These are just signs of this market maturing. For a lot of business customers, EU based hosting is not optional. That includes models. Routing requests via API endpoints in the US is not really acceptable. And anything involving privacy sensitive data of course needs to be handled properly. Sam Altman pinky swearing to be nice doesn't quite cut it in terms of hard guarantees.
EU based legal entities and strong compliance with local laws with some hard SLAs and contractual guarantees is not going to be optional for liability reasons. Provenance of models, their training data, and exact ways they have been instructed to act are also not just nice to haves.
I expect non EU jurisdictions are eventually going to be similarly picky about their AI suppliers and I expect all the big tech providers to adapt to local markets just like they did with cloud infrastructure.
I don't have much experience with Mistral yet. But I may need to get my hands dirty to be able to sell this to some of our customers. We have a few more picky customers in Germany.
Because companies here actively want to avoid breaking the law, as opposed to the U.S where breaking the law is just a matter of paying some $ to the grifter in chief? I always find it funny when Europeans being proactive about that sort of stuff is somehow a bad thing from Americans point of view. Like wanting decent human rights and not having to bend over to megacorps is something we should not have.
Though, if the Americans in question just want to do their grifting in EU, it makes sense why they are upset at that, I guess, because it limits their grifting opportunities.
> companies here actively want to avoid breaking the law
This is hilarious. This reminds me of Soviet propaganda. "No, there was no Chernobyl disaster. Please disregard the corpses. Yes, the centrally-planned economy is doing fantastically, better than expected. Reports of famines and shortages are imperialist propaganda."
(Mind you, the Soviets are not alone here, but the blatant chutzpah of Soviet propaganda is perhaps more conspicuous to the Western eye than the Western varieties of PR and psychological manipulation.)
Thank you for your brilliant demonstration of survivorship bias.
How many people were punished for Enron? For the subprime crisis? Etc.
In the US, you just give a little money for the president's ballroom and you are pardoned. Or you settle out of court because your justice system is crap.
Yes, European companies break the law too. However, the comment this was about literally mocked the companies that are actively trying to follow the law.
So yes, such companies exist and plenty of people see their existence as a good thing rather then something to mock.
It also ultimately a expression of might makes right (sad as this is) and as the current culture supports a decline of western might, it also undoes the law - first international, than domestic. We simply decided to burden our might with these restriction fictions, others feel not at all compelled to follow.
I expect to see further selling out of these laws, as the economic prosperity declines. I can perfectly see german law limiting german companies from developing and selling AI products, while at the same time allowing us companies for a "pay our retires and pension-plans" kickback.
What goalposts? Your sitting president, himself a conman is pardoning fraudsters left and right while he and his family enrich themselves with public money and extortion.
Joke's on you, I'm European not American, so "your president" in this case would be the unelected Ursula and she would technically be a con-woman, not a con-man.
Not sure what your argument was supposed to prove with this cheap jab though.
Let me rephrase this: companies want to avoid breaking the law unknowingly, because their US providers are going to break the law without notice, willingly or unwillingly.
Plenty of corporations are willing to break the rules, but never for free.
> because their US providers are going to break the law without notice, willingly or unwillingly
This is a weird hill to die on because it's not true. I can't find anything to support your world view and if anything evidence points to the contrary. Europe has a deliberately more complex legal framework, usually in the hopes of keeping out foreign competition (although it's dubious whether or not that actually works).
> So Mistral is developing its own data centers, starting with one outside Paris. Mensch projects it will have 200 megawatts of capacity by the end of 2027. Power from France’s state-owned nuclear plants will help, but the buildout could still cost an estimated $5 billion. Mensch tapped oil-rich Abu Dhabi and reportedly sought debt financing to help pay for it.
Though to your point it won't be running until 2027.
Yes, I think the EU is going to be dependent on US tech (other than EUV lithography machines, very cool) for very long time. Even those data centres, while run by Europeans, are still being made with almost entirely US tech. But at least the EU companies can borrow some oil money and buy in the stuff developed by someone else's R&D spend, which is a nice shortcut to have available.
>I think the EU is going to be dependent on US tech (other than EUV lithography machines, very cool
Well, ASML's EUV light sources are based on licensed US IP from Sandia Labs, and manufactured in the US by CYMER, which ASML bought, but they still operate and manufacture out of California, so the EU is not sovereign/independent here (neither is any country).
This doesn't mean much anyway, since despite ASML being European, their machines all go to export and EU doesn't put any of those machine to good use domestically, with the most cutting edge semiconductor fabs on EU soil being the Germany based TSMC fabs on the much older 16 and 12nm nodes, far bigger than the 3nm that Taiwan and US operate domestically.
This is changing somehow. At least on a surface. For example Amazon have created European subsidiary completely managed by Europeans under European company thus under it's local jurisdiction.
1. the 2018 CLOUD Act mandates US companies — and their subsidiaries — to provide information to the US government on demand, regardless of where the data is stored
2. FISA secret courts prevent companies from even saying they where summoned, or telling anyone who or what the case was about (including canaries).
So you won't ever know if your data was handed over to the US government.
The whole point of setting up the EU subsidiary as a separate company that is incorporated in the EU and is managed and staffed by EU citizens is to avoid this.
The purpose of the CLOUD Act was to get at data that was stored outside the US but that was "in the custody, control, or possession of communications-service providers that are subject to the jurisdiction of the United States".
It arose from a situation where an email provider in the US used cloud storage services in several countries to store emails. They were asked for the email of a particular customer and said they did not have to provide it because they had happened to store that customer's mail at a non-US cloud provider.
What the CLOUD Act requires is that:
> A provider of electronic communication service or remote
computing service shall comply with the obligations of this chapter to
preserve, backup, or disclose the contents of a wire or electronic
communication and any record or other information pertaining to a
customer or subscriber within such provider's possession, custody, or
control, regardless of whether such communication, record, or other
information is located within or outside of the United States.
A company incorporated in the EU, even if it is owned by an entity in the US, is not subject to US jurisdiction and so that does not apply. The US owner is subject to US jurisdiction but the data of EU customers of the EU company is not in the US owner's possession, custody, or control.
> The US owner is subject to US jurisdiction but the data of EU customers of the EU company is not in the US owner's possession, custody, or control.
No? Certainly sounds like it is in the US owner's control to me.
But even disregarding that fact. Given that the US government also started hiding what it was doing with FISA courts and forbidding that anyone, including the companies themselves, checks what actually happens ... do you think anyone will believe this? We HAD evidence of US companies refusing to hand over data before CLOUD and FISA, we do not see that anymore. (And that's before we start taking into account more some recent administration's respect for ...)
Of course this is also pretty hypocritical since EU countries have been caught more than once capturing communications of non-citizens. The problem that usually gets mentioned: the Boeing - Airbus fight wasn't a one sided US being untrustworthy to help Boeing.
They should be legally and physically separated and these actions should be then potentially illegal for Europeans so I do not think I'm at least infactual.
But assuming the owner is US company abiding US laws it's safe to assume that data would be transferred to US one way or the another.
The US intelligence machinery spied on Angela Merkel's phone. Do you suppose secretly demanding cooperation for Lawful intercept capabilities in Amazon GmbH is somehow beyond or beneath them?
Also consider that all communication between the European subsidiaries to the HQ is fair game under FISA.
The European leaders would have have no say in it. If the software from Seattle is designed to covertly exfiltrate information, they won't even know it. Even if they review the individual code changes, it can be an obfuscated attack similar to XZ where the code itself is clean, but not so much for the network fabric firmware binary test data.
That's why I used the "somehow". But abiding your logic nothing is ever secured, which is ultimately true, but it could be illegal so detergent here is not the impossibility it self but potentional harsh punishment for breaking the law.
I am a Mistral Le Chat Pro subscriber. I specifically chose to test their offerings because they are European. I don't have the necessary local hardware to run really big models, therefore need to choose a cloud provider if I want LLM action.
I find the antics of Anthropic, OpenAI, Google, Microsoft distasteful and avoid their products where I can.
After testing Le Chat and Devstral-2 for a while, I felt their offering was good enough to stump up some cash for it. I appreciate that many of their models are open weights and Apache 2.0 licensed. In general, I've been happy enough with the service and quality.
Maybe others are better, but I have little reason to change right now. If curiosity gets the better of me, I'll be looking at Qwen, Kimi, GLM, Deepseek, other open weights models, before Anthropic and OpenAI.
Mistral models are definitely good enough. Most people fall for what I call the SOTA Logical Fallacy: whenever there is a 'better model', they think they need to use it, when less-powerful models actually perform the same tasks just as well. (it's an inverse form of the Shifting Baseline Syndrome: every time a new model comes out, people shift their baseline of what is acceptable, despite the fact that a previous baseline was acceptable for the same task)
Devstral Small 2 was (and remains) a particularly strong small coding model, even beating larger open weights. Mistral's "problem" is marketing; other providers ship model updates constantly so they remain in the news and seem like they're "beating" the competition. And it works: people get emotionally attached to brands and models, deciding who's better in the court of popular opinion, and that drives their choices (& dollars).
My biggest issue with Devstral and even their biggest model is that they’re dangerous unless closely directed and reviewed and i mean CLOSELY. Unfortunately mistral models will believe and do anything.
FWIW personally i prefer this. When i tried Qwen3.6 and asked it a few questions, while it did respond, it was ADAMANT i should do something else when i really wanted an answer to the question i made. It felt like when you search something and a stackoverflow answer about what you search for comes up and the most upvoted answer is about using/doing something else - when you want a specific answer to that specific question, not something else.
Meanwhile Devstral Small 2 just answers the damn question.
I don't want to have to convince my computer to do what i want it to do, i want from it to do what i ask it to.
> It felt like when you search something and a stackoverflow answer about what you search for comes up and the most upvoted answer is about using/doing something else - when you want a specific answer to that specific question, not something else.
Don't you think there's usually a good reason for this? Whenever this happened to me, the problem was my ignorance.
I think there is a reason why people do that: trying to steer -those they consider- newbies away from patterns they consider bad, but at the same time this second-guessing can be annoying when you know what you want to do (especially when the original question isn't actually answered yet it comes up in search engine results...).
I can't say if it is a good reason in general, perhaps it is, but it certainly is something i personally find annoying. I think answers should provide an answer to the question asked and then, after that answer was given, they could also give pointers for whatever they consider a better approach and why - this is important, IMO, for a public forum where people of all backgrounds and goals can read the same stuff.
But either way, LLMs IMO should do/provide what they are asked without trying to second guess the user (or at least, there should be LLMs that act like that).
