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.

Will Trump ban Chinese AI?

Axios ran it on Monday. The White House denied it within hours. But the idea poked Dan in the eye, because a real ban would be an own goal on the US economy and a leg-up for everyone else, Europe included.

Mads doesn’t think a straight ban is coming, because you can’t really ban open source. It’s already out there, and banning it edges into banning free speech. What you can do is go after procurement, and here he reached for the old copyleft playbook: open source ships under a licence, and the strictest (the GPL crowd) says if you use this commercially, we own you.

Washington can run the same move in reverse. Not “you can’t use Kimi K3,” but “if Kimi K3 sits anywhere in your supply chain, you can’t sell to the US government.” No free-speech fight, no model ban, just a quiet amputation of a huge slice of the market. And because plenty of private firms sell onward to government, the reach is enormous. If you’re OpenAI or Anthropic, that’s exactly what you’re praying for.

Here’s the bit that surprised Dan. Kimi K3, Moonshot’s new open-weight model releasing on the 27th, runs at roughly $15 a million tokens against $50 for Fable. He’d expected a 10x gap. It isn’t. And Mads’ point cuts deeper: a token isn’t a token. Fable can chew through a problem in fewer tokens than Kimi K3 needs, so the effective cost gap is narrower than the sticker price. The Chinese will grind their way to efficiency eventually, because that’s what they do. The moat is thinner than it looks. And if China pulls back from open weights entirely, which has been discussed, Mads sees Nvidia cleaning up (more on that below).

The Hugging Face hack, and whether it was just PR

Hugging Face got hacked. An OpenAI model did the hacking and sat inside undetected for hours. Dan flagged that Upside called exactly this in episode 87 back in April: an agentic-AI-attributed breach within twelve months. It took about four.

The detail Mads loved is the one worth keeping. When you get hacked, you turn to your trusted AI to work out what’s happening. Hugging Face’s was the good one from Anthropic. But guardrails. Mid-attack, as they leaned on Fable, the classifier fired, decided they weren’t authorised to do cyber-defence work, and shut them out. So they went to the Chinese models to fend off the American one. As Mads put it, isn’t it glorious.

Was it PR? Nobody on the pod was fully sold either way. What’s not PR is the lesson Mads has been hammering for months: don’t build on a single provider. Every switch-off, classifier reroute and badly-timed guardrail is a reminder that the most important suppliers in your stack are becoming instruments of one government’s mood.

Same theme in the AMD-Anthropic deal, $5bn flowing in a satisfying circle. Mads shrugged. This isn’t Lucent and Nortel shovelling money at kit that ends up in a dark room. The gear is being used. The thing to watch isn’t the vendor financing, it’s the monetisation at the front end. Revenue keeps growing, it’s fine. It stops, that’s the problem, and the financing is just the symptom.

What would you do in Number 10?

Wake up tomorrow as the UK’s eighth prime minister in ten years, inside the world we actually work in. What do you do?

First, the scorecard. Burnham has come out swinging with crowd-pleasers. Scrapped digital ID. VAT off domestic electricity. The £2 bus fare cap back. Business rates cuts for pubs and clubs. £340m to end rough sleeping. North Sea oil and gas licensing reopened. And he killed the Department for Science, Innovation and Technology, folding it into Business and the Cabinet Office, with AI minister Kanishka Narayan now allowed to sit in the corner of Cabinet without being in it. Mads reads that as a step backwards, Lomax thinks it’s fine now technology sits inside a proper Whitehall department. At Number 11 is John Healy, the ex-defence minister who once resigned over not getting enough capital to do the job. Easy to file as “unknown,” except he was Gordon Brown’s PPS and spent years in the Treasury under Blair and Brown. A defence hawk. Whether he’s a fiscal one is the open question.

Then Mads set the backdrop, and it’s grim. Had the UK kept growing at its late-70s-to-2008 trend, GDP per head would be about £10,000 higher. A real family, ten grand poorer, every year. Output per head now sits just above Mississippi, and below it once you take London out. Poland is on track to match us within a decade. Debt interest is £110bn, more than the entire education budget. And welfare, at £330bn last year and heading past £400bn by 2031, already exceeds the whole income tax take.

