This week on Upside: the AI infrastructure trade moved past GPUs and into memory, where Micron and SK Hynix both crossed a trillion dollars. Europe started swinging at US cloud providers while its own most important company can't get permission to build. And the "AI is taking the jobs" story took a beating from the actual data.
- Memory is the hottest thing on the planet, and for once the boom might not bust. SK Hynix and Micron crossed a trillion dollars because GPU units are tripling and memory per GPU is growing 80% a year, into an industry nobody dares add capacity to.
- The application layer keeps winning. Snowflake popped 37% on Coco, which runs on Anthropic's Opus but lives or dies on the harness around it. Salesforce got punished for the opposite reason.
- Europe blocked a US acquisition citing the CLOUD Act, stalled its own sovereignty package, and won't let ASML expand because of nitrogen rules tangled up with farmers.
- A million UK young people are now not in education, employment or training, a 12-year high. The data points at remote work and a 53% jump in entry-level hiring costs, not at AI.
- A zero-employee, nine-agent company called Pulsia raised $30m at a quarter-billion valuation. Spell it backwards.
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.
Memory becomes the trade
For two years the AI infrastructure conversation has basically been Nvidia and the GPU. This week it widened, and the most interesting corner is memory.
SK Hynix and Micron both crossed a trillion dollars. Let that sit, because memory has historically been one of the worst businesses in tech. Intel got out of it in 1985 when the Japanese crushed them on price, pivoted to CPUs, and rode that for 35 years. Memory was brutally cyclical. Boom, everyone piles in, bust, the new capacity kills them. The graveyard is long. Three survivors left standing: SK Hynix and Micron as the specialists, Samsung as the powerhouse.
Then GPUs happened and demand started compounding in two directions. Units shipped went from about 3.4 million in 2024 to eight or nine million this year. And as the models got bigger, the memory bolted onto each GPU grew roughly 80% a year. More units, each far hungrier, in an industry that takes years to add a single line.
Here’s the bit that breaks the pattern. Nobody is piling in. The capacity discipline that never held for 40 years is holding now, partly because by the time you stand up new capacity the next generation has arrived, and partly because the GPU makers don’t want their suppliers over-building and dying. China has been hit as hard as anyone by the high-bandwidth memory shortage. They can make GPUs, they can’t get the memory, and that gap doesn’t close before 2028. So margins stay fat, for a business that spent my whole career as a punchline.
The application layer, again
Same theme, software side. We’ve argued for years that applications are where the rubber hits the road, through the whole “models will eat everything” debate. This week gave us the cleanest split yet.
Snowflake popped 37% on Coco, short for Cortex Code. Think Claude Code but pointed at your data warehouse. It writes the SQL and queries your data in real time without hallucinating rows or columns, because the answer has to be true. Coco runs on Opus from Anthropic, but the value isn’t the raw model, it’s the harness Snowflake builds around it. Revenue past a billion within months of launch, and CFO Brian Robbins called it the single biggest driver of the raised forecast.
Salesforce got the opposite treatment. Down around 30% on the year despite a decent quarter, because the market is pricing a move from seat-based SaaS to usage. Agents don’t buy licences. Agentforce is up 200% year on year, which sounds great until you see it’s about a billion of run rate in a company that size. The model shift scares people.
Mads’ framing is the one I keep coming back to. The best model from Anthropic or OpenAI is powerful, but it needs a harness to tame it, and that harness is an application. Which is why the smart firms build in-house. Kirkland and Ellis, the world’s highest-grossing law firm, just put half a billion aside to build their own AI stack. Banks worked this out decades ago. The lawyers are next.
Europe picks a fight, then trips over its own feet
Three sovereignty stories landed in the same week, and together they tell you where Europe is stuck.
One. The Dutch government blocked US IT giant Kindle from buying Solvinity, a Dutch cloud provider that hosts the national online identity platform. First time a European company has cited the US CLOUD Act, the law forcing US firms to hand data to US law enforcement even when it’s stored abroad, as grounds to kill a deal.
Two. The EU’s new tech sovereignty package, which would fence US hyperscalers out of the most sensitive government cloud and AI work, got delayed from the 27th to around the 3rd. Maybe it lands, maybe it goes into the long grass. The fear hanging over it is retaliation. Start banning US suppliers and you invite a ban on European ones, and we have a tech sector to lose.
Three, and this one should sting. ASML, the most important company on the continent, can’t physically expand its facilities around Eindhoven. Not because of money, because of nitrogen rules tangled up with protected areas. It’s knotted to the farmers, because one of the few ways the Netherlands can show a nitrogen reduction is to cut farming. So the country is being asked to choose between its farmers and letting its single best company build. Christophe Fouquet is furious, and fairly so. CEOs from ASML, Airbus, Ericsson, Nokia, SAP, Siemens and Mistral co-signed an op-ed this month: three years after ChatGPT, Europe is still debating regulation while everyone else scales.
Here’s my read. We’re holding the wrong end of the stick. Instead of banning US suppliers and praying we don’t get hit back, point the energy at deregulating so the ASMLs of the world can grow. Let Fouquet build. You don’t win this race by tying your best people up in nitrogen paperwork.
The zero-employee company has arrived
This was inevitable. Pulsia raised $30m at a quarter-billion valuation. Zero employees, nine agents, a Trustpilot score of 2.1. Spell the name backwards if you want the review. They’re talking $10m of ARR, but break it down and it’s under $5m of real subscriptions plus one-offs and ad spend. Apparently even the fundraising was run by AI, so somewhere a bot pitched a bot and got a term sheet. We didn’t do the diligence, but from a distance it smells like 2021.
