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.

This week it was me, Mads Jensen and Priyanka Savjani. Pri had just closed her first term sheet at SuperSeed, roughly two weeks after joining.

The main question was simple. Who is actually paying for AI now? The week’s news gave us plenty to work with: a pulled IPO in Australia, Samsung’s ninefold profit jump, SpaceX borrowing to buy chips, new models from Anthropic, Mistral and Aleph Alpha, and a fresh round of “come back to Europe” talk aimed at Westminster.

Firmus and the hopes-and-dreams discount

Mads raised this one before we recorded. Firmus is an Nvidia-backed Australian AI data centre builder that started out as a crypto miner. It pulled its Sydney IPO the morning we recorded.

The numbers explain why. Firmus wanted up to $5.5bn at a $30bn valuation, two months after a private round at $10bn. It has Nvidia as a backer and Meta among its customers. But in Mads’s words, “95% of the capacity that you are marketing and selling is still on paper.” Forecast debt is $30bn, about six times forecast earnings once the full build is done. On top of that, 58% of shares would have been tradeable from day one.

Investors read that as insiders looking for a quick exit. I think Firmus was chancing its arm.

The bigger point is what comes next. Nscale, London’s big AI data centre company, is lining up a US listing, and Mads expects “some of the similar rumblings”. He wouldn’t call peak AI. He was more willing to call peak IPO for companies that are mostly a business plan.

Samsung, TSMC and who is footing the bill

Two figures to set the scene. Nearly 50% of UK bond issuance this year went to US hyperscalers, so your pension is in this whether you like it or not. And Samsung’s ~$80bn of operating profit last quarter beat Nvidia’s best-ever quarter.

Mads put it in context. Samsung’s sales were $138bn, and in one quarter it made 20 times the operating profit Tesla makes in a year. Almost all of it came from memory. HBM takes three times the manufacturing capacity of DRAM, so every line switched to HBM squeezes supply of the basic stuff. Apple has said memory costs are now hurting its margins, so prices are rising. That means the consumer is paying.

Pri took the other side of the trade. TSMC’s quarterly revenue jumped 51%, and it runs at about 60% operating margins. One company’s spending is another’s revenue, and the spenders are now burning cash. Alphabet spent nearly $45bn on capex in a single quarter, about double the year before, and its free cash flow went negative $5.9bn. Samsung’s quarterly profit alone is about 1.6 times TSMC’s revenue.

The growth is slowing, though. Samsung’s quarter-on-quarter sales growth has gone from 43% to 28% to 14%. Pri asked the question that matters: “And the question is how long are the hyperscalers going to keep writing those checks before shareholders start asking for returns?”

Debt, SpaceX and what CDS spreads are saying

SpaceX is raising $40bn, about $30bn in bonds and $10bn in bank loans, and that money flows on to Nvidia, TSMC and the memory makers. Mads made the link back to Firmus. It couldn’t go public on hopes and dreams. Musk can, because he has made investors a lot of money before.

I keep an eye on credit default swaps because they price credit risk every day. Mads explained it plainly. Oracle, for decades a cash machine, now costs 244 basis points to insure. That’s $244,000 a year on $10m of debt, which is junk territory. SpaceX’s spreads have climbed a lot too.

Haiku 5.5 puts Anthropic back on the Pareto frontier

Haiku has long been Anthropic’s neglected budget model. In June, OpenAI’s Luna took the Pareto frontier for cheaper workloads and Anthropic was “nowhere to be seen”. Haiku 5.5 is roughly 75% cheaper to run, and it put Anthropic straight back in the game.

We don’t trust lab benchmarks much because of benchmaxing, so SuperSeed runs its own tests: more than 30 workloads and 21,000 test cases. Haiku 5.5 became the best model on eight of them immediately. Against Luna on a given workload, Mads saw 20-25% better performance at lower cost, two and a half months after Luna launched.

Pri asked whether this is odd timing before an IPO, since investors usually want pricing power. Mads thinks Anthropic tried pricing power, found it wasn’t there, and is now betting on volume. Revenue figures are guesswork. OpenAI has just put its own number at closer to $50bn, below what was rumoured.

Mistral Large 4, Kolibri and Europe’s two strategies

Mistral Large 4, or “Le Chonk” (Mads thought of Garfield, and I’m still not convinced), is a trillion-parameter model trained from scratch in Europe. It is the first model from Mistral’s €3bn Series D. It’s a serious model. Mistral was also a little cute with its cyber benchmark, giving US frontier models a zero because they refuse hacking tasks.