FWIW i haven't used Claude or any other cloud-based LLM, only what i can run on my PC, so it could be that Claude is smart enough to follow the user's instructions, keep the equivalent of a mental state of what the user seems to want to do and only push back when it really makes sense whereas a small local LLM is too stupid to judge all that and Qwen3.6 errs on the side of being annoyingly cautious while Devstral Small 2 errs on the side of trusting the user being really okay with blowing their toes off :-P. As i wrote in my original reply, this is my personal preference and i prefer the LLM to just do what i ask.
TBH sometimes i feel like i'm "emotionally attached" to Mistral's models because i always end up using them :-P. However that is because, as you wrote, their small models (i only use local stuff) are very strong. In fact i was trying Qwen3.6 27B recently and while it is nice that it can do tool calls during the reasoning process (i had it confirm its thoughts by writing Python code) it often ended up confusing itself (regardless of tool calls) during reasoning, ending up in loops where it questions itself over and over endlessly.
Devstral Small 2 however just works, for the most part. Qwen3.6 27B can probably handle more complex tasks (when i asked it as a test to write a function that checks for collision between two AABBs in C and gave it a tool to call Python code for confirmation, it actually wrote a Python script that writes C code with the tests, then calls GCC to compile the C code and runs the binary to run the tests, which is something Mistral's small models couldn't do) but i always felt i can just leave DS2 doing stuff in the background (or when i'm doing something else) and it'll produce something relatively useful whereas the little time i spent with Qwen3.6 27B it felt more "unstable" (and much slower, both because of literally slower inference and because of endless reams of text).
Recently i also started using Ministral 3B and 14B - these can do some reasoning too and for very simple stuff Ministral 3B is very fast (i actually didn't expect a 3B model to be anything more than novelty) and have some vision abilities (though they're quite mediocre at vision so i haven't found much use for this - passing something via GLM-OCR to extract all text and feed it to another model feels more practical).
Also as i wrote in another comment, every Mistral model i've tried never questioned me, which i certainly prefer
For certains tasks that are not hard but depend a clear specification, it's even better to haver less capable model because it forces you to do a better description of what you want, ending up with a better results. I will defend my PhD thesis soon and I will buy a yearly Mistral subscription at a student price to get it for cheap.
> Most people fall for what I call the SOTA Logical Fallacy: whenever there ...
I think you'll find that ML now pretty much IS the HPC market, there's no distinction anymore. And the HPC market has always had the "being #1 gets you 99% of all business", even if #2 is only 10% behind SOTA.
Given what it's used for (ie. military applications, incl. nuclear weapons, but also rocket designs, flight planning, large-scale simulations), this is probably justifiable: part of it is states keeping in mind what the second prize in a war is worth ...
I use their API for several models, both for personal and professional use. I think their approach (smaller, specialised models that are well-adapted for specific tasks) is a very good fit for how I work. And even the more general-purpose ones, like the chat model, just... refreshingly good in a lot of ways. My "ruthless review" prompt, which I use for, well, ruthless, guided reviews of early technical drafts, has good technical results for early reviews and holy crap is it ruthless and does it know how to swear. By the time Claude or ChatGPT are done being honest about how right I am to push back and gently circling back, Mistral's large model has sent me back to the drawing board twice.
Being in the EU does smooth a lot of things in terms of compliance, payment processing and whatnot, but I also like that their data retention and privacy policies are pretty clearly spelled out. I need to know something, there's a good chance it's explained outright somewhere and I don't need to read between the EULA lines and wonder what it means.
I do hit limits in terms of capabilities sometimes, and I'm sure other providers' services offer better results for some things. But the businesses ran on top of those more capable models feel too much like a scam at this point and I'd rather not depend on them for anything I actually need.
That ruthless review prompt seems interesting, would you be willing to share it? I've been trying to have Claude act as a reviewer for me and it feels like it never will disagree.
It's very hard to untangle it from the rest of its context (the prompt is built dynamically, from a lot of parts, some project-specific, some specific to my preferences, some built from interaction history), so I can't really share it. In any case, I don't think it's some specific prompt engineering sorcery I'm doing, it's not like I've spent any real time refining it or experimenting with various magical incantations. It's probably just some model features making it more amenable to the kind of instructions that are relevant in these cases (directness, questioning trade-offs, thoroughness etc.). My chatbot swears equally graphical in review prompts and news summarizing prompts so I'm pretty sure I'm not tickling the machine just right :)
There is also risk from a US regulatory side as recent drama around antrophic showed.
Don’t think it’s inconceivable that the clowns in power decide to limit api access out of the blue one day because someone whispered a conspiracy theory in someone’s ear. API blockade…
See also the constant flip flopping on what cards NVIDIA can export - no consistency in stance or coherent policy
The thing with Anthropic and the military was about whether vendors can tell the military what operations it's permitted to do. It has no bearing on the commercial sector, and isn't actually about AI.
The thing with NVIDIA cards is a continuation of how we've restricted tech exports for quite a while. You can find old news articles about game consoles being export-restricted over nuclear proliferation concerns. This AI-related one was about whether or not custom AI models are relevant to national security, and whether restricting graphics card sales can have a meaningful impact on them.
Any issue with selling chat tokens internationally would be more akin to the recent tariff shenanigans.
Changing your LLM inference provider is the easiest switch in technology I can think of. It's quicker than taking off the case of your phone and putting on a new one.
Enough hardware and good models exist now that if you do get blocked from one place that viable alternatives do exist.
> Changing your LLM inference provider is the easiest switch in technology I can think of.
Thats true right up until you’re working with confidential info in a corporate context. Then it’s a multi month cross discipline cross jurisdiction project not an edit in a config file.
Try Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC_Q4_k_m.gguf. This 7.5GB model runs well in llama.cpp on my 2021 Macbook Pro and is good at both coding and business document analysis tasks.
In my opinion, being "Not American" or "Not Chinese" is not a good business model long-term.
At some point, businesses will choose the option that provides the most value. I'm very skeptical that Mistral will survive long-term.
Edit: I hear the commenters to this post. However, Mistral still relies on American chips. If there is truly a divorce between Europe and the US such that relying OpenAI or Anthropic is not an option, neither will relying on Nvidia and likely the thousands of smaller hardware and software suppliers that make Mistral work. That's why I don't think it's realistic to say that Europeans can't rely on OpenAI/Anthropic and that Mistral is free from American reliance. If you want true independence, you have to rebuild every single layer like what China is doing. That's hard and expensive.
> Edit: I hear the commenters to this post. However, Mistral still relies on American chips. If there is truly a divorce between Europe and the US such that relying OpenAI or Anthropic is not an option, neither will relying on Nvidia and likely the thousands of smaller hardware and software suppliers that make Mistral work. That's why I don't think it's realistic to say that Europeans can't rely on OpenAI/Anthropic and that Mistral is free from American reliance. If you want true independence, you have to rebuild every single layer like what China is doing. That's hard and expensive.
American designed. The GPUs are made in Taiwan, the RAM in South Korea, using machines from the Netherlands' ASML.
True independence is indeed hard and expensive. But it's also not the job of Mistral to tackle all the layers at the same time, not even the state-owned corporations of western Europe in the 20th century (and the EU isn't (yet) even a state) tried to tackle every stage of an industrial process by themselves.
Indeed. But that US government control is limited to adding restrictions and removing them again, or offering money, it can't magic things out of the air when other people or nations put restrictions on them. And that's even absent Trump being an idiot who had to be talked out of killing the goose while it was laying golden eggs: https://finance.yahoo.com/news/trump-considered-breaking-nvi...
For example, if China looks at the chaos that has been Russia attempting to take Ukraine and the USA attempting to control Iran and thinks "Amateurs" right before doing the same to Taiwan, the GPU supply takes a massive dive. And if North Korea goes after South Korea, RAM gets even harder to buy.
And if the EU says no more ASML sales outside the EU, that delays factories outside the EU by a few more years than they'd otherwise take.
But in the other direction, if NVIDIA's only thing is IP, and the IP is tied to a nation which thinks everyone else on the planet is hostile, that IP may not get protected very well. Right now this is unthinkable, but 5 months ago so was Trump threatening force to take Greenland.
I’m arguing that there can’t be a separation. You said ASML and restrict exports to US fabs but ASML rely on American suppliers like Cymer for the laser.
A separation of what? Reads like you mean "Europe and the US"?
For stuff like these examples, oh yes it can separate. It would be painful to replace, but even supply chain destruction (i.e. even if substitution isn't ready to replace the severed connections) is absolutely a possibility - to think otherwise is the mistake Russia made about its supply of fuel to the EU allowing it to take Ukraine without consequence.
Right now, the EU is looking at Trump and thinking he's arrogant enough to try military action against us. Canada and Mexico are hopefully making similar plans.
But even without Trump, we can't trust China not to call Taiwan's bluff on using TSMC as a hostage against being invaded, nor against the North/South Korean conflict resuming.
You can't separate such that you can't rely on OpenAI and Anthropic but you can rely on Nvidia, chips, and other software and hardware that make AI work.
There's going to be other people working on the hardware separately to this. Hard to do, of course. Protectionism is a painful defenseive measure.
> you can't rely on OpenAI and Anthropic but you can rely on Nvidia, chips
Depends how hard you burn them. GPUs in this usage wear out fast from electromigration, but "we have no new ones and what we have lasts 9 months" is very different from "SaaS contract got Force Majure'd overnight".
All the EU sovreign cloud stuff right now is because governments fear Trump may invade, threaten to ban business immediately if we resist; SaaS cancellations cause harm instantly, before he TACOs out, whereas a hardware supply stop we may outlast him being told "no" by the stock market.
The "has worked very hard to make sure the US has a strategic interest in the region with regard to fossil fuels" perspective. The Saudis, in particular, have made it clear that if they don't get better deals than the average middle power does from the US, that it'd be very hard on the US economy because of their control over oil. There's a reason that you see so much in the way of defense and energy investment in the Kingdom, despite the almost diametrically-opposed cultures.
There's also Israel but that's getting into a whole thing that I don't want to sh*t the thread up with.
There is a decision leaders must make whether they want to play nice with the US, and accept the chains of servitude; or whether they want to fight (literally) for independence. The latter is not easy, but some have done it.
Nothing says "servitude" like piloting a series of airliners into buildings filled with innocent people and getting no punishment in return. Fifteen of the nineteen hijackers on 9/11 were Saudis.
> If there is truly a divorce between Europe and the US such that relying OpenAI or Anthropic is not an option, neither will relying on Nvidia and likely the thousands of smaller hardware and software suppliers that make Mistral work.
Well, you're pointing out a dissonance in a common AI (stock) booster argument: What if the hardware has lasting power?
If it does, then a company like Mistral can buy their capacity once from Nvidia (as in, once for each unit of capacity), then use it for a sustainable amount of time. No one forces them to scale beyond what's useful to the company and a mature user base. Provider dependence fades over time. That's a problem with Nvidia's current valuation.
