Mistral AI raises €3B led by Samsung: how sovereign is it?

Mistral AI 融资 30 亿欧元，三星领投：它到底有多“主权”？

Mistral AI announced a €3 billion Series D on September 8, 2026, at a post-money valuation of more than €21 billion, to build “sovereign, open-weight AI” for Europe. The round was led by Samsung Electronics, which is headquartered in Suwon, South Korea. If you pick an LLM provider for data-residency reasons, or you are deciding whether open weights from a European lab are worth their price, this round changes what Mistral can build and who it answers to.

2026 年 9 月 8 日，Mistral AI 宣布完成 30 亿欧元的 D 轮融资，投后估值超过 210 亿欧元，旨在为欧洲构建“主权级、开放权重的人工智能”。本轮融资由总部位于韩国水原的三星电子领投。如果你是因为数据驻留（data-residency）需求而选择大模型供应商，或者正在权衡欧洲实验室的开放权重模型是否物有所值，那么这轮融资将改变 Mistral 的构建能力及其背后的利益相关方。

TL;DR €3B Series D, post-money above €21B. Mistral calls it “the largest equity fundraising round ever completed by a European technology company” (Mistral). Lead: Samsung Electronics. Co-leads EQT’s Scaleup Europe Fund and PSG Equity; returning investors include a16z, ASML, Nvidia and Salesforce Ventures. CEO Arthur Mensch told CNBC Mistral is on track to pass $1 billion in annual recurring revenue before year end.

简而言之：D 轮融资 30 亿欧元，投后估值超 210 亿欧元。Mistral 称这是“欧洲科技公司有史以来完成的最大规模股权融资”。领投方为三星电子。联合领投方包括 EQT 的 Scaleup Europe Fund 和 PSG Equity；老股东包括 a16z、ASML、英伟达和 Salesforce Ventures。首席执行官 Arthur Mensch 向 CNBC 表示，Mistral 有望在年底前实现超过 10 亿美元的年度经常性收入（ARR）。

Mistral Medium 3.5 (128B, open weights) costs $1.50 in and $7.50 out per million tokens; Artificial Analysis puts the median for comparable open models at $0.14 and $0.40. The same morning, CipherCue measured that 89.6 % of European companies with a CDN sit behind Cloudflare. Sovereignty has a long way to go below the model layer.

Mistral Medium 3.5（1280 亿参数，开放权重）的定价为每百万 token 输入 1.50 美元，输出 7.50 美元；而 Artificial Analysis 的数据显示，同类开放模型的输入输出中位数仅为 0.14 美元和 0.40 美元。就在同一天上午，CipherCue 的监测显示，89.6% 使用 CDN 的欧洲公司都依赖 Cloudflare。在模型层之下，实现“主权”还有很长的路要走。

Who invested in Mistral AI’s €3B round? The cap table is the story. From Mistral’s announcement: Role: Investors Lead: Samsung Electronics Co-leads: Scaleup Europe Fund (EQT), PSG Equity New: Advent, BlackRock (funds and accounts), the Grand Duchy of Luxembourg Returning (selection): a16z, ASML, BNP Paribas CIB, Bpifrance, General Catalyst, Index Ventures, Lightspeed, Nvidia, Salesforce Ventures

谁投资了 Mistral AI 的这 30 亿欧元？股权结构表说明了一切。根据 Mistral 的公告：角色：投资方领投：三星电子联合领投：Scaleup Europe Fund (EQT), PSG Equity 新进：Advent, 贝莱德（基金及账户）, 卢森堡大公国老股东（部分）：a16z, ASML, 法国巴黎银行企业与机构银行部 (BNP Paribas CIB), 法国国家投资银行 (Bpifrance), General Catalyst, Index Ventures, Lightspeed, 英伟达, Salesforce Ventures

There is European money in there: ASML, Bpifrance, BNP Paribas, EQT, and a sovereign state, Luxembourg. There is also Korea, Sand Hill Road, BlackRock and the company that sells Mistral its GPUs. That is not a scandal; a €3B round needs large cheques and there are not many European funds that write them. It does mean “sovereign” describes where the models run and who controls the weights, not who owns the company.