On our tests it reached no Pareto frontier anywhere. On sorting engineering tickets, “it performs a little bit better than the model we use, but it costs forty seven times more.” Mads still sees a clear buyer: European governments and businesses that want a European model they can host themselves, with open weights coming.

Pri set out the sovereignty knot. The UK is being urged to use cheaper models, and the cheapest are Chinese open-weight models like DeepSeek, up to 15 times cheaper to run, or US budget models like Haiku. Neither helps national resilience. Cheaper also doesn’t mean a smaller bill: 60% of UK IT leaders worry about rising AI costs even as unit prices fall.

Germany is going small with Aleph Alpha’s Kolibri, a specialist model for public services. The twist is that Cohere, a Canadian company, agreed to buy Aleph Alpha in April. DeepSeek, meanwhile, is raising $12bn in a round backed by Tencent, roughly four times Mistral’s record round. That is the size of the gap Europe has to close.

Isomorphic, Revolut and the London question

Two stories pulled in opposite directions for London. Isomorphic, the Google DeepMind spinout, is in talks to raise about $2bn at $40-50bn. Pri’s read: it’s priced like a frontier AI lab, with no drug in the clinic and roughly 10 times the valuation of Chai Discovery. It is built in London, and the UK’s Sovereign AI Fund is on the cap table. AI for science is where we have an edge.

Then Revolut. Storonsky says a listing would have a US primary. At $115bn privately, Revolut would rank as about the seventh most valuable bank in Europe. Mads called it a loss for a city that has been a financial centre for centuries.

UK and EU: deals behind the scenes

Andy Burnham has floated a second referendum, and France’s finance minister says Britain is welcome back. Mads thinks rejoining is a great idea, but “a really hard sell” while unease about migration persists. He’d push for the customs union and the Made in Europe scheme instead. His view is that Labour has few other growth levers and doesn’t understand tech well. Pri sees sector deals as the short-term reality: agrifood, the electricity market, a youth visa scheme.

The Tories’ “The Right Way” booklet got short shrift. It flags pension funds’ share of British listed equities falling from 22 percent to under 2 percent, rules out mandating them, and doesn’t offer much else.

I’m more bullish on country-to-country deals, like the UK-Germany industrial tech corridor under the Kensington Treaty, than on another round of referendum politics.

Predictions

• Me: “My prediction this week is that we’re going to see more bilateral and maybe even unilateral agreements, but not EU-driven.”
• Mads Jensen on Mistral Large 4: the open-weight version will come out “probably towards the end of the month”.
• Mads Jensen on ASML’s results next week: “I expect it’ll be good numbers coming out.” He said the same about TSMC’s full results.
• Mads Jensen on rejoining the EU: with migration still unresolved, “that’s just gonna be a really hard sell”.

Deals of the Week

• Vocca (Priyanka Savjani): a Paris and New York startup building AI phone assistants for medical practices. It raised a $20m Series A backed by Norrsken VC, Speedinvest, Heal Capital and London’s firstminute capital. It handles over a million patient conversations a month across 1,500 practices, revenue is up 7x in a year, and 70% of calls are resolved with no human involved. It goes deep on each medical specialty instead of building a generic voice bot.
• Universal Quantum (me): a University of Sussex spinout building quantum computers from trapped ions. It raised a $100m Series A, has live contracts, and its HQ happens to be near my house.

Notable Quotes

“But 95% of the capacity that you are marketing and selling is still on paper.” – Mads Jensen, on Firmus

“So investors aren’t buying a sort of pipeline of drug discovery here. They’re buying the solvable disease platform and they’re buying Demis.” – Priyanka Savjani, on Isomorphic

“And the question is how long are the hyperscalers going to keep writing those checks before shareholders start asking for returns?” – Priyanka Savjani, on the AI build-out

“They have to own up to the big mess they made around Brexit because that is squarely on them.” – Mads Jensen, on the Conservatives

Why This Matters

If you’re an LP, your pension is already funding US hyperscalers through bonds. CDS spreads are the earliest warning you’ll get if that goes wrong. If you’re a founder building on AI, prices are falling fast at every tier, and switching models pays off once you have your own test set. European policymakers have to pick between cheap and sovereign, and right now Mistral is the only option that is both, at a price.