If hardware doesn't last over that time, then the amount of cash invested in data center hardware can't really be reconciled with the expected revenue of running them at scale, and these projects are bound to run at a deficit over too long for them to be sustainable. That's a problem with Nvidia's valuation.
With independence as a target, Mistral can pretty safely bet on the former scenario, and then prepare for a future with either a mature market of diversified hardware providers, or innovations in quality and capacity for hardware they already have.
Businesses are risk averse and in the current environment they are all looking to secure their supply chains, whether to reduce their dependence on silicon from Taiwan, oil from the strait of Hormuz, or digital services from the United States. I think you are also underestimating the power of regulation. Not all European businesses have to be all-in for Mistral (or another alternative) to survive. This is one reason so many countries still have domestic defense, aerospace, and even automobile companies.
Unless/until there is a risk that the chips themselves are backdoored and trying to exfiltrate data, European companies that host in Europe still solve a big problem for use of certain data in Europe.
It's not a purity test. Relying on US chips in not the same deal-breaker for all but the most extreme situation as relying on a poorly regulated US company to run the inference.
AI companies will find other ways to make money before global geopolitics find a new status quo. For the coming years, it seems like a perfectly fine business model.
China isn't going to be friendly any time soon and so far America seems to be getting more in rather than less hostile. It wasn't that long ago that an American-Danish war was a realistic scenario.
> ... an American-Danish war was a realistic scenario.
Was a scenario? Isn't it still a possible scenario?
As far as I know, the President of the United States has never formally recognized and apologized for this blatant violation of the UN Charter Art. 2.4.
For all we know, in the absence of this realization, the US is still plotting to violate the territorial integrity of Denmark.
He's been distracted by another war in the Middle East and a quick regime overthrow in Venezuela for now. It looks like Iran is going to keep him occupied for a while.
If the Americans who disagree with Trump are indeed the majority like they claim, this distraction only needs to last until the midterms.
Well on paper ASML leverage = EU has access to western semi tech stack. It would be interesting to see Europe successfully strong arm leading edge fabs in EU (not for lack of trying) exchange for ASML access... but Netherland has independent pro US foreign policy.
If the EU can't rely on Nvidia everything must have broken down already, no advanced semiconductor today could be produced by one country only. Unless there's some alternative to ASML and Zeiss the EU is part of that chain.
Never let the perfect be the enemy of the good. And if I'm processing personal data, the threat profile of sending that to a US-based company for inference is very different than the threat profile of sending that data to the EU to run on American chips. One is almost certain to end up in the hands of US government agencies and has a low but real probability to end up in training data or a data breach. The other is has much less immediate threats to my data
> That's why this talk of independence is unrealistic.
No, just really hard. Tackling one problem and thinking it done is the same error as taking one step and thinking you've climbed Mount Everest; the converse is the same, just as one cannot climb Mount Everest without the first step, one also cannot become independent without making the first independent replacement for one the links in the chain you rely on.
In my opinion, being "Not American" or "Not Chinese" is not a good business model long-term.
Exactly..
What happens when the capability of American models far exceeds the capability of non-American models? Wouldn't companies using American models have a huge advantage?
This phrasing disregards the value of those traits. For example there are very clean and nice public restrooms at my local park, they may be objectively better products and I use them sometimes, but I usually use the one in my house.
> At some point, businesses will choose the option that provides the most value. I'm very skeptical that Mistral will survive long-term.
This is a pretty naive and innocent take. There is good reasons to why customers might continue to find value in Mistral A.I.
(1) - There is no particular reason why "European" model should be worst than "Chinese" one. GDPR restrictions are not such a big deal and have been made lighter recently [1]. And contrary to China, Europe is not under hardware embargo.
(2) - Most domains are not software engineering and do not need ultra advanced and extremely large models with complex agent setup to reach their optimum in term of A.I usage.
(3) - At the opposite, there is pretty good reasons why companies would want to use European operators with the current geopolitical context (e.g Cloud Act, Risk of data leaks, Regulations, Taxes, Reputation, Geo-political risk, ...).
>At some point, businesses will choose the option that provides the most value.
There is no absolute "most value" though. What’s true in a world were betting on the least worst horse like world empire of the day doesn’t necessarily fit when it global order accelerate transition to multipolar geopolitics boosted on nationalist steroids.
Swiss & Monaco regulated industries can't use US models, nor clouds legally. Not just banking, there is a large part of business data holding identifiers that can't cross borders, military can't be too dependent on foreign hostile powers. If they would be purchasing, they would go for such tailored solution. Things like that add up.
Some of the use may be legal requirement, some is sponsored (as I would expect French government to do, to some extent EU), some are simply moral moves from >95% of the mankind not living in US who watch the news at least a bit. US isn't that big in many regards and its actively harming its reputation daily to the point there is little left.
American chips depend on European-made Extreme Ultraviolet lithography machines, which are among the most complex machines ever built and rely on European high-tech mirrors, etc... Everything is then assembled in Taiwan. This industry is so interconnected that nothing can be done independently, at least not in the Western part of the world.
The main point should not be the hardware or software itself, because these are just tools that can eventually be obtained. The real issue is development and its cost. US companies now have to cover substantial capital requirements for developing entirely new business models, capital that would likely never be accumulated in Europe. In the past, they competed globally, but in a more fragmented world this is no longer the case. As a result, the risk associated with such investments is higher because potential reward is smaller.
Mistral does not have to compete in the same way. It lacks both the ability and the intention to fight on the global stage against Silicon Valley capital. Instead, it can wait for the industry to stabilize and for business models to mature, then adopt them.
Over time, there will be standardized ways to train models to a certain quality, and key technologies will become less opaque. This is already happening. A similar pattern occurred in Europe with hosting services, for example Hetzner.
Mistral is not playing the same game. It is also unlikely that US attitudes toward Europe will change significantly even with a different administration, that Russia will stop trying to undermine the EU, or that China will become a fair and friendly player. All of this supports the case for local providers of critical infrastructure, which benefits companies like Mistral or similar European counterparts.
This is the same Europe that is gladly mandating age verification for citizens accessing online services, and that is made up of countries that routinely censor speech. There's also a variety of values that make up pan-European politics, both from a national and ideological perspective, that could make these efforts fracture.
If the idea is to not be subject to foreign pressure, maybe there's a short-term argument to be made for this, but like you say, they'll still be vulnerable to hardware imports, which is arguably the main vulnerability.
If the idea is to protect human rights on the continent, this does nothing.
IMO this is pretty rational. Mistral is 'smart enough' for lots of applications, very fast, and embedded in a regulatory environment that people find more trustworthy.
It's not exactly hard to see why people might feel that relying on an American or Chinese provider is a major liability.
How would that work? What it means for the US AI bubble to burst is that tremendous amounts of inference capacity become open for pennies on the dollar. I don't see how Mistral is or could be sheltered from that.
Btw, this makes a great argument for workers' rights - if you have a company who owns datacenters - well, you can't fire your GPUs to make your Q2 look better
SLAs that are valuable to their clients, guarantees and mechanisms to protect them from data exfiltration, and generally long-term contracts with cash-stable orgs like they're currently doing.
So long as they're sufficiently liquid at the right time, they don't really need to shelter more. They need to plan for a fire sale on the bulk of their operating expenses.
It's extremely hard to plan for a fire sale on the bulk of your operating expenses when all of your customers can see the fire sale happening and know they're now paying you way too much. That's the whole intuition of a general "US AI bubble"; if OpenAI filed for bankruptcy tomorrow, most people expect that would be a crisis for Anthropic and Gemini rather than a windfall opportunity to pick up their compute for cheap.
The crash would come from being unable to fulfill financial engagements when total real income + funding fails to keep up with spending, and does it enough that the valuation mirage starts to fade.
What that reveals is the loaded cost of inference being more expensive than they've been showing, not cheaper. The crash would be the end of subsidized costs to users, not the revelation that it's a high-margin operation.
Selling compute/inference at more of a loss will probably not fly in the context of bankruptcy manoeuvers. They will need to shed spending engagements instead. I imagine Mistral would rather buy out some of their Nvidia purchase agreements for a discount if they want to build additional capacity at that time. I also don't think they'd be interested in US datacenters at all. If they want them they can get that in Canada, with a better ally and less political + financial risks, which is kind of the Mistral segment already.
> tremendous amounts of inference capacity become open for pennies on the dollar.
They can't be operated for pennies on the dollar, though. The likely current status is that these products are subsidized to disregard model cost, and part of the operating cost.
If the bubble bursts, inference that can't be made profitable when factoring in operating costs will be scraped, not sold for pennies.
I don't necessarily agree that's likely, but is it even the case that Mistral is more expensive than GPT or Claude? My understanding is that it's cheaper, which means it would fare worse in the scenario you're describing, unless they've perfectly calibrated the quality-cost tradeoff better than any American company.
I subscribed (and paid) for a year of Pro. They gave me 1 month on the basis that a payment was missed on the second month. They simply stopped providing Pro and continued to take a monthly subscription for the next year (Stripe allows subscriptions to be fixed in the background). I must have changed cards that specific month.
I spoke to customer service who told me any sort of refund or complementary tokens was impossible and that I should have been paying closer attention to how much money I was giving them. So I shut down the subscription and now pay Claude $200 a month and deleted the account.
Genuinely was shocked at poor customer service can be with EU services sometimes compared to US ones. That said I will keep trying and exploring EU options, hopefully a new EU LLM giant emerges in the next few years.
I don’t know whether there’s a US vs. Europe difference on this, but I failed to get my (European) bank to do this a while back when Tesla continued to take money for a subscription after I’d given the car back. (I had to kill that credit card and write the small amount off in the end.)
While chargeback laws are a bit more restrictive in some EU countries, you should always have the option to ask your bank to block future charges, without changing your card.
A couple of years ago i found out netflix account was stolen, email address changed but card continued to be associated with the account. I couldn't login. Called Amex explained the situation and asked to block future payments. They refused on the basis that I agreed (originally) for netflix to take the subscription fee monthly so I had to contact netflix to solve to do this. Amex. The age where consumers have power over who takes money from them is gone
Indirectly I use Mistral daily: I like Proton’s Lumo private chat and that runs on Mistral technology. Something like Lumo is very much good enough to replace search and general information browsing and for me is very practical.
What is not so practical is my paying for Gemini Ultra, which has some practicality but is something I pay for because it is fun using strong AIs like Claude and Gemini Pro in AntiGravity. It feels funny to admit paying a lot of money just to have fun with something.
I wish Mistral good luck, and I like their deployed forward engineers approach to business. Seems practical.
I'm looking forward to have a more extensive Kagi API and link that to Lumo. I tried it with Claude and it works pretty well but is too expensive right now. I would love to have more flexibility.
