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.

Orbán out, Europe exhales

Hungary was, until this weekend, the most corrupt country in the EU by Transparency International’s reckoning. Orbán started as a smart, progressive, cool operator and spent the last decade as Russia and China’s circuit breaker inside the Union. Ukraine financing, sanctions packages, rearmament loans, all slow-walked or blocked.

Peter Magyar, a renegade from Orbán’s own party, has now trounced his former boss. The seat chart was decisive. Mads put it neatly: people didn’t turf Orbán because they’d cooled on the EU, they turfed him because they wanted to be closer to it. Worth remembering from this side of the Brexit line, because the lazy read is that everybody in Europe is waiting to leave. They are not. In a lot of places the Union is more popular than the national government.

What gets released? Around €90bn of blocked EU funding. Twenty-odd sanctions packages. Ukraine financing. Hungary’s own safe application into European defence investment. Real money moving through a network that’s been sticky for a decade.

Magyar is also talking about turning Hungary from a BMW-and-Chinese assembly line into a “knowledge-based economy.” The Poland playbook. Poland took fifteen years and started with better institutions. This one will take a while. On defence and foreign policy, expect fast flips over the summer. On the 28th regime, don’t pin your hopes here.

Anthropic keeps eating

Six months ago Mads called it. Frontier labs will either become application companies or consolidate into hyperscalers. We’re now watching it play out.

Anthropic’s app builder leaked this week. They’ve hired Workday’s CTO. Mike Krieger (Instagram co-founder, Anthropic’s CPO since 2023) shipped Model Context Protocol, drove the Claude Code revenue ramp, joined Figma’s board under Dylan Field last summer, and has now left it. Reporting says Anthropic is building a design tool to challenge Figma. Krieger is one of the best product operators of his generation, and he is moving fast on a full-stack Anthropic.

If you’re a B2B SaaS founder on a thin wrapper around Claude, the margin story has collapsed. When Anthropic owns the model, the interface, the design, the sales motion, and the enterprise value, they keep the stack.

The European question: who’s our Anthropic? Mistral gets listed in the brackets and is doing better than we feared, picking a solid lane as the AI integrator for European enterprise. Real market. But we don’t have the capital, the operator talent, or the vertical integration to go toe-to-toe on a full-stack play. Meanwhile Anthropic is taking a big new space near King’s Cross, quietly the most bullish London AI story of the quarter.

Mythos vs GPT-5.4 Cyber

Two benchmarks, two very different tests.

GPT-5.4 Cyber scores 76% on Capture the Flag: known patterns, engineered problems, controlled environment, find the token. Mid-level pen tester stuff. Mythos scores 72.4% on exploit-to-vulnerability in the wild. Here’s source code. Find the flaw. Prove the exploit. Real code, no training wheels, and Mythos is a general model rather than a cyber-labelled product.

That’s why Mythos is gatekept behind Project Glasswing with $100m of tokens handed out to vetted testers. OpenAI is shipping. Anthropic is containing. Tells you which one they think is more dangerous.

My prediction: within twelve months we see a major cyber breach publicly attributed, rightly or wrongly, to an AI model. Not necessarily Mythos. It triggers the first AI-specific liability legislation. Start pricing that into your cyber portfolio. On the Chinese distillation angle: right now only a handful of labs have the best models, and there’s a genuine window where the West has a capability advantage. Are we using it? Are we the good guys? Pick your peep show sketch.

The solo unicorn

Sam Altman said it. Matthew Gallagher says he did it. MEDVi, a GLP-1 resale outfit, zero to $400m in year one. $1.8bn projected for 2026. Two employees. $20k of ChatGPT, Claude, Grok, Midjourney, and Runway. 15%+ net margins. Him and his brother.

I’m hedging every number in that paragraph for good reason. Class action suits, grubby marketing, Trustpilot reviews that don’t quite umm meet reality and sales that don’t align with the web traffic. Do your own research.

However. Distribution is a real business. Costco is distribution. Amazon is distribution. Two people, a bunch of tokens, and a wave now underpin $150m a month of economic activity. Even if Gallagher’s numbers are half true, that’s the real story.

Venture gets a longer and fatter barbell. The two and three-person teams running distribution on other people’s rails don’t need us. The teams doing hard frontier work, 30 to 40 people, real engineering, real moats, still will. Brand is one of the last real moats. More of the venture dollar now goes into marketing, go-to-market, and trust. And niche businesses nobody used to fund will survive because two founders have the best lawyer, data scientist, analyst, and coder in their pocket for pennies.

Compute is still scarce

OpenAI cut planned capex from $1.4T to $600bn and everyone shouted “bubble.” Nope. Anthropic has taken pole position and OpenAI can’t justify the old number on their current trajectory. Anthropic is reportedly close to accidental profitability because they literally can’t buy compute fast enough. Great phrase, that.

The real signals are not lab press releases. They’re the picks and shovels. GPU spot prices up roughly 50% in two months. CoreWeave minimum contract one year to three. ASML raised guidance. TSMC, who spend years telling Nvidia to calm down on capacity asks, lifted capex 35% year on year. And somewhere out there, a trainer company (Mads and I thought Allbirds, couldn’t quite place it on air) has pivoted from shoes into selling compute. When shoe people are pivoting to GPUs, you are not living through a bubble correction.

Musk in space

TeraFab. $25bn. Intel as foundry. Based in Austin. Tesla, SpaceX, and xAI as captive customers. 1TW of chip capacity, twenty times global AI compute today, with 80% of it parked in orbit.

The physics is absurd. Cooling one gigawatt in a vacuum needs about 800,000 square metres of radiator. Chips cook without atmosphere to insulate them. Three years, in Musk’s framing, until it’s cheaper to run AI in space than on Earth. Look, a lot of people have lost a lot of money betting against Musk. Starlink is quietly doing around $11bn in margin and is the real business inside SpaceX. He laid off 70% of X and the platform survived. Directionally, orbital compute is probably coming. Three years is PR. He needs a TeraFab story to justify the 2026 raise.

Deals of the week

Synera (Germany), $40m Series B. AI orchestrating the engineer’s toolchain across Siemens, Dassault, PTC, Autodesk, and Ansys for customers including BMW and NASA. Think Claude for engineering. Right down the middle of where European industry needs to go.

Helical (London), $10m round two. A virtual AI lab running drug discovery hypothesis cycles before anything hits the wet bench. Compress the simulation half of that loop and you save quarters.

Notable Quotes

“Anthropic almost becoming accidentally profitable right now because they can’t acquire enough compute quickly enough.” – Mads Jensen

“We want to do business with China, right? We just don’t want to give China a veto over our foreign policy.” – Mads Jensen