His diagnosis is the risk-free society, built brick by careful brick over thirty years, every brick with a good reason. Maxwell robbed the Mirror pension scheme in 1991, so we got the Pensions Act, which pushed funds to value everything as if held in gilts, and UK pension holdings of domestic equity fell from over 50% to about 4%. Grenfell gave us the Building Safety Act and second staircases, and London homebuilding collapsed by around 84%. The 2008 crisis pushed bank core equity from 2% to 7%, and SME lending fell from around 10% of GDP to 6-7%. Each fix was sensible. Stacked together they strangle growth. HS2 is the monument, heading past £600m a mile when Spain builds for £25m and France for £46m. We don’t even have mountains.

The answers split usefully. Lomax, blank slate: shrink the state, keep fiscal credibility (no Liz Truss lettuce moment), attack the welfare bill, get pension capital into growth, court foreign investors with a hard look at corporation tax, and deregulate. Energy first. Nuclear, cheaper industrial electricity, buildable data centres. Because Stargate UK got cancelled and most of last year’s American announcements quietly unwound, for the same reason: it’s too hard to do here.

Dan’s five: rule out a wealth tax on day one, declare war on industrial energy costs, fix planning, give Reform teeth, restructure taxes rather than raise them, and force pension capital into UK assets. Then he went somewhere the other two didn’t, which is pride. He thinks the country lacks a shared sense of getting up and building, and that outside the southeast bubble it’s a harder, different country. Fix the feeling first.

Mads pushed back, gently. Cart before horse. Find the roughly £30bn to balance the current account, rates fall, lending and housebuilding pick up, regulation unwinds, and pride follows the growth. Lomax split the difference with the cleanest frame of the episode: it’s a startup. You need the fundamentals (mostly the economy) and you need the narrative. Blair was great at both. Starmer inherited a 320-plus majority, laid some fundamentals, and whiffed the narrative. Seven prime ministers in ten years is what that looks like.

On pensions the disagreement is real. Dan would force capital home. Mads won’t confiscate savings, and doesn’t need to. We already spend more than £50bn a year on pension tax incentives, so re-point them: less relief for piling into the Magnificent Seven, full tilt for backing UK startups and assets. And judge a good pension manager by long-term returns, not by minimising fees and volatility.

Then the thing this show cares about, AI. Lomax isn’t optimistic. He thinks AI is a vote loser for Burnham, so it won’t be a priority. The specific worry: a framework came in early this year letting central government override local councils on planning for nationally strategic infrastructure, exactly what you need to build data centres over NIMBY objections. Burnham is a devolution man. Hand that power back to the regions and the override dies, the build gets harder, and Matt Clifford’s AI growth zones could quietly unwind. Dan’s more hopeful. We’ll find out.

Predictions

Mads Jensen: if China stops releasing open weights, Nvidia cleans up. Nemotron comes off the leash, Nvidia ships a strong open model, and demand for the one thing you need to run it goes up. Commoditise your complements.

Mads Jensen: humanoids might be a Newton moment, not an iPhone moment. He’s not convinced this wave of humanoid robots sticks. Purpose-built robots in manufacturing, construction and agriculture are where he’s bullish, and robotaxis he thinks will be huge.

Deals of the Week

Mads: Cusp AI. Using AI to discover new materials, claiming to evaluate 300 trillion candidate structures. A $450m Series B led by Kleiner Perkins and NEA, co-founded by Chad Edwards (ex-Quantinuum) and Max Welling, co-inventor of the variational autoencoder that sits under generative AI. The UK government is in the round.

Mads: Arrakis. A Swiss startup, and yes, the Dune reference is the pitch. If data is the spice, Arrakis wants to trade the industrial kind: procurement data, supply chains, project delivery, the stuff locked inside large enterprises. A €30m Series A led by Blossom Capital.

Lomax: Delivery Hero. Not a raise but an ending. Uber is acquiring the listed delivery giant, roughly €14bn, after it scaled to 50 markets across mostly Europe and, in Lomax’ read, hit its ceiling.

Notable Quotes

“That place is called the United Kingdom, my friend.” – Mads Jensen, on where Europe’s builders should be able to build.

“As they turned to Fable to defend themselves, the classifier came out and said you’re not authorised to do that. So they had to go to the Chinese models to defend themselves from the OpenAI attack. Isn’t it glorious?” – Mads Jensen, on the Hugging Face hack.

“Brick by brick, we have been building this risk-free society. And this is what’s killing growth.” – Mads Jensen, on thirty years of well-meaning regulation.

“You need the fundamentals and then you need the narrative.” – Lomax Ward, on governing a country like running a startup.