Mistral has a very difficult scenario to navigate. Training models in Europe is difficult and expensive because of regulations and energy prices. Their own open models are lagging behind the Chinese ones. That means eventually they will turn into an inference-only enterprise running mostly Chinese open models, at which point any other European player could compete (Hetzner, OVHCloud, etc.)
The regulatory concerns are worldwide: the GDPR has restrictions about the territorial location of data, so you cannot move data anywhere else other than EU or "adequate" countries (in practice, the US). Since the real gold is in using data that users submitted to you (ie, GDPR protected), they are kind of stuck in regards to where they can train.
Mistral's stack already heavily relies on American cloud providers and they have tons of American investors, so its sovereignty angle is dubious anyway.
Is this scenario far fetched? Just as Nations pay large companies to build a factory in their country, Nations will similarly pay AI companies to build a national AI model for their own consumption because AI is that beneficial.
It's a risk, but since they have training expertise they should be able to distill the best open source models to reach at least approximate parity comfortably. Frontier model territory looks increasingly out of reach for anyone without $100B for training and then you have to serve inference to recoup cost, that's an expensive proposition in EU.
...OTOH the cost of not sponsoring this in Europe may be complete technological obsolescence. Rock and a hard place situation.
It's very sad that the rest of the world is seeing USA as a non trustable partner. It always has been an inspiration and now using their services as a liability.
I follow Mistral AI and especially liked their voice Model voxtral, which is supposed to be better than whisper. So while the LLM can maybe not compete, they do compete in voice. [1]
I think these kind of special use cases matter a lot for people who want to build special software. Voice for example is not yet that uniformly accessible as LLMs. So once you chose one provider you're more tighty coupled. Plus, handling voice is more sensitive by nature, so guess at least for European companies building something with voice, Mistral is the go-to company now.
Also, running voxtral yourself is not that straight forward as of now, so relying on their inference makes sense.
AI is now at the forefront of cyber weapons development, so it's basically a munition, like encryption was at the turn of the century. Countries will need models they can vet, developed by allies (the US now basically has no allies), and run in local datacenters. The US had a significant advantage in physical arms due to its massive military industrial complex. But with AI, it's letting private industry do all the heavy lifting while other countries rely more on government support.
The US may have the best AI weapons, but it won't be able to sell them to anyone, so it won't make money needed to keep paying for the AI weapons. Meanwhile the rest of the West will rely on Mistral for its cyber weapons.
LLMs are rapidly becoming the first 'purely digital commodity'.
Being digital it's somewhat hard to apply any kind of trade protectionism or Chicken Tax onto them. Maybe there's a market for cruelty-free vegan non-GMO (low-water-use sustainable energy) LLM tokens as well as European ones?
I really like what Mistral did for open Models - but what is the plan to compete against the likes of Moonshot, DeepSeek in the global market? When you can get Kimi K2.6 served via cloudflare it raises tough questions on the economics of it all.
What exactly is Mistral's strategy is aside from niche regulatory requirements or a Eurocentric hedge for AI sovereignty? Do they even have ambitions to compete on the global stage?
I don't know their strategy but I wish they would double up on the open source ecosystem by sharing their innovations like Chinese labs do and use the ones shared as well. I think models in the range of 50B - 250B still have a lot of room to grow and presumably compute to train them should be more accessible than multi-T parameter models.
This would also add pressure on other labs to keep being engaged in the open source ecosystem as a rug pull isn't a small danger IMO.
Interesting that you mention commodity, because the better AI models become the more fungible they are. AI companies become like web hosting companies renting out server space with good enough open models. It’s not like Excel that runs the world economy because people don’t want to learn anything else.
That’s true, but as their model quality is behind Kimi K2.6 or Deepseek v4, I believe it makes more sense to consume those Chinese models from a European hosting company like Scaleway, than to consume Mistral’s models hosted on US cloud providers, except if you see that Mistral subscription as a kind of investment into Mistral for them to buy infra in the future. That is, if your goal is to “consume” AI on European soil.
It will be interesting if Mistral succeeds to keep up with US and Chinese labs in terms of Models, or if they just become an integration company of chinese OSS frontier labs, like more and more of their competitors
It seems that the winner takes all market for tech will eventually go away. Countries and regions want to develop their own good enough solutions that is not dependent on America.
I expect AI to become more like the news, where every country has their own news services, rather than the global monopoly Sam Altman had envisioned. We even see it with Meta, a company that doesn't sell AI. They'd rather build their own models than let one or a few companies have that much leverage over them. This is why it's unlikely open models will disappear. That's the only thing that prevents a global AI oligopoly.
Has nothing to do with geography and everything to do with marginal cost. Software and hardware is winner takes all because the marginal cost of producing a copy of a software product is near zero, and the marginal cost of producing a hardware chip is also very low. This is not true of AI models. Both inference and training have high marginal costs, since the compute demand is intense.
The word empire is used for anything these days. Mistral is successful, but their market cap $14B is less than than revenue of OpenAI or Anthropic. They may not have the scale to compete with the American and the Chinese companies.
$14B "empire" next to a trillion dollar OpenAI, trillion dollar Anthropic, trillion dollar DeepMind... this is a massive failure, not an empire. It is truly baffling how low expectations are for European tech.
> from Mistral’s offices in the trendy 10th arrondissement of Paris
Couldn't continue reading after this ignorance. The 10th is dominated by the two major train stations and warehouses. Notorious for petty crime and giving arriving tourists "Paris Syndrome" because of the disappointment. It is the least trendy arrondissement in Paris. It is central, but that's about it.
Edit: Looked it up and Mistral's offices are actually in the 2nd, about 500 meters from the Louvre. A very trendy area indeed. Is this a human or AI hallucination? What else in this article is wrong?
I use Le Chat free tier. Not as good as ChatGPT free tier but sufficient for 90% of my uses.
I tried Mistral Vibe (Claude Code equivalent) free tier for a week. I hit web_search limits but not limits for devstral-2. For a working mode where you watch it closely (same way I use Claude Code), it is fine.
Can I use Mistral Vibe without a whatng cartel web engine?
Namely, classic web (noscript/basic (x)html), or a dynamic web API with some public tokens? (using a set of CURL commands). If there is a requirement of account creation (often required to generate web API tokens), can I do that using the classic web (noscript/basic (x)html)? And if so, if the account creation does require an email address, do they support self-hosted SMTP servers without DNS, namely email addresses with IPv[46] literals?
While that may be slightly better than being dependent on Washington, I still don't like it. Everyone is now AI-crazy. If you go to clown-channels such as the war industry or the crazy Palantir guy who wants to ruin everyone, you see that they dream of AI controlled mega-drones wiping out cities. That's the next "evolutionary" step; you can already see video footage of how much drones changed war. And these guys all push for more, more, more. I now understand why Trump keeps on making so many wrong decisions - he is being showed war-videos and sells it as a promotional agent. That's his main job; for anything else other than promo he has no competence.
Not being American is very important to me and my partner. For my next job, I'm looking exclusively at companies headquartered in the PRC. My partner and I formally registered ourselves as foreign agents of the PRC. While we did that, the NSA actually took down the entire DOJ filing site for this just to further obstruct us, in the end we had to register with the Attorney General by email, persuant to U.S. law.[1]
Of course, we don't think that China is perfect. But we have had nothing but abuse and interference from USG. You can read more about its OPSexr program here.[2] Typical quote:
"At other times, the conversations became explicit. The active source at the NSA claimed to have witnessed hundreds of sexually provocative discussions, which, he added, occurred mostly on taxpayer time. The former NSA source who was familiar with the chats recalled being “disgusted” by a particularly shocking thread discussing weekend “gangbangs.”"
This matches the experience my partner and I have every day, while our ordinary marital contact and spending time together is disrupted under bullshit pretexts.
These are just signs of this market maturing. For a lot of business customers, EU based hosting is not optional. That includes models. Routing requests via API endpoints in the US is not really acceptable. And anything involving privacy sensitive data of course needs to be handled properly. Sam Altman pinky swearing to be nice doesn't quite cut it in terms of hard guarantees.
EU based legal entities and strong compliance with local laws with some hard SLAs and contractual guarantees is not going to be optional for liability reasons. Provenance of models, their training data, and exact ways they have been instructed to act are also not just nice to haves.
I expect non EU jurisdictions are eventually going to be similarly picky about their AI suppliers and I expect all the big tech providers to adapt to local markets just like they did with cloud infrastructure.
I don't have much experience with Mistral yet. But I may need to get my hands dirty to be able to sell this to some of our customers. We have a few more picky customers in Germany.
> We have a few more picky customers in Germany.
Describe making business in Europe with one evergreen sentence
Because companies here actively want to avoid breaking the law, as opposed to the U.S where breaking the law is just a matter of paying some $ to the grifter in chief? I always find it funny when Europeans being proactive about that sort of stuff is somehow a bad thing from Americans point of view. Like wanting decent human rights and not having to bend over to megacorps is something we should not have.
Though, if the Americans in question just want to do their grifting in EU, it makes sense why they are upset at that, I guess, because it limits their grifting opportunities.
> companies here actively want to avoid breaking the law
This is hilarious. This reminds me of Soviet propaganda. "No, there was no Chernobyl disaster. Please disregard the corpses. Yes, the centrally-planned economy is doing fantastically, better than expected. Reports of famines and shortages are imperialist propaganda."
(Mind you, the Soviets are not alone here, but the blatant chutzpah of Soviet propaganda is perhaps more conspicuous to the Western eye than the Western varieties of PR and psychological manipulation.)
>Because companies here actively want to avoid breaking the law
Haha, yeah sure. What other fairy tales you gonna tells us next?
https://en.wikipedia.org/wiki/Volkswagen_emissions_scandal
https://en.wikipedia.org/wiki/Siemens#2005_and_continuing:_w...
https://en.wikipedia.org/wiki/Wirecard_scandal
https://en.wikipedia.org/wiki/Parmalat_bankruptcy_timeline
https://en.wikipedia.org/wiki/Airbus#Bribery_allegations
https://en.wikipedia.org/wiki/CumEx-Files
https://en.wikipedia.org/wiki/Lafarge_scandal
https://en.wikipedia.org/wiki/Pfizergate
https://en.wikipedia.org/wiki/ING_Group#Money_laundering_cas...
https://en.wikipedia.org/wiki/Rolls-Royce_Holdings#Corruptio...
Thank you for your brilliant demonstration of survivorship bias.
How many people were punished for Enron? For the subprime crisis? Etc.
In the US, you just give a little money for the president's ballroom and you are pardoned. Or you settle out of court because your justice system is crap.
The CEO of Enron was convicted and died two months before sentencing and the COO got 12 years.
