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.

Quick news before the main event

Busy week. Grok 4.5 launched with Cursor, all three flavours of ChatGPT 5.6 are out, and Nvidia shed a trillion dollars, though its stock is still up 1,500% since ChatGPT launched. Prism ML shrunk Alibaba’s Qwen 3.6 model to run on an iPhone 17 Pro, all 27 billion parameters active at once. And Deal Room reckons UK startups raised £17bn in H1 2026, more than double a year ago.

Mads tied off last week: Tesla was tipped to ship 400,000 vehicles and shipped 480,000, its first turnaround print in two years. Samsung posted a 19x jump in quarterly profit and the stock fell 8%. The market has stopped paying for records.

China is distilling, gating, and maybe locking the door

Back in February, Chinese AI models overtook American ones on usage. Not benchmarks, usage. On OpenRouter, the biggest routing platform, Chinese models processed 4.1 trillion tokens in a week against 2.9 trillion for the Americans, a year after the US held 70% of the platform. Today the gap is roughly three to one, with Mads reckoning 61% of tokens now go to Chinese open-weight models. Half of OpenRouter’s users are American, chasing near-frontier performance at a tenth of the cost. Why wouldn’t they?

How did China close the gap so fast? A little naughtiness helped. Anthropic told the US Senate that Alibaba’s Qwen lab ran 29 million exchanges with Claude through 25,000 fake accounts, harvesting outputs to train its own model. That’s distillation. And now Beijing is considering restricting foreign access to its top models. The playbook reads: distil the competition, gate the models, lock the door.

The obvious worry is that this is the rare-earths move again. Mads doesn’t buy it. Open weights are easy to substitute, the models already in the wild stay there, and you can fork something like Nvidia’s Nemotron and build on top. Prism ML is proof. Where Europe is genuinely exposed is the hard stuff, power, data centres and semiconductors, not the models.

Where Mads gets interesting is the game theory. Jensen Huang has always said Nvidia isn’t trying to build the best model, just to sit near the frontier, holding Nemotron back to avoid cannibalising the frontier labs who buy its chips. But those restraints are dissolving, with Anthropic using silicon from Google and Apple and OpenAI taping out its own chip with Broadcom. So if China pulls back and the US labs build their own inference chips anyway, Mads reckons Nvidia floods the market with a commoditised Nemotron and wins. Cheap models sell more GPUs.

Then the tinfoil hat. The industry is waiting on the Anthropic and OpenAI IPOs, hundreds of billions parked in both. If the product is racing toward commoditisation, what does that do to the ability to list? Anthropic is pushing Fable hard because revenue is what gets it to IPO, and the more it climbs into the application layer, the more of its own customers it offends. Ask Dylan Field at Figma. It’s the old IBM and Microsoft problem, the behemoth that upsets the customer base it was built on.

UK pensions finally show up to venture

Dan admits he had happy ears on for this one. The most boring people in Britain did something interesting. Nest, the workplace pension for 14 million Brits with £68bn under management, committed £200m to a VC sleeve with Schroders, growing to £1bn by 2030, starting with Synthesia and Wave. Safe money, but a start.

The scaffolding has been going up for a while. The Mansion House Accord had 17 providers pledge up to 10% of default funds into private markets by 2030, roughly £50bn, half earmarked for the UK. Three years ago DC schemes held about 0.36% in unlisted equities. Last year that doubled, from £800m to £1.6bn. In April the British Business Bank set out to pull £2bn of pension money into venture over five years, and the Pension Schemes Act passed the same month is the gun on the table: play, or we’ll make you. Dan’s take: three years out, pension capital into UK venture will still be single digits, but billions a year, once we get a fund-of-funds structure like France’s Tibi scheme.

Lomax set the backdrop. When the US loosened its rules in the 1970s and 80s to let pension funds into private markets, paired with strong performance, it built US venture into something five to ten times the size of Europe’s. The UK sits on roughly £3 trillion in DC and DB pensions, and only a tiny sliver touches privates. His problem with Nest’s move is where it’s going: Series B, C and D, into companies that could raise anywhere. If you want to move the ecosystem, you want pension money going earlier, into fund-of-funds and managers, not direct deals. But he gets it: Nest needs to show performance, and later-stage markups show up in 24 months rather than the ten years an early-stage fund takes.

Mads was relaxed about the stage. It’s a capital ladder: more money at Series C gives Series B a better market, which gives Series A a better market, which pulls capital down to the seed graduates. Where he drew a hard line is coercion. A forced 10% rule gets satisfied by infrastructure and private credit, and venture might never see a cent. Nest just put £500m into private credit, double its venture cheque. His better model: rather than the government half-playing fund manager through the British Business Bank, build an arm’s-length sovereign wealth fund against the pension liabilities we already carry. Incentives, not the stick.

Is it 1999?

Dan built the segment as a straight fight: ten reasons it is 1999, ten reasons it isn’t, and let the other two pick it apart.

