Hello,

Welcome to Sunday CET.

Sleepy hot continent mostly holidaying - while dealing with the usual summer heat wave, which this year is estimated to cost the European economy €180b, or about 1% of EU GDP and roughly equal to the bloc’s expected growth for the year (around 1.1%).

180b is also roughly the cost of five 1 GW data center projects, such as the one Mistral announced this week to start building by 2030. We have a look at the nitty gritty below - scroll for it and some more.

Enjoy,
Dragos

You read news about VC deals every day - but there’s a significant difference between news and intel:

One sells emotion and engagement and the other what it means for you and what to do about it.

We have built a comprehensive data platform tracking VC intel from Europe, a solid source of professional investment context from 30+ countries, backed by ecosystem intel - high level context curated every week, ideal for busy insiders who can’t keep up with the media noise.

Five things you get with a Watchlist that you won’t find anywhere else:

This week, Mistral announced plans to build up to 1 GW of capacity by 2030. It’s an interesting move coming from an European company that never found foothold on the global market and not entirely unexpected.

There’s at least three reasons to do it - i) exploding sovereign market growth ii) extreme power-grid constraints turning electricity and space into valuable assets and iii) the EU AI Act.

Some numbers first.

Also for context, the European market is an American oligopoly, with AWS, Azure, Google controlling roughly 70%. They operate via two main routes - self-build w/ large campuses they own and operate exclusively (these are not part of multi-tenant co-location availability) + leased co-locations.

That translates into a large fraction of the physical capacity sitting in Europe being controlled or heavily utilized by US companies, even if it primarily serves European customers and regulatory needs.

Herein comes Mistral. First off, it’s no coincidence that the French went public with it just as EU AI Act went into enforcement this month:

Mistral is simply riding the demand side the EU AI act enacted - and plans to do so in three phases - rollout of a 44 MW facility near Paris (2026), make available 200 MW of capacity across France and Sweden (2027) and extend said availability to 1 GW (2030).

Interesting to note that this also means Mistral is pivoting into becoming a neocloud, moving away from being a pure software developer, leaning more toward inference and controlled regional deployment than pure frontier training. Inference has more predictable, recurring demand and is easier to sell under long-term contracts. Training is spikier and more capital-intensive.

To a certain extent, this goes towards the same vertically integrated, infrastructure-first playbook that Elon Musk has deployed across SpaceX and xAI. However, the French are hardly taking the risks Elon is taking.

Taking long-term offtake agreements is a pre-requisite smart financing play for the French. Building 1 GW of AI capacity costs $38b, and requires a bit of financial engineering for a company concurrently in talks to raise $3b in equity at a $20b valuation today - the infra push is likely to be priced in even though technically it shouldn’t be in the operating company, rather in SPVs. Fwiw, they already have at least $1b in debt on the balance sheet (140m in 2024 and 830m in 2026) and make a modest half a billion in sales.

Also interesting to note that energy access is as valuable as financial leverage, and France’s nuclear fleet is an advantage here. Location decisions (Paris region, Nordics, etc.) are driven as much by access to firm, relatively cheap, low-carbon electricity as by regulation or politics.

However, all those things - raising capital and signing offtake agreements - are the easy part. Actually delivering GW-scale AI clusters on time - power, cooling, networking, operations, talent - is extremely hard, particularly for a software development crew. A totally different cup of tea, with very few organisations in Europe having done this at scale - the delivery risk makes for the biggest threat to everything the French plan for the next 3-4 years.

As a closing note, being sovereign of sorts has been Mistral’s only differentiating angle ever since it started building an LLM model, and this is just doubling down.

To make a joke, Mistral is turning from pure software house into a compliance wrapper driven by Macron’s powerful rolodex. It’s also kind of the blueprint for how Europeans have always built generational wealth.

Speaking of the EU AI Act that took effect earlier this month - one sillier side effect of it is the push to mark AI-generated content as such. Providers of covered gen AI systems must make AI-generated or manipulated content detectable by applying machine-readable markings - plus there’s a separate duty for anyone publishing AI-written text on public-interest matters to disclose it.

It’s wrong.