其中确实有欧洲资本：ASML、Bpifrance、法国巴黎银行、EQT 以及主权国家卢森堡。但同时也有韩国资本、硅谷风投（Sand Hill Road）、贝莱德以及向 Mistral 出售 GPU 的公司。这并非丑闻；30 亿欧元的融资需要大额支票，而能开出这种支票的欧洲基金并不多。这确实意味着，“主权”一词描述的是模型运行的位置以及谁控制权重，而不是谁拥有这家公司。

The valuation roughly doubles every twelve months. Total raised is above $6 billion. What does €21 billion buy? Mistral says it works in 20 countries with more than 125 enterprise customers, including Airbus, ASML and HSBC. Mensch told CNBC that revenue will pass $1 billion ARR this year (Quartz via Yahoo Finance). On Hacker News one commenter put Mistral’s revenue at 700 million and wrote that it “is what Anthropic generates in 3 days”. That is the commenter’s figure, not a disclosed one, but the scale gap it points at is real.

其估值大约每 12 个月翻一番。总融资额已超过 60 亿美元。210 亿欧元的估值意味着什么？Mistral 表示，其业务覆盖 20 个国家，拥有超过 125 家企业客户，包括空客、ASML 和汇丰银行。Mensch 向 CNBC 表示，今年营收将突破 10 亿美元 ARR（据 Quartz 转引 Yahoo Finance）。在 Hacker News 上，一位评论者称 Mistral 的营收为 7 亿，并写道“这只是 Anthropic 3 天的收入”。这是评论者的估算而非官方披露，但它所指出的规模差距是真实存在的。

What is sovereign AI, and what Mistral is spending on? “Sovereign AI” in Mistral’s pitch means 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. Mistral describes itself as “the only AI company in the world building the full stack”.

什么是主权 AI，Mistral 把钱花在了哪里？Mistral 宣传中的“主权 AI”是指一家欧洲公司，它自主训练模型、发布开放权重，并在其位于欧洲的自有算力上运行推理，从而使银行或政府部门能够根据欧洲法律保护数据。Mistral 自称是“世界上唯一一家构建全栈技术的 AI 公司”。

The money goes into that stack, per the Yahoo Finance report: an $830 million debt raise earlier in 2026 for a data center at Bruyères-le-Châtel, outside Paris, with 13,800 Nvidia GB300 GPUs; a second site in Sweden, about €1.2 billion; a target of 200 MW of European capacity by the end of 2027, “toward a gigawatt by the end of the decade”. Mensch also said “the amount of compute that we own is going to grow around 100% in the next five years”.

根据 Yahoo Finance 的报道，资金投入到了这一技术栈中：2026 年初筹集的 8.3 亿美元债务用于巴黎郊外 Bruyères-le-Châtel 的数据中心，配备了 13,800 张英伟达 GB300 GPU；瑞典的第二个站点耗资约 12 亿欧元；目标是到 2027 年底在欧洲拥有 200 兆瓦的算力容量，“并在本十年末达到 1 吉瓦”。Mensch 还表示，“我们拥有的算力总量在未来五年内将增长约 100%”。

Two details complicate the word “sovereign”. Microsoft will use capacity from Mistral’s European data centers, and Mistral models ship in Microsoft Foundry and Copilot Studio. The independent European cloud has an American tenant from day one. And on August 24, Mistral announced a strategic collaboration with HUMAIN, “in the hundreds of millions of Euros”, for sovereign AI in Saudi Arabia, where “Mistral will explore using HUMAIN’s data center infrastructure”. Sovereign, in practice, is a product: your data stays under your government’s control, on hardware in your jurisdiction, whichever that is.