The yeses are real. The Shiller CAPE crossed 40 for only the second time in history, the last being December 1999, right before the Nasdaq lost 78%. A leaked Treasury draft likened AI to the dot-com bubble. Over $800bn of AI deals are circular, Nvidia putting $100bn into OpenAI, OpenAI spending it on Nvidia chips. The AI Big Ten are 40% of the S&P. CapEx-to-sales is projected at 34% this year against 32% at the dot-com peak. OpenAI is reportedly losing $12bn a quarter. Nvidia shed almost a trillion in under two months, Oracle is down 40% in a month, and Blue Origin is raising $10bn at a $130bn valuation.

Mads argued it both ways. His anchor: economists have predicted nine of the last five recessions, and timing the top is a fool’s errand because booms run longer than people expect. He noted the BIS, the central bank of central banks, now openly uses the word mania, comparing this to the canal and railway manias, where funders lose their shirts but the world inherits the infrastructure. His tells: SK Hynix listing in the US with a 2x leveraged ETF, Google of all companies issuing equity, retail piling in while corporates sell. But he wouldn’t call the date.

Lomax took the historical scalpel to it. Central banks are paid to be harbingers of calm, and Greenspan warned about irrational exuberance in 1996, three years early. The Nasdaq is up 115% since ChatGPT, where post-Netscape it ran up 700%. His real worry is the ratio: $70-80bn of AI revenue against trillions of CapEx in the ground, a growth rate that has to keep climbing against a commoditising model layer.

Then the noes. The technology already works, unlike Pets.com and Webvan. The leaders are the most profitable companies in history and paid for most of the build-out with free cash flow. Usage is real tokens, not page views. Valuations are less than half the dot-com peak. Revenue is ramping faster than anything before, OpenAI past $20bn ARR and Anthropic past $9bn in three years, where Amazon took a decade. Demand is physically constrained, with governments treating compute as national infrastructure. And if the CapEx overbuilds, the fibre-glut precedent says it still gets used.

Mads’ closing frame is the one that lands. The difference from 1999 is that the technology is generating real revenue much earlier, not eyeball metrics. It may not be profitable yet, but it’s money. Expect a painful consolidation, especially among the frontier labs racing to build a moat before the foundation models commoditise underneath them.

AI Corner

Fable Five is back after its subscription-plan disappearing act, Sonnet 5 has shipped, and ID verification went live on Wednesday. Fable is formidable and, in Mads’ words, chews through credits faster than anything he’s seen. The strategy is a clean experiment: Fable Five sits at the high end, testing whether you can charge 5x for a token-hungry frontier model, while Sonnet 5 mops up the value-conscious middle at a fifth of the price.

A history lesson from Lisbon

Since Dan and Lomax were on the beach, Mads dug up a Lisbon story. In 1755 an earthquake, tsunami and days of fire flattened the city. The good bit is the engineering. Prime minister Pombal had his chief military engineer Manuel de Maia build the gaiola pombalina, a flexible timber cage hidden inside the masonry, tested by soldiers marching around to fake tremors, the first structural earthquake test in history and one of the first industrial prefab methods. Baixa went back up on a clean grid, unusually American among the higgledy-piggledy hills. Lomax’s footnote: Portugal, at 10 million people, punches near the top of Europe on unicorns per capita behind Sweden, home to the likes of Remote.com and Sword Health.

Predictions

Dan Bowyer: Europe can win as the bubble deflates. The last bubble left the fibre that powered two decades of growth. This time the commoditised superintelligence is a global offering, but Dan sees Europe’s opening in real-world applications, industrial depth, and regulated workflows.

Deals of the Week

Mads: Luffy AI. An £8.1m Series A for self-tuning AI control of industrial electric motors. Unglamorous, which is rather the point.

Lomax: Proxima Fusion. The Munich company raised €411m at a €2.5bn valuation, led by XTX Ventures, the largest fusion round ever in Europe from a company only two or three years old. Proof for the recurring Upside thesis that fusion and quantum are where Europe can play, alongside IQM’s IPO in Finland.

Dan: Quantum Diamonds. Another Munich business, £91m raised to scale semiconductor inspection tech, only £15m of it equity. Following the ASML playbook of creating value now.

Notable Quotes

“So what they’re doing is they’re distilling the competition, gating the models, and now potentially locking the door. Pretty genius, an old Chinese playbook.” – Dan Bowyer, on China’s AI strategy.

“I think as leverage goes, it’s different from rare earths because it’s much easier to substitute.” – Mads Jensen, on why a Chinese model lockout worries him less than the metals.

“This week we’ve seen the most boring people in Britain do something I think is quite interesting. Pension funds have started buying venture.” – Dan Bowyer, on Nest’s £200m venture commitment.

“Economists have predicted nine of the last five recessions.” – Mads Jensen, on the difficulty of calling a market top.