The EU is regulating provenance at the point of production when they should worry about distribution and amplification.

AI is a general-purpose production tech not a publishing medium or a social network.

It is a means within the creation process that can take an infinite number of shapes and forms, it’s not a content product that some politician needs to control.

Treating it as a sole media output is not only wrong, but also misleading. The final outcome can be a result of a mix and remix of all sorts of tools and usually only the last step of a longer chain that may have used AI for ideas, research, outlining, drafting, basic editing or proofreading.

If I dictate an article to the AI, have it cleaned up, then translate it in a different language and then substantially rewrite it myself, have the AI correct it and automatically publish - is this AI content? It is, by EU’s understanding of how AI works.

You get the idea - the line is arbitrary once you look at the actual workflow.

NB: Funny enough, the law itself partly acknowledges this distinction: AI-generated public-interest text does not need to be disclosed when it’s been reviewed by humans and a person or organisation assumes editorial responsibility. But that exception does not apply to the separate machine-readable marking obligation imposed on providers, even though in both cases a human is accountable for it.

And the problem is not that, at the end of the day, the EU law implicitly says that AI content is slop - it’s that it turns provenance into a reason for suspicion because the regulation is ostensibly about origin, while people actually care about its quality. It’s a free market, let them decide what’s good for them.

Is there any difference between AI slop and human slop? Just visit Linkedin, it’s the premier playground for AI/human slop competition.

This is where the regulation should work - who hosts the would-be manipulative content, AI-generated or not, should be responsible, not the slop creators, or the tool that helped produce a paragraph. Platforms already have the distribution and moderation power - and the DSA law already has some formal oversight for the large platforms.

But they cannot fully do that because platforms are American and Chinese. And listen, I get it, the EU politicians are nervous about the risks that AI content will lead to manipulation and the deterioration of the information ecosystem - a knee-jerk reaction as the Russians have been using social media massively to impact elections in a successful way in Europe. Note that an AI watermark doesn’t prevent manipulation.

Enforced as law will stigmatize legitimate AI-assisted work and lead to overheads, compliance costs and the emergence of a circumvention-based economy. The result is doing massive collateral damage in a market that’s already hurting, and becomes an extra to an already big tech handicap Europe has versus US and China.

And btw, absence of a mark does not prove something is human-written - a bad rule is worse than no rule.

AI tool sprawl happened fast, and visibility never caught up. Employees paste contracts into ChatGPT, run code through Copilot, and build workflows in apps security never approved.

Harmonic Security classifies every AI interaction by task, tool, and team, so you can see what’s actually happening across approved and unapproved apps alike.

Explore Your AI Usage

We have screened 147+ fundraising deals closed in Europe in the past seven days. We archive/transform deal-related data into an easy-searchable intelligent asset at N9, and email a selection of the interesting ones to our customers every week.

We grade this selection of the ones that matter by the capital event quality, proof vs. price, and investor signal - grades range from C to A+.

This year we graded 830 transactions made by investors in Europe. Below the investors with top grades:

Top investors by A+ grades (24 A+ deals total)

🇺🇸 Nvidia - 7
🇺🇸 General Catalyst - 5
🇸🇬 Temasek - 5
🇺🇸 Lightspeed Venture Partners - 5
🇫🇮 NGP Capital - 4
🇬🇧 British Business Bank - 4

Top investors by A grades (239 A deals total)

🇺🇸 General Catalyst - 20
🇫🇷 Thomas Wolf - 14
🇬🇧 20VC - 14
🇬🇧 Index Ventures - 14
🇬🇧 Plural - 13
🇸🇪 Creandum - 12

Quick facts

The DB with the graded deals here - by country, sector or investor. We send every week our graded transactions to our customers, you can join them too.

Great, will forward • Good • Meh

Thanks for reading! Please send me feedback by hitting reply.

To support my work, upgrade to one of the subscription options.

If this email was forwarded to you, please subscribe, it’s free!

Created every Sunday by @drnovac of Nordic 9 with weekly notes and observations from the European startup ecosystem.

You have received this email as you signed up at Sunday CET or are a Nordic 9 registered user.