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.

Burnham’s Britain

Andy Burnham took Makerfield with more than 50% of the vote. Seventh prime minister in ten years. Continuity for markets, gilts calm, sterling up a touch. The FT was already running his ten-year plan the day after the vote.

Mads’ Starmer scorecard first. Two years, some real work on long-lead stuff. Planning Act, grid queue reform, progress on nuclear. 38 of 50 Matt Clifford AI plan recommendations pushed through. But the £25bn national insurance rise hit employment, gilts got more expensive, G7 investment-level is the lowest in the group, and UK stock market listings are the worst in 30 years. The money did not believe the economy was being fixed.

The deeper problem is the risk-averse reflex compounded over thirty years. Grenfell to the Building Safety Act to a de-facto freeze on London housebuilds. Robert Maxwell to the Pensions Act to UK pension holdings of domestic equity falling from around 50% to about 4%. Nobody can say yes, fifty people can say no. Mads’ line: risk is at the heart of growth, and if you cannot take risk, there is no reward. If the UK had kept growing at pre-2007 pace, GDP per person would be roughly £10,000 higher today.

Enter Burnham. Lomax pulled a Claude-generated summary that landed cleanly: a soft-left interventionist who believes the state should steer the economy, wrapped in northern everyman branding. More charismatic than Starmer, instinctively left of him. The FT quoted a government insider: “unfettered tech boosterism is a vote loser.” Lomax’ concern is that Burnham reads that quote and acts on it. Starmer trod a reasonable line with Microsoft, Nvidia and the Matt Clifford checklist. Burnham may pull back. He wants sovereign, British-owned data centres. He wants distribution and devolution. He is skeptical of the Musks and Zuckerbergs. Good for narrative, bad for the enterprise-scale AI investment the UK needs.

Mads’ upside case: sometimes it takes a chameleon to push reform a technocrat cannot. Starmer got skewered by his own party over a modest welfare reform. If Burnham uses his first three months to lay groundwork for proper reform in the autumn budget, the economy can breathe. Push borrowing and yields spike. Push taxes and the manifesto breaks. Push welfare and he keeps the runway. That is the narrow path.

Is AI creating a new world order?

Three stories from one week. The US lifted the ban on Anthropic’s Fable and Mythos, but not with the ID checks anyone expected. Instead, prompts now route through a government-supervised classifier that flags biological, nuclear or distillation intent and drops a gate. The White House also had OpenAI hold its best model back for approved partners. And Sam Altman suggested handing the US government a 5% stake in OpenAI, worth roughly $42bn, and floated that Anthropic, Google and Meta should do the same. The Trump administration already owns about 10% of Intel.

The frame is the important bit. The American state has moved from regulating AI to gating AI to potentially part-owning AI. Bernie Sanders wants half. Trump wants ten. The horseshoe closes. Mads’ preferred answer, if the state must own anything, is arm’s-length sovereign wealth, not the Treasury on the cap table of the incumbent it is meant to regulate. What Sam Altman is doing here in plain English is buying the incumbency. Give the government paper. Become the default. Get regulated in your favour.

The distillation angle matters underneath. Fable’s new classifier gates prompts that look like reasoning-trace extraction, the exact mechanism through which Chinese open-weight labs have been catching up. Lock the frontier down harder and China builds a parallel ecosystem in open weights while US firms lose access to a huge market. Let the frontier flow and China distills it faster. No clean answer, only trade-offs.

For European founders, the practical takeaway is the one Mads has been repeating: do not build on a single provider. The Anthropic switch-off, the classifier reroute, the 5% political capture story. Each is a reminder that the most important suppliers in the stack are becoming instruments of one government. Multi-model harnesses, open-weight fine-tunes, on-prem deployment.

How trade becomes war

Trump threatened 100% tariffs on any country running a Digital Services Tax, days before the 4 July US-EU deadline. Volkswagen doubled its planned job cuts toward the 100,000 mark. Alphabet’s €4bn EU fine got upheld after eight years. The lines between trade policy, industrial policy and economic warfare are gone.

The numbers frame the fight. The US has a $200bn goods deficit with the EU and an $89bn services surplus. Google and Apple each do more than $100bn a year in EMEA. Meta does around half. Europe already conceded a 15% general tariff, agreed to buy US LNG, agreed to invest. But digital was kept out of the deal, and digital is the one lever where Europe can actually hit back. The EU’s Anti-Coercion Instrument, drafted a few years ago and largely dormant, can hit procurement, services and even how certain IP is treated. Trump knows this. Hence the 100% threat.

