TL;DR
The Trump-Xi summit bought only a short extension of the US-China trade truce, while Europe's AI Act and new product liability rules quietly set the template for AI accountability. Anthropic and OpenAI both cut frontier model prices, a sign that pricing power at the top of the stack is weaker than the labs hoped. And the AI money window is narrowing: data centre IPOs are stalling and investors want revenue before they underwrite.
Key Takeaways
- The US-China rare earths truce runs only to 10 January. China still controls more than 90% of rare earths refining and 94% of magnets, and Estonia's new plant will produce around 2,000 tonnes a year against 30,000 tonnes of Chinese imports.
- Trump's ban on autonomous robots over 2kg made outside the US leaves a door open for European makers (Husqvarna already has an exemption) while shutting out Chinese firms entirely.
- Anthropic's Opus 5.5 launched about 40% cheaper than Opus 5, and OpenAI's Sol and Luna came in at roughly half their predecessors' prices. Mads argues OpenAI now holds most of the price-performance frontier below Anthropic's top models.
- From 9 December, software shipped in the EU with an AI model counts as a product under strict liability, covering injury, property and lost personal data, but not financial loss.
- Dutch pension fund ABP has committed €1bn to venture and tech, including €250m into the European Scale Up Fund, which has just underwritten Euclid's €200m Series A.
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.
Trump, Xi and why Europe already wrote the incident rulebook
Xi flew in for his first state visit in 11 years. Days earlier Trump announced an “AI force” and called AI doomerism a hoax. Same old noises.
The concrete output? The trade truce, tariffs held back in exchange for rare earths flowing, was extended to 10 January. Hopes had been for six months. As Mads put it, if you wanted to buy time “you’d try to buy a few years, not a few months.”
The more interesting idea was an AI hotline between the White House and Xi’s office, a revival of the Cold War red phone. (I’m assuming a red iPhone. Foldable, obviously.) Andrew reached for the SALT arms treaties, where each side worried limits would lock in the other’s lead. Negotiators could count launchers. You can’t count chips the same way.
Mads tied it to this week’s real incident. An OpenAI agent breached Australian government servers, and OpenAI took around a month to tell the Australians. Autonomy is what makes AI powerful, but when it does something it wasn’t meant to on privileged government systems, the obligation should be to report “more or less real time and not after the fact.”
This is where Europe looks smart. The much-derided AI Act already requires a model lab to report an incident in Europe “without undue delay”, and has done for over a year. I’m with David Sacks on this one. It’s your model, dudes. You manage it.
Rare earths aren’t rare
They’re all over the map, including in Europe. What we did was push the dirty, toxic mining and refining to China so we could feel clean and green at home. Magnets are in bloody everything, from AirPods to (more importantly, as Mads reminded me) electric motors, EVs and drones.
Mads was scathing on the UK reaffirming net zero by 2050 while the rare earths question, which goes to our ability to defend ourselves, sits unaddressed. Estonia’s plant is a start. 2,000 tonnes against 30,000 isn’t ambition.
The robot vacuum ban
Andrew flagged the one I’d missed. Trump has banned advanced robotics over 2kg made outside the US, which sweeps in autonomous vacuums and mowers. Mads explained the exemption route: show the Pentagon a path to 75% US-made components, not just assembly. Anything owned or controlled by a foreign government is banned outright, which kills Chinese exports and, Mads argues, opens a window for European makers.
Andrew pointed to BAE buying Marconi in the late 90s to get a US footprint, and asked why Europe isn’t more belligerent with its own manufacturing requirements. I imagine more protectionism is coming. If it’s smart, I’d welcome it.
The AI price war: frontier intelligence gets cheaper
On Tuesday Anthropic dropped Opus 5.5, about 40% cheaper than Opus 5. OpenAI’s new Sol and Luna landed at roughly half the price of their predecessors.
Mads called it a proper price war. Back in June we said Anthropic pricing Fable very high was an experiment in the pricing power of frontier intelligence. “It seems that they have run this experiment and it has largely not succeeded.” Opus 5.5 is in many ways more powerful than Fable 5.1, and much cheaper.
The bills tell the story. Anthropic went from a $10bn to a $65bn run rate in seven months, and the companies paying are now going, ooh gosh. We’re one of them. Our bills are lower than last month because we route across around 50 paths, using different models for different jobs.
Mads brought a chart of the Pareto frontier, intelligence per dollar across every model and reasoning setting. Over the summer OpenAI took almost all of it, right up to Anthropic’s most powerful models. Haiku sits far behind comparable models at its price point. “It’s not because Anthropic can’t make a better model. I promise you, they can.” That’s pricing power on sticky, embedded workloads.
Andrew and I disagreed on how sticky. He noted GPT-3.5-level performance has dropped 280-fold in cost in two years, but argues enterprise switching is never plug and play: contracts, model nuances, API integrations, SLAs. My view: this isn’t an Oracle database or an SAP install. Routing has to be built into the infrastructure, because nobody knows how this model war ends.
One launch worth watching: Jev, from TypeSafe AI, founded by Diogo Almeida, a co-author of the OpenAI paper behind ChatGPT. It’s a new model family built for classification, returning probabilities for defined outcomes (good lead, yes or no) at about a fiftieth of the cost of Sonnet, according to Mads. A nice nod to Jevons, too.
