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.

Anthropic still eating the AI economy

ARR moved from roughly $9-10bn at the start of 2025 to $45bn now, a 4.5x ramp in five months on a base that was already large. The Financial Times reported the figure this morning. The reported $50bn round in motion would value Anthropic at more than $1 trillion, against $380bn at the most recent close. In four months, the company’s market value has more than doubled while ARR has gone vertical.

The capex story finally has a coherent revenue counterparty. Anthropic announced a $200bn five-year commitment with Google Cloud beginning in 2027, equal to roughly 40% of Google’s ~$500bn cloud backlog. Anthropic and OpenAI together now account for more than half of the ~$2 trillion hyperscaler backlog. The arithmetic only works if the AI revenue line is real; for the first time, the line is moving fast enough to make it plausible.

The deeper structural shift is the joint venture model. Anthropic launched a JV with Blackstone, Hellman & Friedman and Goldman Sachs to deploy AI inside enterprise customers; OpenAI announced a parallel structure with TPG, Bain Capital and Brookfield. Two trends are converging: AI capability that needs human integration to extract value, and a private equity industry under pressure to show LPs how AI rescues depressed returns. Roughly 7% of US GDP and $8-9 trillion of PE-backed market cap is the addressable surface; even modest EBITDA uplift across that surface generates significant revenue for the AI side and a renewed fundraising story for the PE side. The implication for European founders is harder. The application layer where European companies traditionally compete is being colonised by the frontier labs themselves, working through US private equity distribution.

ASML overstating moat?

Christophe Fouquet, ASML’s CEO, told a Milken audience this week that “no-one is coming for us.” On the lithography hardware itself, the claim is defensible. ASML’s extreme ultraviolet machines are the product of two decades of work, and competitors are five to ten years behind on the EUV light source alone. Substrate, the Peter Thiel-backed challenger, has raised more than a billion dollars; Fouquet’s read is that it has “huge challenges” before any of that capital produces a competitive machine.

The claim is overstated where it stops being a hardware claim and starts being a full-stack AI claim. The Chinese strategic position does not depend on matching ASML chip-for-chip; it depends on matching it systems-for-systems. Huawei’s CloudMatrix 384 is the proof point. The cluster architecture is designed to compensate for individual chip performance gaps of roughly 9x relative to the latest Nvidia parts; with 384 accelerators networked tightly, system-level performance approaches what Western labs can achieve with far smaller numbers of frontier chips. DeepSeek V4 was trained at least partially on Huawei Ascend hardware. When access to the best hardware is restricted, Chinese teams design around the constraint at the level where their engineering and capital base is strongest.

The contrast with Jensen Huang is instructive. Huang is openly paranoid; he describes himself as working seven days a week against the possibility that someone is catching up. Fouquet is openly confident. The strategic point for European investors is twofold. First, the ASML moat funds the lithography supply chain that the entire Western frontier depends on; betting against it is betting against the supply chain. Second, the moat does not extend to AI capability. A world in which Chinese labs match Western performance through systems-level architecture, while ASML continues to print margins on EUV, is internally consistent. Hardware exposure and AI exposure are now distinct theses.

Brussels takes on Meta (again?!)

Meta has integrated AI into WhatsApp, which has more than 3 billion users globally and functions as core infrastructure across most of the world outside the United States. Brussels did not object to the integration; the objection is that Meta has not allowed competing AI models access to the same surface. Meta proposed that competitors pay to access. Brussels has now rejected that remedy. If the case proceeds, Meta could face fines of up to 10% of global revenue, in the range of $16-20bn, and a precedent of a European platform being forced to provide open access to AI competitors at no cost.

The line between essential infrastructure and proprietary platform is the substantive question. WhatsApp’s reach is utility-like; the investment was private. The right answer depends on a forward view. If messaging platforms expand into payments, identity and commerce, as X is signalling, the case for treating WhatsApp as infrastructure strengthens. UK faster payments and EU open banking produced more competitive financial infrastructure than the US ACH system; competition rules that look heavy-handed in the moment have, in some sectors, produced demonstrably better consumer outcomes.