The China side cuts the other way. Volkswagen used to sell into China as a growth market. It now competes with Chinese EVs that got good faster than the West believed possible. Germany imports more from China than it exports and has done since 2019. Tariffs on Chinese finished goods sound good until you remember that Chinese parts sit deep in the European industrial supply chain. Tariff those and you import inflation into your own economy.

Underneath both fronts is the same structural shift. China’s mercantilist posture treats consumption as a low priority. Manufacture, sell, and buy only what you need to learn how to replicate. That works for them for a while. It stops working for everyone else. The free-trade era assumed everyone would play by the same rules. That assumption has broken.

Quick hits from the rest of the week

Fusion and other flashes. Realta Fusion announced net-positive energy generation (still importing more energy than the reaction outputs, but a milestone). SpaceX showed investors a prototype of the Musk AI device. Iran barred IAEA inspectors from Fordow, Natanz and Isfahan. Gemini 3.5 Pro slipped to July while Google leaked engineers to OpenAI and Anthropic and lost ~$225bn of market cap in a day. Gemma 4 shipped, small enough to run on a laptop.

Market rotation, not meltdown. S&P +9.5% H1, Nasdaq +13%, Russell 2000 +22% (best H1 since 1991). Nasdaq peak-to-trough only 6%. Two trillion added to Micron, Intel and AMD in one quarter. Semiconductors now ~20% of the S&P, four times their weight six years ago. Memory margins above 80%, a first, with South Korea backing a ~$500bn plan to keep capacity coming. Warning: hedge funds sold tech at the fastest pace in a decade while retail bought the dip.

FDEs are back. MIT study: 95% of enterprise AI pilots deliver no measurable profit impact. AWS spent billions this quarter on an FDE unit. Microsoft announced its own billion-dollar version this week. Anthropic, OpenAI and Mistral are going direct into large corporates with smaller integration teams. Mads’ framing: not new. IBM sent engineers to install mainframes in the 1970s. The pendulum swung to package-software configuration in the 1990s. Now the tech is early enough that it swings back.

Predictions

Dan Bowyer: the AI bubble does not pop, it slowly transforms. The rotation this week is the first look at that transformation. Money moves from mega-cap spenders to suppliers, then to the utility layer, then on-prem into the enterprise. Less hectic, more baked in.

Mads Jensen: retail buying while hedge funds sell is a top-of-market signal. “One of the four horsemen of the apocalypse is when the smart money is selling and the less sophisticated money is buying. That’s typically a top of the market signal.”

Deals of the Week

Lomax: Quantum Systems. Munich-based autonomous drone maker, $1.2bn Series D announced Thursday. Founded 2015. Nearly 20,000 missions flown in Ukraine in 2025. Started civilian, went dual-use, now defense-heavy. The FT teased a merger with Stark (backed by Founders Fund) to create a European neo-prime alongside Helsing. Reminder: seed the boring stuff before it becomes hot.

Mads: Equilibri. Prague-based AI lab. Creandom’s largest-ever single bet, Series A at $500m+. Reinforcement-learning trading agents. Founders built Deep Stack, the first AI to beat professional poker players, over 44,000 hands, on a single consumer GPU. Advisors: Murray Campbell (co-built Deep Blue), Michael Bowling (led Alberta Deep Stack), Rich Sutton (Turing Award, father of reinforcement learning). Chess to poker to public markets.

Notable Quotes

“A soft-left interventionist who believes the state should steer the economy, wrapped in northern everyman branding.” – Lomax Ward, quoting a Claude-generated summary of Burnham.

“Unfettered tech boosterism is a vote loser.” – FT government insider, cited by Lomax Ward, on the political challenge for UK AI policy.

“Risk is at the heart of growth. If you can’t take risk, there’s no reward.” – Mads Jensen, on Britain’s risk-free society compounding.

“One of the four horsemen of the apocalypse is when the smart money is selling and the less sophisticated money is buying.” – Mads Jensen, on the tech market rotation.

“It’s just what you’d expect.” – Mads Jensen, on why the FDE boom is a return to how the industry always worked.