown case. The personal naming of Altman bothered him more, and me too. Limitation of liability is load-bearing for entrepreneurship, and the bar for stripping it should stay high.

Deal of the week

Mads picked Stark, the Berlin defence-tech company raising €300m at €2.5bn. They make exactly what the Pope told everyone not to make last week: autonomous loitering munitions, plus Minerva, an AI command layer. Autonomous weapons are coming whether we like it or not, because the other side will build them. The only question is how far you lean in.

Mine was Gigaton, formerly Carbonry, a lovely boring business that helps cement, steel, glass and chemicals firms run cleaner, more efficient systems. They’ve raised $26m to build out their control systems. I love an industrial efficiency play.

Predictions

I’ll go first, since Mads doesn’t love them. Europe can’t achieve total AI self-sufficiency, and few would argue otherwise. But it will achieve strict sovereign control, and it’ll do it by splitting in two. A state tier, where government and critical infrastructure run solely in the EU on the likes of Mistral and Aleph Alpha, hosted on native cloud. And a commercial tier running tamed, legally compliant US frontier models from Anthropic and OpenAI on hived-off AWS or Azure. I think we see that two-tiering through 2027.

The follow-on is a change of mind for me. I no longer think Chinese open-source models make real inroads into regulated, data-touching workloads, not where patient records are involved. Mads pushed back, fairly: coding is the most important workload right now, and plenty of Chinese open-weight models sit in coding stacks already. So maybe in the unregulated MVP layer, just not near the patient records. And the way Matt framed the model market stuck with me: not commoditised like electricity, but like airlines, low-cost carriers at one end and premium charters at the other.

Notable Quotes

“Some of our very best performing deals we’ve paid a premium for, because it’s been exposure to a great company with a lot of growth ahead of it.” – Matt Russell, VenCap

“We’ve always been too conservative on the winners. Time and time again, we’ve undermodeled it.” – Matt Russell, on venture’s biggest outcomes