A forced-access remedy creates a second-order problem: liability. When a third-party AI agent acts inside WhatsApp, ownership of the resulting fraud, content moderation breach or security failure becomes contested. The platform argues it is now a neutral pipe; the third party argues it cannot police what users do; regulators need a designated responsible party. A clean remedy on access has to be paired with a clean remedy on responsibility; the EU has not yet articulated the second. For European founders, the upside scenario is a WhatsApp opened up by regulatory action becoming a distribution channel for any AI agent that can integrate. A protracted enforcement fight, in contrast, slows everyone.

Quick hits from the rest of the week

Mythos and Washington’s AI safety pivot. Mythos prompted Dario Amodei to fly to the White House, and Trump’s team is advancing a 15-page draft executive order that would create FDA-style pre-deployment review for frontier models. CAISI has signed Google, Microsoft and xAI; Anthropic is not yet in the formal framework. An anti-regulation administration is treating frontier AI as strategic infrastructure on par with biotech and nuclear. Two readings are both defensible: necessary oversight after Mythos demonstrated unsafe capability, or incumbent capture that disadvantages smaller labs and open-source projects.

ARM and Spotify. ARM is up 84% year-to-date on the strategic pivot from licensing to designing AGI-class CPUs. The 1:8 historical CPU-to-GPU ratio is moving toward 1:1 as agentic workloads pull computation back to the CPU. Spotify printed 10% premium revenue growth and lost 5% on the week, with the miss in ad-supported revenue and rising customer acquisition cost. The market is pricing the AI-disruption tail on Spotify and the AI-tailwind tail on ARM.

UK Infinity Fusion Consortium. Bill Gates is lead financial backer; Type One Energy provides stellarator technology developed in Wisconsin; UK partners contribute superconducting magnet expertise from Tokamak Energy. Target is a 400 MW plant by the mid-2030s, supported by the new Critical National Infrastructure planning framework. The UK’s deeper question is whether it can stop relying on US risk-takers to anchor strategically important domestic projects.

Predictions

Mads Jensen: Anthropic does not IPO in the near term. With private markets throwing capital at the company at a trillion-plus valuation, the board can defer the decision. A board meeting on the IPO question is expected this month.

Andrew Scott: ASML’s monopoly does not last forever. “It’s only a matter of time. It’s not if, it’s when.” Given the size of the market and China’s strategic imperative to catch up and overtake the US, the position is strong today and finite over time.

Andrew Scott: UK strategic vision does not improve until a change of government. A coherent strategy across fusion, quantum, AI and defence requires the UK to take risk and commit capital against generation-spanning goals. On Andrew’s read, the current government has not shown that capacity, across the SDR and other sectors.

Deal of the Week

SAP acquires Prior Labs for €1bn. Prior Labs is a 15-month-old German company building tabular foundation models, pre-trained AI for the structured data that lives in spreadsheets, databases and enterprise systems. The seed round was €9m in February 2025, led by Balderton, Atlantic Labs and XTX Ventures. The price tag reflects two convictions: that enterprise data is overwhelmingly tabular, and that SAP needs proprietary AI capability tied to its data estate. A clean European speed-to-exit on a frontier-adjacent technical bet.

Notable Quotes

“If they keep going at this rate and you project that out over the next two years, then Anthropic will be a $250-300 billion revenue company.” – Lomax Ward.

“These people are not idiots; in fact, they are the opposite of that. They are throwing an insane amount of resource at this.” – Lomax Ward, on Chinese systems-level workarounds.

“The bottleneck has shifted from model capacity to enterprise implementation.” – Andrew Scott, on the Anthropic and OpenAI joint ventures.

“Frontier AI is no longer just a tech product; it is strategic infrastructure, alongside nuclear and biotech.” – Andrew Scott, on the Trump draft executive order.