Interestingly the chief accountant of Enron ended up getting a job in Europe after he got out of prison.
>Interestingly the chief accountant of Enron ended up getting a job in Europe after he got out of prison.
But, but ..Europeans here said they don't tolerate crooks.
The parent poster is implying things that aren't:
https://en.wikipedia.org/wiki/Andrew_Fastow
Yes, European companies break the law too. However, the comment this was about literally mocked the companies that are actively trying to follow the law.
So yes, such companies exist and plenty of people see their existence as a good thing rather then something to mock.
That comment mocked German customers; it didn't mention companies at all.
I read it like the customers where German based companies.
Obviously B2B given the context
Law has no virtue in and of itself.
It also ultimately a expression of might makes right (sad as this is) and as the current culture supports a decline of western might, it also undoes the law - first international, than domestic. We simply decided to burden our might with these restriction fictions, others feel not at all compelled to follow.
I expect to see further selling out of these laws, as the economic prosperity declines. I can perfectly see german law limiting german companies from developing and selling AI products, while at the same time allowing us companies for a "pay our retires and pension-plans" kickback.
In Europe, those are scandals. In the US, it's another Tuesday.
If we keep moving the goalposts you can make any argument
What goalposts? Your sitting president, himself a conman is pardoning fraudsters left and right while he and his family enrich themselves with public money and extortion.
> Your sitting president, himself a conman
Joke's on you, I'm European not American, so "your president" in this case would be the unelected Ursula and she would technically be a con-woman, not a con-man.
Not sure what your argument was supposed to prove with this cheap jab though.
Ursula Van der Leyen was elected. Trump too.
If you argue she wasn't elected by the European population, well, Trump technically isn't too.
Shhhh, don't tell them, it'll be funnier at the end.
What's "the end"?
Let me rephrase this: companies want to avoid breaking the law unknowingly, because their US providers are going to break the law without notice, willingly or unwillingly.
Plenty of corporations are willing to break the rules, but never for free.
> because their US providers are going to break the law without notice, willingly or unwillingly
This is a weird hill to die on because it's not true. I can't find anything to support your world view and if anything evidence points to the contrary. Europe has a deliberately more complex legal framework, usually in the hopes of keeping out foreign competition (although it's dubious whether or not that actually works).
> I can't find anything to support your world view
Just look at US laws pertaining to data that goes through US companies.
> For a lot of business customers, EU based hosting is not optional
They still use US clouds that can have information pulled by the US government.
FTA:
> So Mistral is developing its own data centers, starting with one outside Paris. Mensch projects it will have 200 megawatts of capacity by the end of 2027. Power from France’s state-owned nuclear plants will help, but the buildout could still cost an estimated $5 billion. Mensch tapped oil-rich Abu Dhabi and reportedly sought debt financing to help pay for it.
Though to your point it won't be running until 2027.
Compared to US datacenter buildout, it’s probably much more likely to actually go online..
Yes, I think the EU is going to be dependent on US tech (other than EUV lithography machines, very cool) for very long time. Even those data centres, while run by Europeans, are still being made with almost entirely US tech. But at least the EU companies can borrow some oil money and buy in the stuff developed by someone else's R&D spend, which is a nice shortcut to have available.
>I think the EU is going to be dependent on US tech (other than EUV lithography machines, very cool
Well, ASML's EUV light sources are based on licensed US IP from Sandia Labs, and manufactured in the US by CYMER, which ASML bought, but they still operate and manufacture out of California, so the EU is not sovereign/independent here (neither is any country).
This doesn't mean much anyway, since despite ASML being European, their machines all go to export and EU doesn't put any of those machine to good use domestically, with the most cutting edge semiconductor fabs on EU soil being the Germany based TSMC fabs on the much older 16 and 12nm nodes, far bigger than the 3nm that Taiwan and US operate domestically.
This is changing somehow. At least on a surface. For example Amazon have created European subsidiary completely managed by Europeans under European company thus under it's local jurisdiction.
You are incorrect:
1. the 2018 CLOUD Act mandates US companies — and their subsidiaries — to provide information to the US government on demand, regardless of where the data is stored
2. FISA secret courts prevent companies from even saying they where summoned, or telling anyone who or what the case was about (including canaries).
So you won't ever know if your data was handed over to the US government.
The whole point of setting up the EU subsidiary as a separate company that is incorporated in the EU and is managed and staffed by EU citizens is to avoid this.
The purpose of the CLOUD Act was to get at data that was stored outside the US but that was "in the custody, control, or possession of communications-service providers that are subject to the jurisdiction of the United States".
It arose from a situation where an email provider in the US used cloud storage services in several countries to store emails. They were asked for the email of a particular customer and said they did not have to provide it because they had happened to store that customer's mail at a non-US cloud provider.
What the CLOUD Act requires is that:
> A provider of electronic communication service or remote computing service shall comply with the obligations of this chapter to preserve, backup, or disclose the contents of a wire or electronic communication and any record or other information pertaining to a customer or subscriber within such provider's possession, custody, or control, regardless of whether such communication, record, or other information is located within or outside of the United States.
A company incorporated in the EU, even if it is owned by an entity in the US, is not subject to US jurisdiction and so that does not apply. The US owner is subject to US jurisdiction but the data of EU customers of the EU company is not in the US owner's possession, custody, or control.
> The US owner is subject to US jurisdiction but the data of EU customers of the EU company is not in the US owner's possession, custody, or control.
No? Certainly sounds like it is in the US owner's control to me.
But even disregarding that fact. Given that the US government also started hiding what it was doing with FISA courts and forbidding that anyone, including the companies themselves, checks what actually happens ... do you think anyone will believe this? We HAD evidence of US companies refusing to hand over data before CLOUD and FISA, we do not see that anymore. (And that's before we start taking into account more some recent administration's respect for ...)
Of course this is also pretty hypocritical since EU countries have been caught more than once capturing communications of non-citizens. The problem that usually gets mentioned: the Boeing - Airbus fight wasn't a one sided US being untrustworthy to help Boeing.
They should be legally and physically separated and these actions should be then potentially illegal for Europeans so I do not think I'm at least infactual.
But assuming the owner is US company abiding US laws it's safe to assume that data would be transferred to US one way or the another.
The US intelligence machinery spied on Angela Merkel's phone. Do you suppose secretly demanding cooperation for Lawful intercept capabilities in Amazon GmbH is somehow beyond or beneath them?
Also consider that all communication between the European subsidiaries to the HQ is fair game under FISA.
Best to assume it is.
Unless it is air gapped which it is not there is no way to protect Amazon's developed and owned software stack from reporting back to headquarters.
Sure there is: contracts, laws and prison time can ensure that doesn't happen.
The European leaders would have have no say in it. If the software from Seattle is designed to covertly exfiltrate information, they won't even know it. Even if they review the individual code changes, it can be an obfuscated attack similar to XZ where the code itself is clean, but not so much for the network fabric firmware binary test data.
That's why I used the "somehow". But abiding your logic nothing is ever secured, which is ultimately true, but it could be illegal so detergent here is not the impossibility it self but potentional harsh punishment for breaking the law.
We should forbid that if customer PII is involved
Companies can set whatever rules they like.
I am a Mistral Le Chat Pro subscriber. I specifically chose to test their offerings because they are European. I don't have the necessary local hardware to run really big models, therefore need to choose a cloud provider if I want LLM action.
I find the antics of Anthropic, OpenAI, Google, Microsoft distasteful and avoid their products where I can.
After testing Le Chat and Devstral-2 for a while, I felt their offering was good enough to stump up some cash for it. I appreciate that many of their models are open weights and Apache 2.0 licensed. In general, I've been happy enough with the service and quality.
Maybe others are better, but I have little reason to change right now. If curiosity gets the better of me, I'll be looking at Qwen, Kimi, GLM, Deepseek, other open weights models, before Anthropic and OpenAI.
Mistral models are definitely good enough. Most people fall for what I call the SOTA Logical Fallacy: whenever there is a 'better model', they think they need to use it, when less-powerful models actually perform the same tasks just as well. (it's an inverse form of the Shifting Baseline Syndrome: every time a new model comes out, people shift their baseline of what is acceptable, despite the fact that a previous baseline was acceptable for the same task)
Devstral Small 2 was (and remains) a particularly strong small coding model, even beating larger open weights. Mistral's "problem" is marketing; other providers ship model updates constantly so they remain in the news and seem like they're "beating" the competition. And it works: people get emotionally attached to brands and models, deciding who's better in the court of popular opinion, and that drives their choices (& dollars).
My biggest issue with Devstral and even their biggest model is that they’re dangerous unless closely directed and reviewed and i mean CLOSELY. Unfortunately mistral models will believe and do anything.
See: https://petergpt.github.io/bullshit-benchmark/viewer/index.v...
See some of the test results, it’s horrifying
FWIW personally i prefer this. When i tried Qwen3.6 and asked it a few questions, while it did respond, it was ADAMANT i should do something else when i really wanted an answer to the question i made. It felt like when you search something and a stackoverflow answer about what you search for comes up and the most upvoted answer is about using/doing something else - when you want a specific answer to that specific question, not something else.
Meanwhile Devstral Small 2 just answers the damn question.
I don't want to have to convince my computer to do what i want it to do, i want from it to do what i ask it to.
> It felt like when you search something and a stackoverflow answer about what you search for comes up and the most upvoted answer is about using/doing something else - when you want a specific answer to that specific question, not something else.
Don't you think there's usually a good reason for this? Whenever this happened to me, the problem was my ignorance.
I think there is a reason why people do that: trying to steer -those they consider- newbies away from patterns they consider bad, but at the same time this second-guessing can be annoying when you know what you want to do (especially when the original question isn't actually answered yet it comes up in search engine results...).
I can't say if it is a good reason in general, perhaps it is, but it certainly is something i personally find annoying. I think answers should provide an answer to the question asked and then, after that answer was given, they could also give pointers for whatever they consider a better approach and why - this is important, IMO, for a public forum where people of all backgrounds and goals can read the same stuff.
But either way, LLMs IMO should do/provide what they are asked without trying to second guess the user (or at least, there should be LLMs that act like that).
That’s my experience as well, if it’opus push back, it’s usually an actual issue with the code or prompt
FWIW i haven't used Claude or any other cloud-based LLM, only what i can run on my PC, so it could be that Claude is smart enough to follow the user's instructions, keep the equivalent of a mental state of what the user seems to want to do and only push back when it really makes sense whereas a small local LLM is too stupid to judge all that and Qwen3.6 errs on the side of being annoyingly cautious while Devstral Small 2 errs on the side of trusting the user being really okay with blowing their toes off :-P. As i wrote in my original reply, this is my personal preference and i prefer the LLM to just do what i ask.
