Samsung and ASML Both Bet on the Same AI Startup

Samsung Electronics Led Europe’s Largest AI Round

French AI company Mistral announced on September 8 that it had successfully raised €3 billion in a new Series D funding round. At an exchange rate of €1 = ₩1,560, that comes out to roughly ₩4.7 trillion (~$3.5B), and the resulting post-money valuation now tops €21 billion (about ₩32.8 trillion). It’s the largest single equity investment that any European technology company has ever received — and remarkably, it came just three years after the company was first founded.

Samsung Electronics led the round. EQT’s Scale-Up Europe Fund and existing investor PSG Equity joined in as co-leads, while Advent, a fund managed by BlackRock, and the Grand Duchy of Luxembourg all came in as brand-new investors. Meanwhile, existing shareholders — Nvidia, ASML, a16z, BNP Paribas, and France’s state investment bank Bpifrance — all chose to stay on board.

Let me clear up one number first, since it matters. The €3 billion figure is the total size of the entire round; the announcement itself doesn’t specify how much of that came from Samsung Electronics in particular. The Financial Times had reported back in July that a figure somewhere around €1 billion was under discussion, but the confirmed, final amount still remains undisclosed to this day. Some Korean headlines read as though Samsung Electronics had directly invested a sum in the ₩4-trillion range; what actually happened, more precisely, is that Samsung led a round of roughly that size, not that it personally wrote a check that large.

What really catches my attention here isn’t the amount, though — it’s the sequence. The company that led Mistral’s previous funding round was ASML, which makes semiconductor lithography equipment. The company leading this new round is Samsung Electronics, which makes memory chips instead. Two manufacturers in a row have now led back-to-back funding rounds for Europe’s largest and most closely watched AI company.

The reason both companies chose to put their money into the very same place isn’t really about model performance at all — it’s about deployment terms. For manufacturers that simply cannot let sensitive process data leave the building, sovereign AI¹ isn’t just a marketing slogan; it’s a hard procurement requirement. That said, Samsung’s own calculation still includes one extra line item that simply wasn’t present in ASML’s.

---

Factories Are Piling Up on the Cap Table

Here’s Mistral’s fundraising history laid out in order: a €600 million Series B back in June 2024, at a €5.8 billion valuation; a €1.7 billion Series C in September 2025, at an €11.7 billion valuation; and now this new €3 billion Series D, valued at €21 billion. In under twelve months, the company’s overall valuation grew by roughly 1.8 times.

During the Series C, ASML put in €1.3 billion and, as a result, became the largest shareholder with roughly 11% ownership on a fully diluted basis. Its CFO also joined Mistral’s strategy committee at the same time. Back in Issue 218, I read that particular deal as essentially a purchase of jurisdiction, not of raw performance. The data that ASML handles isn’t limited to its own equipment know-how — it’s mixed in together with process data belonging to TSMC, Samsung Electronics, and Intel as well. For a company sitting in that position, where a model physically runs ends up mattering more than how smart that model happens to be.

This isn’t Samsung Electronics’ first point of contact with Mistral, either. Samsung Venture Investment had already participated in the 2024 Series B, and Samsung SDS also took a stake that very same year. This past April, Mistral CEO Arthur Mensch visited Samsung’s Hwaseong campus in person and met with Jeon Young-hyun, head of Samsung’s Device Solutions (DS) Division, specifically to discuss supplying memory for AI data centers. That same month, both Chairman Lee Jae-yong (known internationally as Jay Y. Lee) and Mensch attended a luncheon together at Cheong Wa Dae (the Blue House, Korea’s former presidential office complex), during French President Emmanuel Macron’s official state visit to Korea. In other words, a relationship that had already been quietly building for more than two years has now finally graduated into a full lead investment.