有两个细节让“主权”一词变得复杂。微软将使用 Mistral 欧洲数据中心的算力，且 Mistral 模型已集成在 Microsoft Foundry 和 Copilot Studio 中。这个“独立的欧洲云”从第一天起就有了美国租户。此外，8 月 24 日，Mistral 宣布与 HUMAIN 达成“数亿欧元”的战略合作，旨在沙特阿拉伯实现主权 AI，其中“Mistral 将探索使用 HUMAIN 的数据中心基础设施”。实际上，“主权”是一种产品：即你的数据保留在政府的控制之下，并存储在你所在司法管辖区的硬件上，无论那个司法管辖区是哪里。

Mistral Medium 3.5: price and benchmarks. For developers, the product that matters is the model. Artificial Analysis lists Mistral Medium 3.5 as a 128-billion-parameter dense model with open weights under a Modified MIT license, a 256k context window and about 149 output tokens per second. It scores 15 on their Intelligence Index, well above the median of 8 for comparable models. Their summary also says it is “particularly expensive when comparing to other open weight models of similar size”.

Mistral Medium 3.5：价格与基准测试。对于开发者而言，最重要的产品是模型本身。Artificial Analysis 将 Mistral Medium 3.5 列为拥有 1280 亿参数的稠密模型，采用修改版 MIT 许可证，支持 256k 上下文窗口，输出速度约为每秒 149 个 token。它在他们的智能指数（Intelligence Index）中得分为 15，远高于同类模型 8 分的中位数。他们的总结还指出，该模型“与同等规模的其他开放权重模型相比，价格特别昂贵”。

What that means for a bill, as illustrative arithmetic (list prices, no caching): Workload: 10M input + 2M output tokens Input | Output | Total Mistral Medium 3.5 ($1.50 / $7.50) | $15.00 | $15.00 | $30.00 Median comparable open model ($0.14 / $0.40) | $1.40 | $0.80 | $2.20

这对账单意味着什么（以示例计算，标价，无缓存）：工作负载：1000 万输入 + 200 万输出 token 输入 | 输出 | 总计 Mistral Medium 3.5 ($1.50 / $7.50) | $15.00 | $15.00 | $30.00 同类开放模型中位数 ($0.14 / $0.40) | $1.40 | $0.80 | $2.20

Running the whole Intelligence Index on Medium 3.5 cost Artificial Analysis $1,159.93. Dev Twitter was less polite. Theo wrote: “It’s also been 3 years since they had a model worth looking at for any reason at all”. Youssof Al Toukhi quote-tweeted a chart he says is Terminal-Bench, with Medium 3.5 at 0 %, behind a 27B Qwen: The defence on Hacker News came from davedx: “People dump on them because they’re not benchmaxxxing which is pretty shortsighted”. Mistral’s customers are banks and governments that cannot legally send data to a US region, and for them a European model is a compliance answer, not a leaderboard entry. Another commenter, nik736, reported the opposite from their own workloads: “In our business benchmarks their Mistral Medium 3.5 with reasoning is worse than Gemma 4 31B and Glimmer 30B.” What I’d take from this as a developer: if data residency is your hard requirement, Mistral is one of the few serious options and the o…

在 Medium 3.5 上运行整个智能指数测试花费了 Artificial Analysis 1,159.93 美元。开发者推特上的评价则没那么客气。Theo 写道：“距离他们发布任何值得一看的模型已经过去 3 年了”。Youssof Al Toukhi 转发了一张他称之为 Terminal-Bench 的图表，显示 Medium 3.5 的得分为 0%，落后于 270 亿参数的 Qwen。Hacker News 上的辩护者 davedx 则表示：“人们抨击他们是因为他们没有在基准测试上‘刷分’，这太短视了”。Mistral 的客户是那些在法律上无法将数据发送到美国地区的银行和政府，对他们而言，欧洲模型是合规的解决方案，而不是排行榜上的名次。另一位评论者 nik736 则根据自己的工作负载给出了相反的反馈：“在我们的业务基准测试中，他们的 Mistral Medium 3.5（带推理功能）表现比 Gemma 4 31B 和 Glimmer 30B 还要差。”作为开发者，我的结论是：如果数据驻留是你的硬性要求，Mistral 是少数几个严肃的选择之一，而……