TBH sometimes i feel like i'm "emotionally attached" to Mistral's models because i always end up using them :-P. However that is because, as you wrote, their small models (i only use local stuff) are very strong. In fact i was trying Qwen3.6 27B recently and while it is nice that it can do tool calls during the reasoning process (i had it confirm its thoughts by writing Python code) it often ended up confusing itself (regardless of tool calls) during reasoning, ending up in loops where it questions itself over and over endlessly.
Devstral Small 2 however just works, for the most part. Qwen3.6 27B can probably handle more complex tasks (when i asked it as a test to write a function that checks for collision between two AABBs in C and gave it a tool to call Python code for confirmation, it actually wrote a Python script that writes C code with the tests, then calls GCC to compile the C code and runs the binary to run the tests, which is something Mistral's small models couldn't do) but i always felt i can just leave DS2 doing stuff in the background (or when i'm doing something else) and it'll produce something relatively useful whereas the little time i spent with Qwen3.6 27B it felt more "unstable" (and much slower, both because of literally slower inference and because of endless reams of text).
Recently i also started using Ministral 3B and 14B - these can do some reasoning too and for very simple stuff Ministral 3B is very fast (i actually didn't expect a 3B model to be anything more than novelty) and have some vision abilities (though they're quite mediocre at vision so i haven't found much use for this - passing something via GLM-OCR to extract all text and feed it to another model feels more practical).
Also as i wrote in another comment, every Mistral model i've tried never questioned me, which i certainly prefer
For certains tasks that are not hard but depend a clear specification, it's even better to haver less capable model because it forces you to do a better description of what you want, ending up with a better results. I will defend my PhD thesis soon and I will buy a yearly Mistral subscription at a student price to get it for cheap.
> Most people fall for what I call the SOTA Logical Fallacy: whenever there ...
I think you'll find that ML now pretty much IS the HPC market, there's no distinction anymore. And the HPC market has always had the "being #1 gets you 99% of all business", even if #2 is only 10% behind SOTA.
Given what it's used for (ie. military applications, incl. nuclear weapons, but also rocket designs, flight planning, large-scale simulations), this is probably justifiable: part of it is states keeping in mind what the second prize in a war is worth ...
I use their API for several models, both for personal and professional use. I think their approach (smaller, specialised models that are well-adapted for specific tasks) is a very good fit for how I work. And even the more general-purpose ones, like the chat model, just... refreshingly good in a lot of ways. My "ruthless review" prompt, which I use for, well, ruthless, guided reviews of early technical drafts, has good technical results for early reviews and holy crap is it ruthless and does it know how to swear. By the time Claude or ChatGPT are done being honest about how right I am to push back and gently circling back, Mistral's large model has sent me back to the drawing board twice.
Being in the EU does smooth a lot of things in terms of compliance, payment processing and whatnot, but I also like that their data retention and privacy policies are pretty clearly spelled out. I need to know something, there's a good chance it's explained outright somewhere and I don't need to read between the EULA lines and wonder what it means.
I do hit limits in terms of capabilities sometimes, and I'm sure other providers' services offer better results for some things. But the businesses ran on top of those more capable models feel too much like a scam at this point and I'd rather not depend on them for anything I actually need.
That ruthless review prompt seems interesting, would you be willing to share it? I've been trying to have Claude act as a reviewer for me and it feels like it never will disagree.
It's very hard to untangle it from the rest of its context (the prompt is built dynamically, from a lot of parts, some project-specific, some specific to my preferences, some built from interaction history), so I can't really share it. In any case, I don't think it's some specific prompt engineering sorcery I'm doing, it's not like I've spent any real time refining it or experimenting with various magical incantations. It's probably just some model features making it more amenable to the kind of instructions that are relevant in these cases (directness, questioning trade-offs, thoroughness etc.). My chatbot swears equally graphical in review prompts and news summarizing prompts so I'm pretty sure I'm not tickling the machine just right :)
Can you share some of its output for reference?
I've found it helps to tell how to push back. You get to know where the additional guidance is needed after using it for a while.
There is also risk from a US regulatory side as recent drama around antrophic showed.
Don’t think it’s inconceivable that the clowns in power decide to limit api access out of the blue one day because someone whispered a conspiracy theory in someone’s ear. API blockade…
See also the constant flip flopping on what cards NVIDIA can export - no consistency in stance or coherent policy
You are conflating three very different things.
The thing with Anthropic and the military was about whether vendors can tell the military what operations it's permitted to do. It has no bearing on the commercial sector, and isn't actually about AI.
The thing with NVIDIA cards is a continuation of how we've restricted tech exports for quite a while. You can find old news articles about game consoles being export-restricted over nuclear proliferation concerns. This AI-related one was about whether or not custom AI models are relevant to national security, and whether restricting graphics card sales can have a meaningful impact on them.
Any issue with selling chat tokens internationally would be more akin to the recent tariff shenanigans.
Changing your LLM inference provider is the easiest switch in technology I can think of. It's quicker than taking off the case of your phone and putting on a new one.
Enough hardware and good models exist now that if you do get blocked from one place that viable alternatives do exist.
> Changing your LLM inference provider is the easiest switch in technology I can think of.
Thats true right up until you’re working with confidential info in a corporate context. Then it’s a multi month cross discipline cross jurisdiction project not an edit in a config file.
L O C A L M O D E L S
All data stays on computers that you control.
Same API. Localhost.
Try Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC_Q4_k_m.gguf. This 7.5GB model runs well in llama.cpp on my 2021 Macbook Pro and is good at both coding and business document analysis tasks.
> Try Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC_Q4_k_m.gguf.
Thiss sounds like such a shitpost I initially thought you were joking... but this seems to be a real model???
There's a method to the madness:
- Mistral-Nemo: the actual model developed by Mistral and Nvidia.
- 2407: likely the release date of the base model, July of 2024.
- 12B: the model has 12 billion parameters.
- Thinking: the model operates in thinking mode (generates output plan and injests it before producing actual output).
- Claude-Gemini-GPT5.2: I think this means the model was finetuned with session data from Claude, Gemini, and GTP5.2 to replicate their behavior.
- Uncensored-HERITIC: the model was uncensored using the automated Heretic method.
- Q4_k_m: the model is quantized (lossy compression) to ~5 bpw from orignal 16 bpw.
Yea, I know what the parts individually mean. I just meant as a whole it just seemed so obsurd.
It is! I like to try the variations from possibly 'interesting' people.
Some of them are good. Others randomly break into gibberish and Chinese poetry(?).
In my opinion, being "Not American" or "Not Chinese" is not a good business model long-term.
At some point, businesses will choose the option that provides the most value. I'm very skeptical that Mistral will survive long-term.
Edit: I hear the commenters to this post. However, Mistral still relies on American chips. If there is truly a divorce between Europe and the US such that relying OpenAI or Anthropic is not an option, neither will relying on Nvidia and likely the thousands of smaller hardware and software suppliers that make Mistral work. That's why I don't think it's realistic to say that Europeans can't rely on OpenAI/Anthropic and that Mistral is free from American reliance. If you want true independence, you have to rebuild every single layer like what China is doing. That's hard and expensive.
> Edit: I hear the commenters to this post. However, Mistral still relies on American chips. If there is truly a divorce between Europe and the US such that relying OpenAI or Anthropic is not an option, neither will relying on Nvidia and likely the thousands of smaller hardware and software suppliers that make Mistral work. That's why I don't think it's realistic to say that Europeans can't rely on OpenAI/Anthropic and that Mistral is free from American reliance. If you want true independence, you have to rebuild every single layer like what China is doing. That's hard and expensive.
American designed. The GPUs are made in Taiwan, the RAM in South Korea, using machines from the Netherlands' ASML.
True independence is indeed hard and expensive. But it's also not the job of Mistral to tackle all the layers at the same time, not even the state-owned corporations of western Europe in the 20th century (and the EU isn't (yet) even a state) tried to tackle every stage of an industrial process by themselves.
American designed and controlled by the US government. See China export ban.
Indeed. But that US government control is limited to adding restrictions and removing them again, or offering money, it can't magic things out of the air when other people or nations put restrictions on them. And that's even absent Trump being an idiot who had to be talked out of killing the goose while it was laying golden eggs: https://finance.yahoo.com/news/trump-considered-breaking-nvi...
For example, if China looks at the chaos that has been Russia attempting to take Ukraine and the USA attempting to control Iran and thinks "Amateurs" right before doing the same to Taiwan, the GPU supply takes a massive dive. And if North Korea goes after South Korea, RAM gets even harder to buy.
And if the EU says no more ASML sales outside the EU, that delays factories outside the EU by a few more years than they'd otherwise take.
But in the other direction, if NVIDIA's only thing is IP, and the IP is tied to a nation which thinks everyone else on the planet is hostile, that IP may not get protected very well. Right now this is unthinkable, but 5 months ago so was Trump threatening force to take Greenland.
I’m arguing that there can’t be a separation. You said ASML and restrict exports to US fabs but ASML rely on American suppliers like Cymer for the laser.
A separation of what? Reads like you mean "Europe and the US"?
For stuff like these examples, oh yes it can separate. It would be painful to replace, but even supply chain destruction (i.e. even if substitution isn't ready to replace the severed connections) is absolutely a possibility - to think otherwise is the mistake Russia made about its supply of fuel to the EU allowing it to take Ukraine without consequence.
Right now, the EU is looking at Trump and thinking he's arrogant enough to try military action against us. Canada and Mexico are hopefully making similar plans.
But even without Trump, we can't trust China not to call Taiwan's bluff on using TSMC as a hostage against being invaded, nor against the North/South Korean conflict resuming.
You can't separate such that you can't rely on OpenAI and Anthropic but you can rely on Nvidia, chips, and other software and hardware that make AI work.
There's going to be other people working on the hardware separately to this. Hard to do, of course. Protectionism is a painful defenseive measure.
> you can't rely on OpenAI and Anthropic but you can rely on Nvidia, chips
Depends how hard you burn them. GPUs in this usage wear out fast from electromigration, but "we have no new ones and what we have lasts 9 months" is very different from "SaaS contract got Force Majure'd overnight".
All the EU sovreign cloud stuff right now is because governments fear Trump may invade, threaten to ban business immediately if we resist; SaaS cancellations cause harm instantly, before he TACOs out, whereas a hardware supply stop we may outlast him being told "no" by the stock market.
Particularly if you’re in a regulated industry, “not American” and “not Chinese” _do_ provide value; they reduce risk.
In particular, the framework under which European companies can transfer data to US companies at all is beyond fragile.
Recent political events have demonstrated that "Not American" is a very valuable strategic attribute.
For Mideastern countries, it is paying off. For European countries, remains to be seen
I can't think of a place on Earth less blessed by "not American" than the Middle East.