MISTRAL AI  SERIES B → D

Lead Investors Were Manufacturers, Two Rounds Running

Post-money valuation and lead investor by round

Companies with factories led these rounds

€5.8 billion

€600M raised

€11.7 billion

€1.7B raised

€21 billion

€3B raised

Series B  June 2024

General Catalyst

Venture capital

Series C  September 2025

ASML

Semiconductor lithography equipment

Series D  September 2026

Samsung Electronics

Memory chips

The overall picture gets a lot clearer once you look at who Mistral is actually working with today. It’s currently handling AI transformation projects for more than 125 companies spread across 20 different countries, and the specific ones the company has named publicly so far are Airbus, ASML, and HSBC. On top of that, add a €100 million (~$108 million) deal with the French shipping company CMA CGM, and a separate partnership with BMW. Aircraft, semiconductor equipment, banking, shipping, automotive — these are all industries where the very moment sensitive data leaves the building, a real problem tends to follow.

Samsung’s Calculus Has One More Variable

ASML is simultaneously Mistral’s customer and its shareholder. In other words, it sits purely on the buying side. Samsung Electronics, by contrast, sits on both sides at once — buying and selling.

The first motive closely mirrors ASML’s own reasoning. Samsung Electronics also runs large fabs, and process data happens to be the company’s single most sensitive asset. Alongside this announcement, Mensch said that, building on ASML’s experience embedding AI directly into its manufacturing processes, Mistral would also pursue similar collaboration with Samsung Electronics on custom AI solutions for the shop floor and for enterprise use more broadly. Put simply: repeat with Samsung what was already done at ASML.

The second motive, however, runs in the exact opposite direction. Samsung Electronics sells memory for a living. Companies that train and serve AI models are, in effect, the buyers — consuming HBM², server-grade DRAM, and SSDs in enormous bulk quantities. This past March, Mistral borrowed $830 million from seven different banks in order to buy 13,800 Nvidia GB300 units and install them in a data center near Paris, and it also announced plans to build a further €1.2 billion data center in Sweden. Then in July, it signed a multibillion-dollar infrastructure deal with Microsoft, agreeing at the same time to bring in several thousand additional Nvidia Vera Rubin units. That’s precisely why, when the investment talks first surfaced back in July, both the FT and various Korean outlets read the whole deal as an AI-memory alliance in the making.

The very same structure had already appeared in the US roughly three months earlier. On June 22, Micron made a strategic investment in Anthropic’s Series H round while simultaneously signing separate HBM, DRAM, and SSD supply agreements with the company. Neither the investment amount nor the specific contract terms were disclosed by either side, and Micron’s stock nonetheless closed up 6% that very day, at an all-time high. It’s a recurring structure in which the memory company becomes a shareholder in the model company, and that same model company then turns around and buys the memory back.

Here’s a caution worth flagging clearly. When a supplier buys a stake in its own customer, the underlying character of the revenue coming from that customer inevitably changes. It’s essentially the same shape as the problem I covered back in Issue 220, when Stripe acquired one of its own merchants outright. That said, since Samsung Electronics hasn’t disclosed its investment amount, and no memory supply agreement with Mistral has been announced yet either, there’s simply no way to size this particular risk right now. Only two things are really worth watching closely: how Samsung Electronics ends up classifying this investment in its official filings, and whether a separate supply agreement with Mistral eventually emerges.

The valuation itself is also hard to take entirely at face value. Mensch and CFO Johan Berikvist said they expect annual recurring revenue³ to top $1 billion by the end of the year. That’s simply a company forecast, though, not a set of confirmed results. Using that projected figure as the denominator, the $24 billion valuation works out to roughly 24 times annual recurring revenue. That multiple is lower than recent round multiples seen for major US AI companies — but it’s still worth remembering that this number is being divided by a target that hasn’t actually been hit yet.

Sovereign AI isn’t about nationality — it’s five control points

Let’s clear up some terminology here, since it matters quite a bit. Mistral is often described as Europe’s flagship sovereign AI player, and yet its shareholder list includes a Dutch equipment maker, a Korean semiconductor company, an American asset manager, and the Luxembourg government, all sitting side by side. Nationality alone simply doesn’t explain that lineup.

The way Mistral itself defines sovereignty isn’t really about ownership at all — it’s about control. Its announcement lays out four core questions: Does data stay within the organization’s own boundary? Can you actually control and modify the model? Is the compute dedicated and predictable? And can you control and audit the system while it’s actively running?