AI is getting blamed for things AI didn’t do
Imagine hearing your CEO call you “lower value human capital.” That’s roughly where Bill Winters at Standard Chartered landed at a conference this week, with 82,000 staff mostly in back-office roles and a plan to cut.
Set against that, the UK just hit a million NEETs, young people not in education, employment or training. A 12-year high, and 61% have never had a paid job. So, AI? Convenient, but no.
Banks have automated forever. Lloyds cut 9,000 in 2014 off a billion-pound digital investment. JPMorgan’s COIN saved 360,000 lawyer-hours in 2017. This is vanilla tech plus the performance management companies do every few years. And the AI-washing is measurable: 59% of hiring managers admit they frame cuts as AI-driven because it plays well with stakeholders.
The better explanation came from a John Burns-Murdoch piece in the FT. The drop in junior hiring correlates far more strongly with remote work than with AI. Juniors are brilliant but they don’t know how the business runs, and that training used to happen in the office. Add a 53% rise in the cost of an entry-level hire over six years, and businesses are quietly deciding it isn’t worth it. AI is a compounder, not the cause.
Quick hits
Opus 4.8. Anthropic shipped it the day before we recorded and the vibe checks are great. Same week, it eclipsed OpenAI as the most valuable frontier lab at around $965bn, with talk of a $50bn run rate and a first profitable quarter. Mads called it the most important company in the world.
The Pope. On the 25th he published a 42,000-word AI encyclical, 135 years to the day after Rerum Novarum answered the Industrial Revolution. Anthropic’s Chris Olah stood next to him, not OpenAI, not Google. Six principles, one with teeth: no autonomous weapons. Which is also the one I’d bet is dead on arrival. If the other side builds them and you don’t, the theology loses to the security argument.
Predictions
I’ll go first, since Mads doesn’t love them. Somewhere in this cohort of 15 to 20 AI-native unicorns, one blows up publicly by the end of next year. A full 72-hour catastrophic collapse, and I’d put my money on the foundation-model wrapper category, where the labs either already ship the product or are about to. A dot-com moment, remembered.
Mads, dragged into it, offered two: the West will build autonomous weapons because he can’t see how we don’t, and memory stays a hot, structural trade through next year.
Deal of the week
Mads picked Cognition, the team behind Devin, which raised $1bn at a $26bn post led by Lux Capital and General Catalyst. Why does Devin exist when you’ve got Claude Code and Codex? Enterprise. Inside a regulated bank you need audit trails, permissions, GovCloud and certification, the long regulatory tail Anthropic has said it won’t build. So Devin is used by Goldman Sachs, Mercedes, the US Army, the US Navy, the Treasury and NASA’s Jet Propulsion Lab. That niche isn’t going away, it’s getting bigger.
Mine was EDGY, a Dutch energy-tech startup out of Utrecht. Cash-flow positive, over 500MW under management as the largest independent energy optimiser in the Netherlands, a $7.5m raise after roughly 900% growth. I love an energy startup. If you’re early-stage and doing clever things with electrons, come find me.
Notable Quotes
“It’s super interesting how memory has suddenly become the hottest thing on the planet.” – Mads Jensen
“In the Netherlands, you’re basically forced to choose between the farmers and letting Europe’s most important company expand and build facilities.” – Mads Jensen, on ASML
“We can’t find evidence yet that AI is displacing masses of workers. The data just doesn’t stack up.” – Mads Jensen
Frequently Asked Questions
Why are memory chip makers like Micron and SK Hynix suddenly worth so much?
AI demand is compounding in two directions. GPU shipments roughly tripled from 3.4 million units in 2024 to eight or nine million this year, and the memory fitted to each GPU is growing around 80% a year. Because new capacity takes years to build and nobody is over-investing, supply stays tight and margins stay high, which pushed both Micron and SK Hynix past a trillion dollars.
What is Snowflake Coco and why did the stock jump?
Coco (Cortex Code) is an agentic data assistant inside Snowflake that queries your data warehouse in real time and writes the SQL without hallucinating rows or columns. It runs on Anthropic's Opus, but the value sits in Snowflake's surrounding harness. It passed a billion dollars of revenue within months of launch, and the CFO called it the largest driver of the raised forecast, which is why the stock popped about 37%.
Why did the Dutch government block the Kindle and Solvinity deal?
The Netherlands blocked US IT giant Kindle from acquiring Solvinity, a Dutch cloud provider that hosts the national online identity platform. It was the first time a European company cited the US CLOUD Act, which compels US firms to hand data to US law enforcement even when stored abroad, as grounds to block an acquisition.
Why can't ASML expand its factories in the Netherlands?
ASML is constrained by Dutch nitrogen emission rules around protected areas, not by money. The issue is tied to farming, because reducing agriculture is one of the main levers the Netherlands has to cut nitrogen, so the country is effectively choosing between its farmers and letting its most important company expand.
Is AI causing the rise in UK youth unemployment?
The evidence points elsewhere. The UK hit a million NEETs, a 12-year high, but a John Burns-Murdoch analysis in the FT found the drop in junior hiring correlates more strongly with remote work, which removes the on-the-job training juniors need. The cost of an entry-level hire has also risen 53% in six years. On the podcast, Mads argued AI is a compounder rather than the cause.