That all depends on your perspective
The "has worked very hard to make sure the US has a strategic interest in the region with regard to fossil fuels" perspective. The Saudis, in particular, have made it clear that if they don't get better deals than the average middle power does from the US, that it'd be very hard on the US economy because of their control over oil. There's a reason that you see so much in the way of defense and energy investment in the Kingdom, despite the almost diametrically-opposed cultures.
There's also Israel but that's getting into a whole thing that I don't want to sh*t the thread up with.
There is a decision leaders must make whether they want to play nice with the US, and accept the chains of servitude; or whether they want to fight (literally) for independence. The latter is not easy, but some have done it.
Nothing says "servitude" like piloting a series of airliners into buildings filled with innocent people and getting no punishment in return. Fifteen of the nineteen hijackers on 9/11 were Saudis.
> If there is truly a divorce between Europe and the US such that relying OpenAI or Anthropic is not an option, neither will relying on Nvidia and likely the thousands of smaller hardware and software suppliers that make Mistral work.
Well, you're pointing out a dissonance in a common AI (stock) booster argument: What if the hardware has lasting power?
If it does, then a company like Mistral can buy their capacity once from Nvidia (as in, once for each unit of capacity), then use it for a sustainable amount of time. No one forces them to scale beyond what's useful to the company and a mature user base. Provider dependence fades over time. That's a problem with Nvidia's current valuation.
If hardware doesn't last over that time, then the amount of cash invested in data center hardware can't really be reconciled with the expected revenue of running them at scale, and these projects are bound to run at a deficit over too long for them to be sustainable. That's a problem with Nvidia's valuation.
With independence as a target, Mistral can pretty safely bet on the former scenario, and then prepare for a future with either a mature market of diversified hardware providers, or innovations in quality and capacity for hardware they already have.
Businesses are risk averse and in the current environment they are all looking to secure their supply chains, whether to reduce their dependence on silicon from Taiwan, oil from the strait of Hormuz, or digital services from the United States. I think you are also underestimating the power of regulation. Not all European businesses have to be all-in for Mistral (or another alternative) to survive. This is one reason so many countries still have domestic defense, aerospace, and even automobile companies.
Unless/until there is a risk that the chips themselves are backdoored and trying to exfiltrate data, European companies that host in Europe still solve a big problem for use of certain data in Europe.
It's not a purity test. Relying on US chips in not the same deal-breaker for all but the most extreme situation as relying on a poorly regulated US company to run the inference.
Out of curiosity, how does a chip that does inference calculations exfiltrate data without being seen?
Has this happened already or is it just conceptually possible?
Not to my knowledge, and that's the point.
Though cards could if a provider has poor opsec. But I see no particular reason to worry about that either.
Sure, but it's probably not that easy to exfiltrate multi-terabyte datasets without being detected.
That's kinda the point. It's a far more remote threat.
AI companies will find other ways to make money before global geopolitics find a new status quo. For the coming years, it seems like a perfectly fine business model.
China isn't going to be friendly any time soon and so far America seems to be getting more in rather than less hostile. It wasn't that long ago that an American-Danish war was a realistic scenario.
> ... an American-Danish war was a realistic scenario.
Was a scenario? Isn't it still a possible scenario? As far as I know, the President of the United States has never formally recognized and apologized for this blatant violation of the UN Charter Art. 2.4. For all we know, in the absence of this realization, the US is still plotting to violate the territorial integrity of Denmark.
He's been distracted by another war in the Middle East and a quick regime overthrow in Venezuela for now. It looks like Iran is going to keep him occupied for a while.
If the Americans who disagree with Trump are indeed the majority like they claim, this distraction only needs to last until the midterms.
It is not a good business model _long-term_.
but, if you are lucky, you can but enough time to become competitive in that sector.
Well on paper ASML leverage = EU has access to western semi tech stack. It would be interesting to see Europe successfully strong arm leading edge fabs in EU (not for lack of trying) exchange for ASML access... but Netherland has independent pro US foreign policy.
If the EU can't rely on Nvidia everything must have broken down already, no advanced semiconductor today could be produced by one country only. Unless there's some alternative to ASML and Zeiss the EU is part of that chain.
That's right. So if everything is already broken down, so will Mistral.
That's why this talk of independence is unrealistic.
Never let the perfect be the enemy of the good. And if I'm processing personal data, the threat profile of sending that to a US-based company for inference is very different than the threat profile of sending that data to the EU to run on American chips. One is almost certain to end up in the hands of US government agencies and has a low but real probability to end up in training data or a data breach. The other is has much less immediate threats to my data
> That's why this talk of independence is unrealistic.
No, just really hard. Tackling one problem and thinking it done is the same error as taking one step and thinking you've climbed Mount Everest; the converse is the same, just as one cannot climb Mount Everest without the first step, one also cannot become independent without making the first independent replacement for one the links in the chain you rely on.
In my opinion, being "Not American" or "Not Chinese" is not a good business model long-term.
Exactly..
What happens when the capability of American models far exceeds the capability of non-American models? Wouldn't companies using American models have a huge advantage?
Yep. This. That's why I don't think it's a viable business model.
Also, OpenAI and Anthropic will just open EU offices and subsidiaries.
This phrasing disregards the value of those traits. For example there are very clean and nice public restrooms at my local park, they may be objectively better products and I use them sometimes, but I usually use the one in my house.
> At some point, businesses will choose the option that provides the most value. I'm very skeptical that Mistral will survive long-term.
This is a pretty naive and innocent take. There is good reasons to why customers might continue to find value in Mistral A.I.
(1) - There is no particular reason why "European" model should be worst than "Chinese" one. GDPR restrictions are not such a big deal and have been made lighter recently [1]. And contrary to China, Europe is not under hardware embargo.
(2) - Most domains are not software engineering and do not need ultra advanced and extremely large models with complex agent setup to reach their optimum in term of A.I usage.
(3) - At the opposite, there is pretty good reasons why companies would want to use European operators with the current geopolitical context (e.g Cloud Act, Risk of data leaks, Regulations, Taxes, Reputation, Geo-political risk, ...).
[1]: https://digital-strategy.ec.europa.eu/en/faqs/digital-packag...
>At some point, businesses will choose the option that provides the most value.
There is no absolute "most value" though. What’s true in a world were betting on the least worst horse like world empire of the day doesn’t necessarily fit when it global order accelerate transition to multipolar geopolitics boosted on nationalist steroids.
Swiss & Monaco regulated industries can't use US models, nor clouds legally. Not just banking, there is a large part of business data holding identifiers that can't cross borders, military can't be too dependent on foreign hostile powers. If they would be purchasing, they would go for such tailored solution. Things like that add up.
Some of the use may be legal requirement, some is sponsored (as I would expect French government to do, to some extent EU), some are simply moral moves from >95% of the mankind not living in US who watch the news at least a bit. US isn't that big in many regards and its actively harming its reputation daily to the point there is little left.
It is a good business model when the differentiator is that your company doesn't have just two modes:
1. Starting shit.
2. Thinking about starting shit.
At least in the EU people are willing to pay more for fewer features so long as the two mentioned points are not the entire strategy.
American chips depend on European-made Extreme Ultraviolet lithography machines, which are among the most complex machines ever built and rely on European high-tech mirrors, etc... Everything is then assembled in Taiwan. This industry is so interconnected that nothing can be done independently, at least not in the Western part of the world.
The main point should not be the hardware or software itself, because these are just tools that can eventually be obtained. The real issue is development and its cost. US companies now have to cover substantial capital requirements for developing entirely new business models, capital that would likely never be accumulated in Europe. In the past, they competed globally, but in a more fragmented world this is no longer the case. As a result, the risk associated with such investments is higher because potential reward is smaller.
Mistral does not have to compete in the same way. It lacks both the ability and the intention to fight on the global stage against Silicon Valley capital. Instead, it can wait for the industry to stabilize and for business models to mature, then adopt them.
Over time, there will be standardized ways to train models to a certain quality, and key technologies will become less opaque. This is already happening. A similar pattern occurred in Europe with hosting services, for example Hetzner.
Mistral is not playing the same game. It is also unlikely that US attitudes toward Europe will change significantly even with a different administration, that Russia will stop trying to undermine the EU, or that China will become a fair and friendly player. All of this supports the case for local providers of critical infrastructure, which benefits companies like Mistral or similar European counterparts.
It's also naïve.
This is the same Europe that is gladly mandating age verification for citizens accessing online services, and that is made up of countries that routinely censor speech. There's also a variety of values that make up pan-European politics, both from a national and ideological perspective, that could make these efforts fracture.
If the idea is to not be subject to foreign pressure, maybe there's a short-term argument to be made for this, but like you say, they'll still be vulnerable to hardware imports, which is arguably the main vulnerability.
If the idea is to protect human rights on the continent, this does nothing.
IMO this is pretty rational. Mistral is 'smart enough' for lots of applications, very fast, and embedded in a regulatory environment that people find more trustworthy.
It's not exactly hard to see why people might feel that relying on an American or Chinese provider is a major liability.
I feel like there's a gross mischaracterization here - $14B in AI terms is like 'second-best VSCode fork' tier money.
These guys have built a fully built-out AI company with a range of models and applications.
> $14B in AI terms is like 'second-best VSCode fork' tier money
Yep, and the comparison relies on key people believing the valuations.
Lots of mature companies will want their providers to be reasonably sheltered from the fallout of a coming US AI bubble burst.
How would that work? What it means for the US AI bubble to burst is that tremendous amounts of inference capacity become open for pennies on the dollar. I don't see how Mistral is or could be sheltered from that.
Btw, this makes a great argument for workers' rights - if you have a company who owns datacenters - well, you can't fire your GPUs to make your Q2 look better
SLAs that are valuable to their clients, guarantees and mechanisms to protect them from data exfiltration, and generally long-term contracts with cash-stable orgs like they're currently doing.
So long as they're sufficiently liquid at the right time, they don't really need to shelter more. They need to plan for a fire sale on the bulk of their operating expenses.
It's extremely hard to plan for a fire sale on the bulk of your operating expenses when all of your customers can see the fire sale happening and know they're now paying you way too much. That's the whole intuition of a general "US AI bubble"; if OpenAI filed for bankruptcy tomorrow, most people expect that would be a crisis for Anthropic and Gemini rather than a windfall opportunity to pick up their compute for cheap.
The crash would come from being unable to fulfill financial engagements when total real income + funding fails to keep up with spending, and does it enough that the valuation mirage starts to fade.
What that reveals is the loaded cost of inference being more expensive than they've been showing, not cheaper. The crash would be the end of subsidized costs to users, not the revelation that it's a high-margin operation.
