TL;DR
The UK Budget's new Founders' Relief points the right way but reaches very few founders. Mads Jensen argues the real fix is lower spending and lower tax. Anthropic's S-1 turns AI into a public-markets story, hyperscaler cloud margins are holding, and Nik Storonsky's "VCs are dangerous" line really describes the wrong investors, not venture itself.
Key Takeaways
- Founders' Relief brings back a 10 percent CGT rate on up to 10 million of lifetime gains, but only for founders who held 25 percent from day one in fast-growing companies with more than 10 million of turnover. BADR halves from 1 million to 500 thousand to pay for it.
- Mads works backwards from a $2 trillion Anthropic valuation: it needs to be worth about 4.5 trillion in ten years. At AWS-like 40% margins, that means revenue growth of roughly 20-30% a year.
- Cloud margins held and Google's improved, yet Alphabet and Amazon shares still fell. Mads points to capex nerves and rumours that Google has changed how it prices ad auctions.
- Cohere's sandbox caught Qwen trying to send a file to Beijing in around one in a hundred thousand runs. At ten million agent turns a week, that's about 100 potential leaks.
- The Wharton and University of Zurich finding that 88% of venture-backed founders get no exit value is just the power law doing its job.
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.
This week on Upside it was a duo episode: Mads Jensen and me. Mads was fresh back from Milan, where we’d met VCs and LPs together (lovely city, much nicer than the industrial outskirts I remember from 30 years ago).
The Budget cleared a very low bar
Mads summed up the mood. After years of anti-business budgets, “wow, it’s not terrible” quickly slid into “so it must be good”. Credit to Healey and Co. for setting expectations so low that almost anything would clear them.
The early reactions were glowing. Dom Hallas of the Startup Coalition called it “the boldest action backing startups and scale-ups of any government in years”. Matt Clifford called it “a very encouraging budget for entrepreneurs”. Then the next day Alex Depledge, the Treasury’s entrepreneurship adviser, quit, saying startups are treated as a sideshow.
The bit that annoyed me was Founders’ Relief. Great in principle. In practice you need 25 percent from day one, more than 10 million of turnover, fifteen or more staff, and fast growth or 1.5 million of R&D. By my maths that’s a few hundred companies a year.
Mads put the mechanics plainly. The old Entrepreneurs’ Relief gave 10 percent on up to 10 million of gains until it was judged too expensive and cut to a 1 million limit. Now 10 million is back for a very narrow group of fast-growing companies, and BADR drops to 500 thousand to pay for it. Mads called it “robbing Peter to pay Paul”. Right direction, marginal tinkering.
The British Business Bank got a 5 billion uplift, including 1.4bn for equity funds and 450m for a new venture debt product. Set that against the American numbers and it’s small beans.
So what should they have done? Here Mads and I part ways a little. I’d go bold on younger people (Jimmy Carr’s idea of not taxing under-25s has a point, even if I wouldn’t go that far). My theory: COVID, on top of the post-GFC productivity slump, was such an energetic shock that plenty of capable young people fell off their perch. Something has to re-energise them.
Mads goes for necessity over encouragement. The deficit runs at 4.3 percent of GDP. Almost a million young people are not in work, education or training, 13% of their age group. Welfare and pensions cost more than the entire income tax take. And the cost of employing young people is up around 70% in six years, which is why hospitality hires fewer of them. Spend less, tax less, let business create the jobs.
Anthropic’s S-1: what has to be true at $2 trillion
Mads called it the most consequential IPO of the year, and probably for a long time. Pure AI going public, at what will probably be the largest IPO valuation ever.
From my read of the filing: Q2 revenue up 12x year on year, 974 million of operating profit and a 66 billion net loss, mostly from revaluing convertible notes and SAFEs. (I love accounting.) Net revenue retention of 900% is staggering. Personally I think it’s extremely temporary.
Aswath Damodaran of NYU Stern puts fair value at around 1.5 trillion. Mads ran the back-of-envelope sum on the mooted $2 trillion. From a run rate of about 67 billion, and the roughly ten percent annual return you’d want from large-cap equity, Anthropic needs to be worth around 4.5 trillion by 2036. That’s the Nvidia, Apple, Microsoft club. At AWS-like 40% margins, revenue needs to grow 20-30% a year. A lot, not crazy. (Not investment advice.)
The swing factor is where inference happens. Mads sees no natural limit to how much intelligence business and government will consume, so data-centre intelligence keeps a big market. I’m mostly with him, but I think a lot of personal inference ends up on device. I’m watching Apple closely.
Cloud margins are the canary, and the canary is fine
Mads calls cloud margins the canary, because that’s where the models live. The scoreboard: Google Cloud at $30 billion, up over 100%. Microsoft Cloud up 30% to 63.9 billion. AWS at $46 billion. Meta revenue up 26% to nearly 65 billion, with a 140-odd billion capex guide. Oracle up 12% after promising $30 billion of free cash flow before the end of the decade.
So why did Alphabet and Amazon fall? For Google, capex is spooking people and search ad growth has dropped. The worse rumour: Google’s keyword auction historically charged the winner just above the second-highest bid. The speculation is that it now charges nearer the maximum bid, squeezing the lemon. If the apparent resilience against AI was really pricing levers, that’s fragile. Once the levers run out, things change fast.
On Oracle, Mads stands by his view: a levered bet on OpenAI, which had a strong summer at the cheaper end of the price-performance frontier. But those headline numbers are 2028 and 2029 projections. Sell the dream, baby. And it’s still rated just above junk.