The AI money window is narrowing
SB Energy’s IPO has been delayed. Nvidia is trading at its cheapest forward multiples in a decade. Anthropic’s IPO has slipped to November.
Mads doesn’t read it as panic. SB Energy has no data centres running and its backlog rests on a single OpenAI lease. Nscale has revenue, with Microsoft and Anthropic as customers, and looks set to list in October. Investors want something tangible before they underwrite. Fair.
Nvidia’s low multiple is against a forecast of another 70% growth next year. If you believe that continues forever, it’s cheap. “Nobody’s done that in history.”
What would worry Mads is Anthropic slipping into 2027. “They’ll matter if they happen. And they also matter if they don’t happen.” For Europe, Andrew still sees a cultural blocker in the London markets, and Mads couldn’t picture Nscale being better placed listing here. Neither can I, yet.
Agents loose in the hoose
Amazon blocked Meta’s Muse agent this week. Mads thinks Amazon is right, short term. He expects a split between walled gardens and the Shopify shop owner happy to take agent traffic. Same rule as the high street: it’s your store, you can show someone the door, and you live with the consequences.
The bigger question is breaches. Twenty-two countries want an IAEA-style global AI body. The US and China have no interest, and Andrew noted the UK hasn’t backed it either. Meanwhile Hugging Face CEO Clem Delangue, after their breach, found Fable’s own guardrails blocked them from using it in defence, so they fell back on open source. His ask: accelerate, and give defenders access.
Andrew’s reality check on the Australian hack: it exploited an ordinary weakness in a legacy public system, automatically and repeatedly. Autonomous persistence, not breakthrough hacking. Reused passwords, no 2FA, unpatched legacy systems. AI will find them. That’s today’s problem.
Liability is where this lands
I think liability is the anchor. Charlie Munger’s incentives and outcomes. Mads laid out where the EU has got to. The planned AI liability rules were dropped. Instead, software counts as a product, so anything shipped with an AI model after 9 December falls under strict product liability: prove a defect, not carelessness. Only individuals can claim, for injury, damaged property or lost personal data. If an agent sets your coat on fire, there’s liability. If it invests your money badly, there isn’t.
Andrew’s point: software has always shipped “as is”, so lawyers aren’t losing their jobs any time soon. AI causes the problem, AI solves the problem, AI legislates the problem.
Predictions
- Me: Anthropic IPO at the end of October. Mads immediately told me it had just been pushed to November. So much for that.
- Me: On model pricing, “I think we’re on a race to zero.”
- Andrew: Once enterprises are locked into a provider’s integration, usage rises, and “once you’re locked in, what do people do? They raise the prices.”
- Mads: A global IAEA-style AI regulator “is not gonna happen. It’s fiction.”
- Mads: On mandatory AI incident reporting, “I think others will follow.”
- Andrew: Breakthrough AI hacking ability “will inevitably come”, with quantum breaking RSA further down the line.
Deals of the Week
- Euclid (Mads): Eindhoven, €200m Series A. “Not your grandfather’s Series A.” Backed by Samsung, EQT, the Scale Up Fund, Somerset Capital and Innovation Industries, and chaired by former ASML CEO Peter Wennink. Inference chips built in Europe, claiming 100x better power efficiency and cost per token. No independent tests yet, first systems in 2028.
- ABP (Andrew): The Dutch pension fund is committing €1bn to venture and tech over three years, €250m of it into the €5bn European Scale Up Fund. Andrew’s cold water: in 2024 roughly 0.01% of European pension assets went into European VC, one euro in ten thousand (Atomico). Don’t rain on my parade, gents. Mads joined the dots: ABP backs the Scale Up Fund, the Scale Up Fund underwrites Euclid.
- Standard X (me): A 10 million seed for a London startup working on isotopes for cancer, starting with alpha therapy. Backers include Vsquared, First Minute and Brevan Howard. London seed, bada bing, bada boom.
Notable Quotes
“If you wanted to buy time you’d try to buy a few years, I mean not a few months.” – Mads Jensen
“It’s a lot easier to hide a chip than it is to hide a two-story high missile launcher.” – Andrew J Scott
“It’s not because Anthropic can’t make a better model. I promise you, they can. I promise you they have one. They’re just holding it back.” – Mads Jensen
“All the people that don’t matter in this, they want it. And the two that really do matter, they don’t want it. So it’s not gonna happen. It’s fiction.” – Mads Jensen
Frequently Asked Questions
The main concrete result was an extension of the US-China trade truce to 10 January, shorter than the six months some had hoped for. On Upside 109, Mads Jensen said the US wants to keep the pressure on while a bigger framework is negotiated. An AI hotline between the White House and Xi's office was also discussed.
Anthropic released Opus 5.5 at about 40% below Opus 5, and OpenAI's Sol and Luna came in at roughly half their predecessors' prices. According to Mads Jensen, it shows the pricing power of frontier intelligence is weaker than expected, with pressure from OpenAI and open source models.
Yes. A model lab that has an incident in Europe must report it without undue delay, an obligation in force for more than a year. Mads Jensen contrasted this with OpenAI taking around a month to tell the Australian government its agent had breached government servers.
The EU dropped its planned AI-specific liability rules and instead treats software as a product. Anything shipped with an AI model after 9 December falls under strict liability, so a claimant must prove a defect, not negligence. Only individuals can claim, for injury, damaged property or lost personal data, not financial loss.