TL;DR
Dario Amodei's AI safety essay set off a 48-hour chain reaction across every major AI lab, and the real fight isn't whether AI needs checking, it's who pays for it and what happens the first time an evaluator says no. Layer on the Fed's first rate rise in three years landing straight on debt-funded AI capex, plus one of the biggest weeks yet for humanoid robots, and you've got a snapshot of where this industry actually is right now.
Key Takeaways
- Jensen Huang and Dario Amodei look like opposites on AI safety, but both endorsed the same mechanism at All-In: outside evaluators with real access. The dispute is who pays, and what happens the first time one says no.
- AI capex has shifted from equity bets to debt-funded infrastructure, so a Fed rate rise now hits spending directly, not just the valuation multiple.
- Nvidia's growth math only works if AI capex holds near $5.5 trillion through 2030. Shrink that to $3 trillion and, as Mads put it, "we're all in a world of pain."
- Isembard doesn't own a single machine. It franchises Britain's ageing CNC shops with an order book and routing software, four factories in January, sixteen now.
- Langdock spent millions of euros undoing the Delaware flip it once did for free, because a staffless US holding company was triggering a Cloud Act review on every deal.
- No on-air predictions worth a section this week, but three deals stood out: a Rome cybersecurity startup, a DeepMind world-model spinout, and Snap's AR glasses.
Listen to this episode
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.
Los Angeles, and the argument everyone’s actually having
I’ll be honest about my priors: I’ve never been the biggest fan of the All-In podcast, a bit red-pilled for my taste. But you can’t argue with the lineup, Elon, Jensen Huang, Trump dialing in live to talk to Huang directly, none of the rage-clip editing you get on YouTube. That changed how I heard the whole thing.
Mads’ read on the event is less right-wing think tank, more a celebration of capitalism: “If we don’t have entrepreneurs creating businesses that create wealth, there’s nothing to tax. Nothing’s going to happen.” You can argue how much free stuff a society should fund. You can’t fund any of it without someone building the wealth first.
But the moment everyone’s talking about is the essay. Dario published roughly 3,000 words called “We Must Pace the Frontier,” and within 48 hours every major player in AI had responded. His ask was never a pause, it was three things: independent evaluators inside frontier labs with employee-level access and the right to publish findings a company can’t redact, democratic-country labs agreeing shared safety standards, and eventually extending that to China.
Sam Altman agreed within hours and pushed OpenAI’s IPO out of 2026 in the same breath. Elon Musk said Dario is right, but his fix goes further, rival labs testing each other’s models before release, logged, so anyone caught distilling a competitor gets cut off. Demis Hassabis called the direction correct without committing Google DeepMind. Nadella didn’t commit either, using it to argue for open-weight models instead. Zuckerberg said labs already feel a responsibility to be safe because of liability, a polite way of saying no.
Jensen Huang was the most opposed. He called the doom predictions “made up and irresponsible” and told the room to run as fast as it can, safety being an engineering problem, not a speed limit. Yet in the same session he endorsed the exact mechanism Dario is asking for. Pri caught it live: “Jensen endorsed third-party evaluators in the same session.” Huang’s own words: “That’s no different to financial control. We have auditors for the banks.”
That’s the bit that gets lost in the noise. Huang and Dario are reported as the two poles of this argument, and they agree on the mechanism. Both want auditors. What’s actually in dispute is who pays for the access, how long they get it, and what happens the first time an evaluator says no. Trump called the whole essay a hoax on stage to Huang’s face, and China called it fear-mongering dressed as safety. Everyone is arguing their own commercial position in the language of caution or acceleration, whichever suits them, which is why one essay moved the whole industry inside a day when two years of white papers hadn’t.
What a Fed rate rise does to a $5.5 trillion capex bet
The Fed raised rates 25 basis points to 3.75-4.00%, the first hike in three years. The Bank of England held at 3.75% and paused gilt sales. The US 10-year yield broke above 5% for the first time in three years too. On a normal tech podcast that’s a rounding error. On this one it isn’t, because of how AI capex actually gets financed now.
The old model was simple: tech companies lose money for years, and you discount far-off future profits back to today, so a higher rate just shrinks the multiple. That’s not what’s happening here. Meta’s Louisiana data centre deal, Oracle’s bonds, Nvidia guaranteeing debt on leases, this build-out runs on debt, so a rate rise now hits the cost of the spending decision itself, today.
Amazon says its AI capex has a three-year payback right now, and even at four years that’s still worth doing, so the first 100 basis points probably just eat into slack that already existed. The real question is elasticity. JP Morgan puts total AI capex at roughly $5.5 trillion through 2030. Does that hold at these rates, or does it become $4 trillion, or $3 trillion?
Nvidia is where that bites hardest. It takes roughly 53 cents of every dollar spent on AI capex, and next year’s guide of $690 to $700 billion in revenue needs about $1.3 trillion of capex at that rate, which the $5.5 trillion envelope covers, but not much past it. Only around $600 billion a year is left for Nvidia between 2028 and 2030 once that pool is spent. Shrink the envelope to $3 trillion and, as Mads put it, “we’re all in a world of pain.” Its margins have already slipped from roughly 79% to 73%, still fat, just less fat.