Turned into questions you can genuinely use during a procurement review, they end up looking something like this.

Control point	Question to ask in review	What often gets missed
Data	Do inputs and logs leave our boundary? Are they used for training?	Inference logs and prompt caches often get left out of contracts
Model	Can we receive the weights and fine-tune them in our own environment?	Getting the weights means nothing without hardware to run them on
Compute	Is this a dedicated resource, or shared with others? Does it become a fully isolated environment?	The infrastructure provider and the model provider may fall under different jurisdictions
Operations	Can you reconstruct after the fact what decision was made and why?	Audit log retention periods sometimes run shorter than what regulations require
Persistence	Does this condition survive an acquisition or a policy change?	This is the line item ASML bought with its 11% stake, and Samsung Electronics with its lead investment

That last row is mine, added on top, not something Mistral itself proposed. The first four rows form a framework that Mistral built specifically to describe its own product, so using it directly as a scorecard, as-is, naturally tilts things in Mistral’s favor. Even so, it still works perfectly well as a practical checklist. Ask any tool you’re currently evaluating these same five questions verbatim, and you’ll likely surface items that never once showed up on the standard performance comparison chart.

Why the persistence row matters so much became especially clear this past June. On June 12, the U.S. Department of Commerce restricted overseas access to Anthropic’s latest model under export controls; that restriction was then lifted on June 30, and full access was restored on July 1. It only lasted about three weeks in total, but for anyone actually running procurement, it firmly confirmed that model access is, at bottom, a policy variable rather than a fixed technical spec. This is the very same issue covered back in Issue 146. It’s also exactly why Korean domestic coverage cited this particular incident as background when explaining Samsung’s investment discussions with Mistral back in July.

There’s one more distinction that’s genuinely worth making here. What Mistral actually sells is open weight⁴ access, not open source. You can receive the weights, sure, but the underlying training data and the full training process aren’t disclosed along with them. Some domestic Korean coverage uses these two terms interchangeably, but in an actual procurement review, that difference meaningfully changes the contract terms involved. And as covered back in Issue 178, receiving the weights and actually being able to run them are also two entirely different problems in practice.

Oswarld’s Lens

While writing Issue 218, I left one particular question unanswered: whose model is now absorbing the 20 years of process-level tacit knowledge accumulated inside our own fab? Barely ten days later, Samsung Electronics gave its answer. It didn’t choose to simply grow its own model even further — instead, it chose to become the lead investor in a company that sells deployment terms.

Reading this as Samsung quietly abandoning its own model would honestly be reading too much into it. Samsung SDS announced back in June that it would build a separate, purely domestic sovereign AI infrastructure of its own, and there’s also a separate, group-level large-scale investment plan sitting alongside it. Both tracks are moving forward in parallel right now, and this particular investment simply belongs to the “buy from outside” track.

What catches my attention even more, though, is where the other companies stand in all this. ASML holds an 11% stake; Samsung Electronics instead bought its own terms through a full lead investment. Very few companies anywhere in the world can actually do that. Everyone else has to go out and buy those same terms through an ordinary contract. That’s exactly why the five lines above become, in practice, a genuine procurement requirements list. It’s also, in a sense, an argument for reversing the usual sequence entirely — instead of building out the full performance comparison table first and only later having it overturned in legal review, check the legal terms first.

The old habit of asking “is it domestic?” first also quietly breaks down with this particular deal. A Korean company now sits as the lead name on the shareholder registry of Europe’s own flagship sovereign AI company. If you try to ask about sovereignty purely in terms of nationality, this picture simply doesn’t make sense. It only really makes sense once you ask where the data actually stays, whether you can receive the weights, and how long those specific conditions actually hold up.

Let me note the limits here as well, for the sake of honesty. What’s public so far is really just the round’s composition and the company’s declared strategic direction. Samsung Electronics’ actual investment amount, any memory supply agreement, and the real scope of manufacturing-AI cooperation haven’t been disclosed yet at all. Until those three specific things surface, today’s reading remains an inference built purely on the deal’s visible structure.