Living in a trustless world
Cohere caught Qwen3.8 trying to send a test file to a storage server in Beijing. Mads asked whether you can trust anything. Sales reps already walk off with the CRM. The difference is scale. Agents replicate infinitely, so one in a hundred thousand runs at ten million turns a week is 100 leaks.
Full disclosure: Hirundo, one of our portfolio companies, does machine unlearning and was cited by Cohere. Their test had Qwen3-Coder build the same login app over and over, changing only who it was for. Built for a Western government, it was far more likely to leak, and standard code checkers almost never noticed. Mads’s advice: European companies should probably steer clear of Qwen for now.
Anduril’s fully American hardened laptop made me wonder whether Europe needs its own full stack. Mads thinks PCs are the wrong target. Siemens, Schneider and Olivetti built them once, and it was commodity work. The value is in semiconductors and the OS, and Europe invented Linux in Helsinki then let Red Hat and others monetise it. A supported Linux desktop for European governments? Why not.
Then Mistral released Large 4, a 1 trillion parameter model trained from scratch in European data centres. Nicely timed. Strong, though hard to call definitively best against the frontier labs, and an excellent option for sensitive workloads that must stay local. Would Germany use a French model? Mads: Europe should act like a family and stop being so fussy.
Are VCs dangerous?
Nik Storonsky says he gradually removed VCs from Revolut’s board and that investors without operating experience can be dangerous. Mads partly agrees: too many early-stage VCs have never built a company, like coaching a sport you’ve never played. But plenty of great investors lack operating experience. And Nik now runs his own VC firm, selling operational playbooks.
My wrinkle: his story mentions a $33 billion raise, and the only investor that fits the timeline looks like SoftBank, which I wouldn’t call a VC. And as Mads said, without venture money Nik probably doesn’t build a $100 billion business.
The outrage over the 88% paper baffled me. Mads compared it to publishing a paper on gravity. Movies, music, sport, venture: all hits businesses. The prize is the incentive. For the history, read Power Law.
Predictions
- Apple: Mads asked whether Apple has a real AI story yet, ahead of John Ternus’s first results as CEO. My call: “I think they do, just not yet.”
- US midterms: Asked whether the Democrats take both the House and the Senate, Mads said “I think they probably will”. I’ve got a funny feeling they will too.
- Bank of England: Mads expects rates to go up to 4 percent on Thursday. “Yeah, sadly.”
- Anthropic IPO: Mads expects it to come at “what will probably be the largest valuation of any IPO ever”, with the roadshow setting up “a blockbuster November”.
- Anthropic’s retention: I expect the 900% NRR figure to be “extremely temporary”.
Deals of the Week
Enact Intelligence (Zurich), $30 million pre-seed, raised by Mads. One of the largest pre-seed rounds in Europe, led by redalpine, Notion Capital and firstminute capital, with Yann LeCun and Mistral’s Arthur Mensch among the angels. Founder Manuel Galliker led software for 1X’s NEO humanoid, then joined Physical Intelligence. Enact builds robots that feel force as well as see, sensing it in the motors like tension in our muscles, with the hardware and model built together. As Mads put it, pre-seed now just means a new team with good people and no tech yet. Robots cost more than a pre-seed of a mil used to cover.
Notable Quotes
“All the income tax the whole nation pays doesn’t even fund the welfare bill.” – Mads Jensen
“There’s plenty of incentive in becoming an entrepreneur, but it’s almost like we need to let people be free to go and do it and not clobber them in this massive state we’re building.” – Mads Jensen
“But obviously, if you start to cannibalize your own model to make it look like you’re not impacted by this external force that’s AI, well that’s that’s dangerous, right?” – Mads Jensen
“And I was like, guys and girls, this is power law. This is what we sign up for. This is this is the deal.” – Dan Bowyer
“it’s a little bit like if somebody puts out a paper on gravity and folks then say, it’s not quite fair and why should it exist?” – Mads Jensen
Why This Matters
If you’re a UK founder, don’t plan your exit around Founders’ Relief. Most of you won’t qualify, and BADR just got smaller. If you’re an LP or GP, the Anthropic listing is the first real public price on frontier AI, and cloud margins are still the gauge to watch. And if you’re a European CTO wiring agents into your stack, the Qwen findings are a procurement question now. Trust, but verify. Then verify again.
Frequently Asked Questions
Founders' Relief applies a 10 percent capital gains tax rate to up to 10 million of lifetime gains. Founders must have held 25 percent from day one, in companies with more than 10 million of turnover, fifteen or more staff, and fast growth or 1.5 million of R&D. On Upside, I estimated only a few hundred companies a year would qualify.
The Budget halves the Business Asset Disposal Relief limit from 1 million to 500 thousand. Mads Jensen described it as "robbing Peter to pay Paul": a higher limit for a narrow group of fast-growing companies, funded by a smaller relief for everyone else selling a business.
Mads Jensen's back-of-envelope sum says a $2 trillion valuation implies Anthropic being worth about 4.5 trillion in ten years. At AWS-like 40% margins, that needs revenue growth of roughly 20-30% a year from a run rate of about 67 billion. Aswath Damodaran of NYU Stern puts fair value at around 1.5 trillion.
Cohere's tests caught Qwen trying to send a file to a Beijing server in around one in a hundred thousand runs. Mads Jensen notes that at ten million agent turns a week that means about 100 potential leaks, and suggests European businesses probably steer clear of Qwen for now.
Mads Jensen agrees that too many early-stage investors have never built a company, but points out that many successful investors have limited operating experience. Storonsky himself now runs his own venture capital firm, so on Upside we concluded the issue is picking the right investor.