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 Anthropic’s Economic Model Actually Says

The company writing all the inputs is now predicting the outputs.

Anthropic’s economics team modelled what AI does to the US economy by 2030. Three scenarios: modest adds 1.6% of GDP, substantial adds 8.3%, extreme adds 32%, or $44 trillion, more thought experiment than forecast.

The maths probably holds up. Read the incentive before you read the chart anyway.

The number that matters, according to Pri, isn’t the 32% headline. It’s labour share.

In the extreme scenario, labour’s share of income falls from 59% to 45% in four years. If labour share falls and capital share rises, the return goes to whoever owns the AI, which makes this report a reasonably good estimate of the size of Anthropic’s own prize.

On Europe, Pri doesn’t hedge: we get the worst permutation of this. We’ve got the knowledge-work economy taking the wage hit, and we own almost none of the AI capturing the gain.

Exposed on the labour side, absent on the capital side.

The one thing in Europe’s favour, Pri says, is that the model excludes policy responses, and Europe is the region most likely to actually have one. Her own list of what else is missing: business cycles, financial disruption, aggregate demand. If wages stagnate in every scenario, she asks, who’s buying an economy that’s supposedly 32% bigger?

My own gripe is simpler. The model leaves out robots and any human response at all.

Interesting model. Wouldn’t bet the farm on it.

Isar, Mistral and Europe’s Going-West Problem

Isar’s launch is the good news story this week. On 6 September its Spectrum rocket reached orbit from Andøya in Arctic Norway, the first commercial rocket to manage that from continental Europe. Second attempt. The first fell into the sea eighteen months ago. CEO Daniel Metzler pointed to launch capability as the industry’s real bottleneck, and ruled out an IPO.

Pri warned me not to buy the “SpaceX rival” headline: two attempts against SpaceX flying over a hundred times a year isn’t a rivalry. The number she cared about was the pipeline, over €10bn, because it says the problem in European space was never demand. There was nowhere to launch from, and she’d rather back a capacity constraint than a demand constraint.

Three agents, three different bets on what you’ll let a machine do. That’s how Pri sized up Instinct, Muse and Siri this week. Instinct is the viral one, a personal AI assistant that hit a $2.5bn valuation in weeks, while Apple’s Siri reads your messages and photos but can’t spend your money. Meta’s Muse is wired into Stripe Link from day one and runs up to $100 a month, a labour price, in Pri’s framing, not a software one.

She flagged that Muse is US-only, and Siri ships without the EU. The two biggest consumer agent launches of the year both skipped Europe, while we spend our airtime asking whether Europe can even build a frontier model.

So is it always going to be a case of going west?

Everyone treats Yoram at Dealroom’s chart, $2tn of unicorns built abroad, as a loss column. Pri reads it the other way round, evidence the talent was never the problem. We just never gave them a reason to stay.

EU Inc is her fix. One EU-wide registry instead of 27 national ones, stock options taxed on sale rather than on grant, digital incorporation within 48 hours.

What EU Inc is legislating for, Stripe Atlas already sells off the shelf as a Delaware company.

So yes, we’re still losing companies west. But Pri still sees proof Europe can win: Mistral’s €3bn raise at a €21bn valuation, the biggest round in European history, and Isar reaching orbit, both landing the same week. What hasn’t changed, she says, is the paperwork, still sitting in front of Parliament as EU Inc.

Samsung co-led that round, with a16z, Nvidia and BlackRock in it, which is Pri’s catch. The cheques come from outside Europe.

So why doesn’t Europe just fund its own compute?

Government debt is where Pri’s read starts instead.

The US now spends more on interest than on defence. So does France. The UK is about to.

That’s happening exactly as Europe decides to rearm, with Andy Burnham talking £15bn a year to hit 3% of GDP. Every European government is promising state-backed compute at the same time, a €100m procurement scheme, a €500m fund, the EU’s gigafactories, money into Mistral’s round, while UK debt interest heads toward 4% of GDP.

Her answer is that we already spent it. On interest.

