Build1 publisher3 min readPublished
Mistral takes 3 billion euros from a Samsung-led round to build compute it owns in Europe
Mistral AI raised 3 billion euros in a Series D led by Samsung Electronics, at a post-money valuation above 21 billion euros. Buyers who choose it for data residency get a label that covers where models run, on an API priced at 11 to 19 times the open-model median.
The Engineer · Build desk
What happened
- EQT's Scaleup Europe Fund and PSG Equity co-led the round, and returning investors include a16z, ASML, Nvidia and Salesforce Ventures.
- Mistral's build-out includes a data center outside Paris with 13,800 Nvidia GB300 GPUs, funded by $830 million of debt, plus a Swedish site of about 1.2 billion euros.
- Microsoft will use capacity from Mistral's European data centers, and Mistral models ship in Microsoft Foundry and Copilot Studio.
- On August 24 Mistral announced a Saudi Arabia collaboration with HUMAIN worth hundreds of millions of euros, under which it will explore using HUMAIN's data centers.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Buyers who chose Mistral for residency have to write the serving site into the contract, because the vendor name alone no longer fixes the jurisdiction a request runs under.
- cost The API premium is charged per token, so it grows with volume; self-hosting the open weights swaps it for GPU capacity the buyer has to run and pay for itself.
- exposure Procurement rules that score vendor ownership will find a Korean lead investor, US funds and the company's GPU supplier on the cap table, wherever the hosting sits.
The dev.to write-up of the round summarises Mistral's definition of sovereign AI. It is a European company that trains its own models, publishes open weights and runs inference on compute it owns in Europe, so a bank or a ministry can keep data under European law [5]. The write-up concludes that the word describes where the models run and who controls the weights, and not who owns the company [6]. The cap table fits that reading. European holders include Bpifrance, BNP Paribas and the state of Luxembourg, next to BlackRock and a lead investor from Suwon, South Korea [4][2]. Nvidia turns up twice, once as a returning investor and once as the supplier of the GPUs in the new site outside Paris [3][7].
The round pays for the owned-compute part of that definition [7]. Arthur Mensch, the chief executive, told CNBC that Mistral is on track to pass $1 billion in annual recurring revenue before year end [17]. He also said "the amount of compute that we own is going to grow around 100% in the next five years" [8]. Owned compute settles where Mistral's own endpoints run. It does not settle where every copy of the model runs. Through the Microsoft arrangement, and the HUMAIN plan if it proceeds, the same model can reach a buyer from a Microsoft catalogue or from Saudi hardware [9][10].
Then there is the price. Medium 3.5 lists at $1.50 per million input tokens and $7.50 per million output tokens. Artificial Analysis puts the median for comparable open models at $0.14 and $0.40 [11]. The ratios are about 10.7 on input and 18.75 on output [1]. Artificial Analysis's summary says the model is "particularly expensive when comparing to other open weight models of similar size" [14]. It also scores Medium 3.5 at 15 on its Intelligence Index, where the comparable median is 8 [13]. That score is about 1.9 times the median [2]. Running the whole index on the model cost Artificial Analysis $1,159.93 [15].
Those figures are list prices without caching [16]. The premium a buyer actually pays depends on the mix of input and output tokens. Traffic made of long prompts and short answers pays close to the 10.7 input ratio. Generation-heavy traffic pays close to 18.75. The index score reflects Artificial Analysis's task mix. A buyer's extraction or retrieval jobs need their own evaluation before the 1.9 figure goes into a procurement memo.
Medium 3.5 is a 128-billion-parameter dense model with a 256k context window, about 149 output tokens per second, and open weights under a Modified MIT license [12]. The write-up does not describe what the modifications to the MIT terms restrict.
Residency also has layers below the model. On the morning of the announcement, CipherCue measured that 89.6% of European companies with a CDN sit behind Cloudflare [18].
What to watch
- Whether Mistral or Microsoft states which data centers serve Mistral models in Foundry and Copilot Studio, and whether that is Mistral's own European capacity.
- Whether the HUMAIN exploration becomes a signed deployment on Saudi hardware, putting the same weights under a third operator.
- A cut to Medium 3.5's $1.50 and $7.50 list prices toward the $0.14 and $0.40 medians for comparable open models.