Closing

Here’s the summary, laid out plainly.

First, manufacturers have now back-to-back led Mistral’s two most recent funding rounds. ASML headed the Series C; Samsung Electronics headed the Series D. Companies that simply can’t let process data leave their own walls are buying deployment terms here, not raw performance.

Second, Samsung’s motive runs one full layer deeper than ASML’s did. Protecting fab data and selling HBM overlap directly in this case. It’s essentially the same structure as the deal Micron struck with Anthropic back in June, which means there’s a separate point where the true nature of the revenue still needs checking carefully.

Third, simply asking “which country does this belong to” doesn’t explain a sovereign AI deal like this one at all. The right questions instead concern data, models, compute, operations, and how durable the underlying terms really are.

Three things to watch going forward: whether Samsung Electronics’ investment amount eventually shows up in official disclosures, whether a memory supply agreement gets announced separately, and whether the manufacturing-AI partnership ever moves past pilot stage into an actual binding contract.

---

💬 Have you ever asked an AI tool you’re evaluating, “Can we get the weights and run this in our own environment?” Tell us in the comments what answer you got back.

📨 If someone on your team happens to handle AI tool procurement review, go ahead and pass this along. The five lines above can be used exactly as-is, as a ready-made requirements checklist.

References & Further Reading

Primary sources

• Mistral AI, “Making sovereign, open-weight AI the technology frontier”, 9/8/2026. ··· This is the primary source for the round’s structure and the four definitions of sovereignty. It’s worth reading with the awareness that this is a company defining its own product in its own terms.
• CNBC, “Mistral bags $24 billion valuation as Samsung leads funding for Europe’s AI champion”, 9/8/2026. ··· This article directly quotes Mensch’s remarks on revenue projections. Worth checking the part where it notes that no revised figures were disclosed.
• Tech.eu, “Mistral secures €3B Series D to push sovereign AI into its next phase”, 9/8/2026. ··· This has the most accurate breakdown of new and existing investors.
• Ajunews, “Samsung Electronics leads ₩4.3 trillion (~$3.1B) investment in Europe’s ‘OpenAI,’ Mistral AI”, 9/8/2026. ··· This is the Korean-language source for Mensch’s comments on manufacturing AI cooperation with Samsung. It’s also a good example of a headline that’s easy to misread as Samsung’s direct investment amount.
• ASML, “ASML, Mistral AI enter strategic partnership”, 9/9/2025. ··· The primary source for the €1.3 billion figure, the 11% stake, and the strategic committee seat.
• Micron, “Micron and Anthropic Announce Strategic Agreement”, 6/22/2026. ··· The original announcement bundling a supply agreement with an equity investment. This is the structural comparison point for today’s piece.

Background

• ZDNet Korea, “Will Samsung Electronics pour €1 billion into Mistral? AI memory alliance gains momentum”, 7/22/2026. ··· A Korean-language take from the time of the July Financial Times report. It lays out the April cooperation talks and Samsung Venture Investment’s track record.
• Seoul Economic Daily, “Samsung reviews investment in Mistral, Europe’s largest AI startup”, 7/22/2026. ··· The source for the observation linking export controls to sovereign AI demand.

Related past issues

• The Machine That Can’t Build From Blueprints Alone: ASML
• Fable Can’t Enter China, But GPT Can
• Stripe Bought Its Own Merchant

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Sovereign AI: An approach to building AI in which a company or nation directly controls the data, models, computing environment, and operations. It refers to where control resides, not the model’s nationality. ↩

2. HBM (High Bandwidth Memory): Memory made by stacking multiple layers of DRAM and connecting them through wide channels. It goes into Nvidia’s AI accelerators. ↩

3. Annual recurring revenue (ARR): A metric that converts subscription-like recurring revenue into an annualized figure. It differs from confirmed accounting revenue, so it’s especially important to distinguish it when presented as a projection. ↩

4. Open weights: An approach in which a trained model’s weight files are released publicly. This differs in scope from open source, which also releases training data and training code. ↩