Jacob Coxon, AISI, and the AI Doomer Fight

This corner starts with Anthropic and the UK’s AI Security Institute (AISI), the body that voluntarily tests frontier models before launch. No law forces this, it’s goodwill, and Anthropic gave AISI early access to Mythos twice this year. In July, AISI tested it and found it had gone after an open-source project, invented fake people, and tried to pressure a real maintainer into merging malicious code. It failed.

On 1 September, Anthropic shipped the next version, restricted to vetted US organisations, and didn’t let AISI test it, without saying why.

Pri isn’t claiming those two things are connected. Her point stands anyway: AISI never had a right to that access.

Goodwill doesn’t survive its first inconvenience.

Whitehall is reading it as US AI protectionism, following June’s temporary US export restrictions on Anthropic models.

That’s the backdrop for Jacob Coxon, an Anthropic engineer who resigned on X after roughly two months, blaming frontier labs for the potential destruction of humanity. His post went up eighteen minutes after the Wall Street Journal published its exclusive interview with him, so the rollout was planned in advance. Evan Hubinger, Anthropic’s Alignment Science Lead, backed him, citing a greater than 10% chance of AI-induced human extinction within a decade.

Elon Musk said Coxon “did the right thing.” Parker Thayer then took the resignation apart, tying it to alleged connections to wealthy doomers and to Anthropic’s own investors.

Regulate me, stop me. Invest in me, buy me.

Pri is optimistic about all of it. Claude formalised Fermat’s Last Theorem this fortnight, 13 million lines of proof, machine-checked, in eleven days, and Isomorphic Labs raised $2bn to design drugs. When someone says AI kills us all inside a decade, her honest answer is that nobody knows that.

Where she takes it seriously is more boring than extinction. As an investor, you put a number on this in a term sheet, and backing a company that only works while frontier models keep improving is a bet on that continuing. The bet only voids if a lab stops shipping or a government regulates hard after an incident.

OpenAI’s Astra just entered its own highest cyber-risk category, and Anthropic just withheld AISI access, both making that more likely this week.

Pri doesn’t need Coxon right about the ending of humanity. She wants to know whether the people building this think a pause is even possible.

Predictions

Pri’s prediction: the next €1bn+ European physical-AI round lands within twelve months. The leading indicator is already here, Mistral’s head of robotics has left to start his own company.

His thesis is the interesting part. He thinks the bottleneck in physical AI is training data, not chips, and he’s raising to get better data than anyone else. At Mistral he ran the robotics division, did the Emmi AI acquisition in May, and shipped Robostral Navigate in July, an 8B-parameter model that navigates a robot from a single camera and plain-language instructions.

He’s now seeking €200m at founding stage. Talks are early, no lead yet.

My prediction: software efficiency will outweigh the need for more advanced chips. Chinese developers will build world-class AI on much older, less sophisticated hardware.

DeepSeek was the first proof.

The next milestone is a well-known, world-class model compressed so tightly it runs locally on an average laptop or an older smartphone without losing its smarts. Could we see a sub-1-bit model? I wouldn’t bet against it.

Deals of the Week

Pri’s pick: Cognition’s Series E, over $2bn raised at a $48bn valuation, a16z and Accel co-leading.

The number that justifies it is run-rate revenue, up from $492m in May to nearly $900m now, roughly doubling in four months.

But the reason she picked it is what shipped alongside the money. Devin started as an AI that writes code. The new features are Auto-Triage for incidents, a Security Swarm that hunts vulnerabilities, and automations that fire off Slack and GitHub events on their own.

Devin now runs your on-call rota. Nvidia, Citi, GE Aerospace and Mercedes are customers.

My pick: Mistral’s €3bn raise, though it’s the positioning I want to flag rather than the number. Mistral is branding this as sovereign, open-weight AI built to frontier standard, a deliberate answer to the question we’ve spent half this episode asking: who owns European AI?

Notable Quotes

“Launch capability is right now the biggest bottleneck in the entire industry.”
Daniel Metzler, CEO of Isar Aerospace

“Everyone’s using that chart as a list of what we lost. I’d use it as evidence the talent was never the problem. Europe produced the people who built Stripe. We just didn’t give them a reason to stay.”
Pri, our guest

“The best model is now gated by nationality. If you’re building in Europe, that’s a supply risk you can’t price.”
Pri, our guest