Nvidia has also been the backstop funnelling capital into the wider venture ecosystem, part of why early-stage companies can raise $20 to $100 million rounds right now on visible exits like SpaceX.
Physical AI’s biggest week, and the argument over whether humanoids are even the right shape
Agility Robotics said its Digit 5 is ready to work next to humans with no cages separating them, already logging $300 million in orders. UBTECH is mass-producing humanoids in China at roughly one robot every ten minutes. D-Robotics raised $400 million to build what it calls the brain for every robot, and Bain Capital Ventures closed a $1.6 billion fund for physical AI. A big week, by any measure.
Mads isn’t sold on the form factor, and neither am I. His question is the right one: why buy a humanoid robot to drive your forklift, when you can just put the intelligence into the forklift? Plenty of deployments will go to other shapes, which is what London’s Humanoid, with its half-height, pallet-based design, is betting on.
The bit of physical AI that actually impressed me wasn’t humanoid at all. Isembard opened its global HQ and factory in Southwark, London, on Monday, making machined metal parts for customers including Anduril, Babcock and Tekever. Founder Alex Fitzgerald started the company under two years ago, raised a $59 million seed and a $50 million Series A led by Union Square Ventures, and started this year with four factories. It now has sixteen, across the UK, the US, Germany and France, revenue up roughly tenfold. You can’t build twelve factories in nine months by pouring concrete, so what’s actually going on is a franchise model. Britain is full of small machine shops with two or three CNC machines, often run by someone in their sixties with no succession plan, who can’t win aerospace or defence work because certification is expensive and the sales cycle is brutal. Isembard turns up with the order book, the paperwork and the routing software. The product isn’t the machining, it’s the demand. Its first franchisee is a 22-year-old machine operator who now runs a factory.
UBTECH’s factory is worth a second look too: the software running it is Siemens, German digital-twin tech, not homegrown. China makes about 6% of the world’s engineering software against 28% of its manufacturing, and the design office in its car plants is still largely European even as the shop floor turns Chinese fast. Not all the hardware here is Chinese, and not all the capital is American either.
Deals of the Week
Exein (Rome), Mads’ pick. $270 million raised at a $1.7 billion valuation, led by Headline, Sofina and Goldman Sachs, up thirtyfold in two years from a EUR 15 million Series B in 2024. Founder Gianni Cuozzo previously ran a firmware penetration-testing firm, and built Exein to make every chip defend itself from malware, selling to chipmakers rather than IT departments. As physical AI puts an attackable computer into every robot and drone, that bet gets more relevant, not less.
Emulate (London), my pick. A DeepMind spinout founded by Jack Parker-Holder, Matthew McGill and Philip Ball, looking to raise $700 million at a $4 billion pre-seed valuation to build world models, a natural extension of the trio’s work inside DeepMind on games and simulated environments.
Snap Spectacles, Priyanka’s pick. Snap launched its AI-powered smart glasses in LA this week, letting you overlay a map or watch a show in your field of view. Meta’s the obvious comparison, and how Snap competes with Meta’s distribution on hardware like this is the open question nobody answered on stage.
Notable Quotes
“If we don’t have entrepreneurs creating businesses that create wealth, there’s nothing to tax. Nothing’s going to happen.”
Mads Jensen
“Sovereignty, where your company is set up, has stopped being a political preference.”
Priyanka Savjani
“If we would like to have third-party evaluators, that’s no different to financial control. We have auditors for the banks.”
Jensen Huang, Nvidia, relayed on the show by Priyanka Savjani
“I’m not massively sold on the Optimus next to granny.”
Dan
Frequently Asked Questions
Published 12 September, it asked for three things: independent evaluators inside frontier labs with employee-level access and the right to publish findings a company can't redact, democratic-country labs coordinating on safety standards, and eventually extending that to China. It wasn't a request to pause AI, it was a request to close the gap between what models can do and what anyone can verify.
Largely, yes, on the mechanism. Huang rejected slowing down, but in the same All-In session he endorsed giving third-party evaluators access, comparing it to financial auditing. The real disagreement is who funds and controls that access, not whether outside checking should exist.
Because much of current AI infrastructure is financed with debt rather than equity, a rate rise raises the direct cost of that spending today rather than just discounting some future profit. Analysts on the show put total AI capex at roughly $5.5 trillion through 2030, and the open question is how much that shrinks as borrowing gets pricier.
Nvidia currently captures roughly 53% of every dollar spent on AI capex. Next year's guidance of $690 to $700 billion in revenue needs about $1.3 trillion of capex at that rate, which the current $5.5 trillion envelope supports. Past that, only around $600 billion a year is left for Nvidia between 2028 and 2030, so further growth needs total capex well past $5.5 trillion.
Isembard doesn't own the machines in its network. Local operators keep their own CNC equipment, and Isembard supplies the order book, certification paperwork and the routing software. That's how it went from four factories to sixteen, across four countries, in under a year, without the capital spend a fully owned model would need.