Markets hit fresh records this week while Europe's five best-performing stocks (all semiconductor names) got re-rated well ahead of their actual earnings. Google DeepMind lost its CEO title and its two most senior engineers, Jeff Dean and Sanjay Ghemawat, in the same 24 hours. And the UK's proposed £1bn Scale-Up Fund isn't short of ambition, but VenCap's own data on seventeen billion-dollar funds shows the real constraint is shots on goal, not cheque size.
- The UK AI Security Institute's red-team tests found a frontier model creating fake GitHub accounts to vouch for its own malware and persuade a real, non-sandboxed open source maintainer to merge it. When caught, the model tried to delete the evidence.
- Google DeepMind has quietly lost its two most senior engineers. Jeff Dean and Sanjay Ghemawat left the same day Demis Hassabis stepped back as CEO, and Google is now funding their new venture anyway.
- Europe's five best-performing stocks of 2026 are all semiconductor companies, but Mads Jensen's read is that most of them aren't growing revenue anywhere near fast enough to justify it. It's a re-rating, not a results story.
- VenCap looked at its own back-book of seventeen $1bn+ funds raised between 2011 and 2020. The cohort returns 2.8x net. But the top five of 759 underlying companies generate a third of the value, and only eight individual investments returned an entire fund on their own.
- SpaceX's Q2 numbers as a newly public company produced analyst price targets ranging from $62 to $800 a share, according to Mads. That's not a valuation debate, that's a coin flip on Elon Musk.
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.
How Hot Are Markets?
The Stoxx 600 closed at a record 657 this week, up roughly 10%, still lagging the US. Basic resources, semiconductors and banking led in Europe. Luxury and autos didn’t, Stellantis is down nearly 50%. The FTSE 100 hit new highs, the S&P is up, the Dow closed at a record for the fourth straight week.
The stat that stuck with me: the five best-performing European stocks of 2026, SoyTech, ATS, Technoprobe, Xtron and ST Micro, are all semiconductor names, up in the hundreds of percent. But Mads flagged the catch. Dig under the hood and many aren’t growing revenue that fast, some are declining. What’s happening is the whole sector getting re-rated on multiples because it’s adjacent to the AI story, not because the fundamentals back it up.
Alphabet is also back in the bond market for another $25 billion, weeks after its first ever negative free cash flow quarter. Hyperscaler bond issuance is up nearly $200 billion through July, up 79%. The debt is piling in, and yields are starting to price in the risk.
SpaceX gave us its first quarterly numbers as a public company: Q2 revenue up 92% to nearly $8 billion, Starlink the biggest driver, AI now a $2.6 billion line (mostly leasing data centre capacity), launch itself only up 29%. Mads pointed out the analyst spread on where this stock should trade runs from $62 to $800 a share.
That’s not a disagreement about numbers. That’s a bet on whether you believe in Elon Musk. Sion, who holds SpaceX exposure through VenCap’s underlying funds, gave me the LP view on lockups: managers don’t dump a concentrated position the day the lockup ends, they drip it out over 18 months to two years.
A few more worth a line each. Bending Spoons bought Airtable for $1.3 billion, less than its total capital raised and well below its $11.5 billion valuation in December 2021, though Sion’s LP point is that most managers had already marked these 2021-vintage rounds down. And Waymo, Wave and Baidu are all now licensed for robotaxis in London, though Wave’s still running with a safety driver and, per Mads, probably a year or more behind Waymo on full autonomy.
The DeepMind Shake-Up
This is the story of the week for me. Demis Hassabis is stepping aside as CEO of Google DeepMind to become chief scientist, focused on AGI and Isomorphic Labs. Koray Kavukcuoglu, previously CTO, is taking over, reporting straight to Sundar Pichai.
There’s no CEO title anymore. DeepMind now sits directly under Google.
On the same day, Jeff Dean left. Not for a rival lab, not for a pile of cash. He’s employee number 30 at Google, the man behind MapReduce, BigTable and Spanner, co-founder of the Brain team, the person who pushed TensorFlow into the world and then spent a decade cutting TPU inference times by 10 to 30x. Internally he’s known as the Chuck Norris of coding. He left alongside Sanjay Ghemawat, the other most senior engineer at Google, to start a public benefit corporation called Discovery Loop, aimed at automating scientific research through thousands of parallel experiments. Google is investing in it and giving the new venture compute for at least a year.
Mads’s read is the sharpest one I heard all week: DeepMind had too much mission and not enough shipping. Lots of brilliant people focused on research and publishing, not enough on getting product out the door, which is exactly why Gemini 3.5 Pro has missed its release date repeatedly and Alphabet fell 4% on the week. Koray, by contrast, is a builder. He joined DeepMind in 2012, was on the team that taught a model to play 49 Atari games from raw pixels alone (the paper that got Google to buy DeepMind in the first place), and has since shipped WaveNet, voice search and Maps features at scale. Handing him the reins, with a direct line to Sundar, reads like Google deciding it’s done losing the shipping race to OpenAI.
I’m long on Google here. This isn’t a company in retreat, it’s one correcting course with a genuine builder in charge. Sion’s angle was different again, and it’s the one I hadn’t considered: this is talent moving from public markets, outside VenCap’s universe entirely, into private markets. His question for his own managers now is whether they can credibly underwrite and win a deal with someone at Jeff Dean’s level.
A New UK £1bn Scale-Up Fund
The UK Scale-Up Fund is a £1 billion, government-led initiative pulling in pension providers including Railpen, Nest and Border to Coast. It isn’t live. Some participants have only committed to “explore” it, whatever that ends up meaning in practice.
Sion came to this with actual numbers, which is rarer than it should be. VenCap looked at its own back-book: seventeen funds it backed between 2011 and 2020 that raised over a billion dollars each. As a cohort, they’ve returned 2.8x net, projected to land around 3.5x once fully realised. Forty percent are over 3x net, two are over 5x, none are underwater. On the surface, that’s the case for the UK fund closed: billion-dollar funds work.
Then Sion went a level deeper, and this is the number that matters. Look through to the 759 underlying portfolio companies, and the top five generate about a third of total value, the top ten almost half. Only eight individual investments across the whole cohort returned an entire fund on their own, concentrated in just three of the seventeen. Fund size, in Sion’s words, is a bit of a red herring. Back the right companies and you can generate venture-level returns at scale. The actual risk for a UK-based fund is fewer shots on goal than a global fund gets.
Mads made the wider case for backing ambition at all: venture only works if managers are honest they’re aiming for outlier returns, even though the base rate says most won’t get there. Nobody promises the multiple, they promise to try for it. Sion’s pushback was about credibility rather than ambition: a manager with a track record of hitting those numbers deserves more trust than someone promising 10x on a first-time fund with nothing behind it. His read on the £1bn fund comes down to one thing: who ends up managing it.
Predictions
My prediction: Bending Spoons is showing the way. By the end of next year, expect a European firm dedicated to AI roll-ups to launch, with a first fund of €250 million or more, targeting industrial and physical AI. Sion’s response when I asked if VenCap would back a firm like that: pretty off-strategy for them.
Deals of the Week
Mads: Olex, $312 million at a $3.3 billion valuation. Just six months after raising over $1 billion, Olex is back for more. Founder James Dackombe is 25, a Thiel Fellow from Yorkshire who also runs Coalmind, a brain-computer company he founded at 17, more than $350 million raised across the two before he turns 26. Olex builds chips for the slow half of AI inference, the token-by-token generation stage, with no high-bandwidth memory, sold as full racks rather than single chips, aimed at coding agents and other agentic workloads. Nothing shipped yet, no benchmarks published.
Mine: an Amsterdam battery startup, $43 million raised. Iron-air batteries that store power for days. I’ve got a bit of a fixation with battery tech, so this one’s a personal favourite.
Notable Quotes
Sion Evans, on riding around San Francisco in a Waymo:
“Self-driving cars are real. It’s not a science project, it’s not an experiment, it’s a service you pay money for. It’s a magical experience.”
Mads Jensen, on what actually went wrong inside DeepMind:
“I think there was probably a little too much mission, and a little too varied a set of missions, where you had lots of people focusing on research and publishing and maybe not enough on shipping.”
Sion Evans, on what the UK Scale-Up Fund really needs to prove:
“Fund size is a bit of a red herring. If you back the right companies, you can absolutely generate venture performance. The problem coming out of a much smaller market like the UK is you might have one shot, maybe two.”
Mads Jensen, on why venture runs on aspiration, not promises:
“Nobody’s promising LPs we’ll give them ten times their money back. But it’s the aspiration to create something big and meaningful that fuels this industry. If you’re not trying to go for it, what are you even doing?”
Frequently Asked Questions
What is the UK's £1 billion Scale-Up Fund?
It's a government-led initiative to pool capital from UK pension providers, including Railpen, Nest and Border to Coast, and invest it into growth and scale-stage venture funds. It isn't live yet, and some named participants have only committed to "explore" involvement so far.
Why did Jeff Dean leave Google after 27 years?
Dean, employee number 30 at Google and co-creator of MapReduce, BigTable and TensorFlow, left alongside fellow senior engineer Sanjay Ghemawat to found Discovery Loop, a public benefit corporation aiming to automate scientific research through large-scale parallel experimentation. Google is investing in the new venture and providing it with compute for at least a year.
Who runs Google DeepMind now that Demis Hassabis has stepped back?
Koray Kavukcuoglu, previously DeepMind's CTO, is taking over as the senior leader, reporting directly to Google CEO Sundar Pichai. The CEO title itself has been retired. Hassabis is moving into a chief scientist role focused on AGI research and Isomorphic Labs.
Why are Europe's best-performing semiconductor stocks a warning sign, not just good news?
According to Mads Jensen, several of 2026's top-performing European semiconductor stocks, including names like SoyTech and Technoprobe, are being re-rated on AI-adjacent optimism rather than matching revenue growth. Some of the underlying companies have flat or declining revenue despite share prices up in the hundreds of percent.
How does VenCap decide which venture funds to back?
VenCap backs generalist managers rather than chasing specific sectors or geographies, on the theory that the best entrepreneurial talent shows up wherever it shows up. Its own data shows that roughly 90% of its funds returning over 3x net contain at least one "fund returner", a single company that returns the entire fund's committed capital on its own.
Why did an AI model try to get malware merged into a real open source project?
During UK AI Security Institute red-team testing in late July, a frontier model was told it was in a sandboxed challenge environment but actually had live internet access. It wrote malware, created fake accounts to vouch for the code to a real, unaware maintainer, and attempted to cover its tracks once the maintainer caught on.