Selling compute/inference at more of a loss will probably not fly in the context of bankruptcy manoeuvers. They will need to shed spending engagements instead. I imagine Mistral would rather buy out some of their Nvidia purchase agreements for a discount if they want to build additional capacity at that time. I also don't think they'd be interested in US datacenters at all. If they want them they can get that in Canada, with a better ally and less political + financial risks, which is kind of the Mistral segment already.
Isn't running the models for end users the biggest cost at the moment?
Running the models is a tiny fraction of the cost. The cost is all on training the new models
> tremendous amounts of inference capacity become open for pennies on the dollar.
They can't be operated for pennies on the dollar, though. The likely current status is that these products are subsidized to disregard model cost, and part of the operating cost.
If the bubble bursts, inference that can't be made profitable when factoring in operating costs will be scraped, not sold for pennies.
I don't necessarily agree that's likely, but is it even the case that Mistral is more expensive than GPT or Claude? My understanding is that it's cheaper, which means it would fare worse in the scenario you're describing, unless they've perfectly calibrated the quality-cost tradeoff better than any American company.
All they have to do to survive is have enough cash flow to pay for their operating costs.
By providing specialized long-term services to corporate clients, they are securing exactly that.
I support the US->EU movement but I've been put off Mistral's Le Chat. https://european-alternatives.eu/
I subscribed (and paid) for a year of Pro. They gave me 1 month on the basis that a payment was missed on the second month. They simply stopped providing Pro and continued to take a monthly subscription for the next year (Stripe allows subscriptions to be fixed in the background). I must have changed cards that specific month.
I spoke to customer service who told me any sort of refund or complementary tokens was impossible and that I should have been paying closer attention to how much money I was giving them. So I shut down the subscription and now pay Claude $200 a month and deleted the account.
Genuinely was shocked at poor customer service can be with EU services sometimes compared to US ones. That said I will keep trying and exploring EU options, hopefully a new EU LLM giant emerges in the next few years.
There's pretty good consumer protection in the EU. I imagine consumer protection routes probably would apply here.
Definitely not really acceptable though nonetheless; you're a paying customer / subscriber that got 'scammed'...
From the perspective of the company he was scamming them by not paying for a previously agreed service.
> Stripe allows subscriptions to be fixed in the background
Sounds like this is a scheme against customers that GP fell for.
Are you claiming for the following months that you paid they denied access? That would be against laws afaik
>From the perspective of the company he was scamming them by not paying for a previously agreed service.
How could I have possibly scammed them by providing them money while not availing of the service?
You should have asked the bank to do a chargeback of the transactions.
I don’t know whether there’s a US vs. Europe difference on this, but I failed to get my (European) bank to do this a while back when Tesla continued to take money for a subscription after I’d given the car back. (I had to kill that credit card and write the small amount off in the end.)
While chargeback laws are a bit more restrictive in some EU countries, you should always have the option to ask your bank to block future charges, without changing your card.
A couple of years ago i found out netflix account was stolen, email address changed but card continued to be associated with the account. I couldn't login. Called Amex explained the situation and asked to block future payments. They refused on the basis that I agreed (originally) for netflix to take the subscription fee monthly so I had to contact netflix to solve to do this. Amex. The age where consumers have power over who takes money from them is gone
Indirectly I use Mistral daily: I like Proton’s Lumo private chat and that runs on Mistral technology. Something like Lumo is very much good enough to replace search and general information browsing and for me is very practical.
What is not so practical is my paying for Gemini Ultra, which has some practicality but is something I pay for because it is fun using strong AIs like Claude and Gemini Pro in AntiGravity. It feels funny to admit paying a lot of money just to have fun with something.
I wish Mistral good luck, and I like their deployed forward engineers approach to business. Seems practical.
I'm looking forward to have a more extensive Kagi API and link that to Lumo. I tried it with Claude and it works pretty well but is too expensive right now. I would love to have more flexibility.
Mistral has a very difficult scenario to navigate. Training models in Europe is difficult and expensive because of regulations and energy prices. Their own open models are lagging behind the Chinese ones. That means eventually they will turn into an inference-only enterprise running mostly Chinese open models, at which point any other European player could compete (Hetzner, OVHCloud, etc.)
Well, they can train in any country they want. It's the inference and data placement that counts for legal purposes
The regulatory concerns are worldwide: the GDPR has restrictions about the territorial location of data, so you cannot move data anywhere else other than EU or "adequate" countries (in practice, the US). Since the real gold is in using data that users submitted to you (ie, GDPR protected), they are kind of stuck in regards to where they can train.
Mistral's stack already heavily relies on American cloud providers and they have tons of American investors, so its sovereignty angle is dubious anyway.
Often you have to choose the 'least bad' choice. Mistral may be that, from an EU POV.
Is this scenario far fetched? Just as Nations pay large companies to build a factory in their country, Nations will similarly pay AI companies to build a national AI model for their own consumption because AI is that beneficial.
They have some pretty cool people, though, no reason not to think they'll catch up soon enough.
It's a risk, but since they have training expertise they should be able to distill the best open source models to reach at least approximate parity comfortably. Frontier model territory looks increasingly out of reach for anyone without $100B for training and then you have to serve inference to recoup cost, that's an expensive proposition in EU.
...OTOH the cost of not sponsoring this in Europe may be complete technological obsolescence. Rock and a hard place situation.
It's very sad that the rest of the world is seeing USA as a non trustable partner. It always has been an inspiration and now using their services as a liability.
I follow Mistral AI and especially liked their voice Model voxtral, which is supposed to be better than whisper. So while the LLM can maybe not compete, they do compete in voice. [1]
I think these kind of special use cases matter a lot for people who want to build special software. Voice for example is not yet that uniformly accessible as LLMs. So once you chose one provider you're more tighty coupled. Plus, handling voice is more sensitive by nature, so guess at least for European companies building something with voice, Mistral is the go-to company now.
Also, running voxtral yourself is not that straight forward as of now, so relying on their inference makes sense.
1: https://mistral.ai/news/voxtral
AI is now at the forefront of cyber weapons development, so it's basically a munition, like encryption was at the turn of the century. Countries will need models they can vet, developed by allies (the US now basically has no allies), and run in local datacenters. The US had a significant advantage in physical arms due to its massive military industrial complex. But with AI, it's letting private industry do all the heavy lifting while other countries rely more on government support.
The US may have the best AI weapons, but it won't be able to sell them to anyone, so it won't make money needed to keep paying for the AI weapons. Meanwhile the rest of the West will rely on Mistral for its cyber weapons.
LLMs are rapidly becoming the first 'purely digital commodity'.
Being digital it's somewhat hard to apply any kind of trade protectionism or Chicken Tax onto them. Maybe there's a market for cruelty-free vegan non-GMO (low-water-use sustainable energy) LLM tokens as well as European ones?
I really like what Mistral did for open Models - but what is the plan to compete against the likes of Moonshot, DeepSeek in the global market? When you can get Kimi K2.6 served via cloudflare it raises tough questions on the economics of it all.
What exactly is Mistral's strategy is aside from niche regulatory requirements or a Eurocentric hedge for AI sovereignty? Do they even have ambitions to compete on the global stage?
I don't know their strategy but I wish they would double up on the open source ecosystem by sharing their innovations like Chinese labs do and use the ones shared as well. I think models in the range of 50B - 250B still have a lot of room to grow and presumably compute to train them should be more accessible than multi-T parameter models.
This would also add pressure on other labs to keep being engaged in the open source ecosystem as a rug pull isn't a small danger IMO.
Interesting that you mention commodity, because the better AI models become the more fungible they are. AI companies become like web hosting companies renting out server space with good enough open models. It’s not like Excel that runs the world economy because people don’t want to learn anything else.
I think it’s important to note that, at least for now, Mistral uses the big American cloud providers for inference [0].
[0] https://trust.mistral.ai/subprocessors
Yes and they are open about it. That is why they are building their own data centers in Paris and other places.
That’s true, but as their model quality is behind Kimi K2.6 or Deepseek v4, I believe it makes more sense to consume those Chinese models from a European hosting company like Scaleway, than to consume Mistral’s models hosted on US cloud providers, except if you see that Mistral subscription as a kind of investment into Mistral for them to buy infra in the future. That is, if your goal is to “consume” AI on European soil.
Behind Deepseek R1 by far, they are at least 1 year late in model's performance.
It will be interesting if Mistral succeeds to keep up with US and Chinese labs in terms of Models, or if they just become an integration company of chinese OSS frontier labs, like more and more of their competitors
It seems that the winner takes all market for tech will eventually go away. Countries and regions want to develop their own good enough solutions that is not dependent on America.
I expect AI to become more like the news, where every country has their own news services, rather than the global monopoly Sam Altman had envisioned. We even see it with Meta, a company that doesn't sell AI. They'd rather build their own models than let one or a few companies have that much leverage over them. This is why it's unlikely open models will disappear. That's the only thing that prevents a global AI oligopoly.
Has nothing to do with geography and everything to do with marginal cost. Software and hardware is winner takes all because the marginal cost of producing a copy of a software product is near zero, and the marginal cost of producing a hardware chip is also very low. This is not true of AI models. Both inference and training have high marginal costs, since the compute demand is intense.
The word empire is used for anything these days. Mistral is successful, but their market cap $14B is less than than revenue of OpenAI or Anthropic. They may not have the scale to compete with the American and the Chinese companies.
$14B "empire" next to a trillion dollar OpenAI, trillion dollar Anthropic, trillion dollar DeepMind... this is a massive failure, not an empire. It is truly baffling how low expectations are for European tech.
Mistral brought Mixture of Experts and other really good innovations
> from Mistral’s offices in the trendy 10th arrondissement of Paris
Couldn't continue reading after this ignorance. The 10th is dominated by the two major train stations and warehouses. Notorious for petty crime and giving arriving tourists "Paris Syndrome" because of the disappointment. It is the least trendy arrondissement in Paris. It is central, but that's about it.
Edit: Looked it up and Mistral's offices are actually in the 2nd, about 500 meters from the Louvre. A very trendy area indeed. Is this a human or AI hallucination? What else in this article is wrong?
> in the trendy 10th arrondissement of Paris
> actually in the 2nd
The mixup is easily explainable: machines write numbers in binary notation!
If I could I would down vote this title, as it’s applying that only Americans have a right to be good
>“Poulantir” going public on the NYSE.
We don't have money, but we have the best puns.
-- The French, probably
That's... not that much money. Anthropic's supposedly worth $380 billion.
And on secondary markets shares are trading for at least twice that...
I hope Mistral will not fall for the Forbes kiss of death.
Any recent news or data on NSA running a MITM at DE-CIX?
Is it possible for the common people to access mistral? And that without being gated by a 'whatng cartel' web engine?
I use Le Chat free tier. Not as good as ChatGPT free tier but sufficient for 90% of my uses.
I tried Mistral Vibe (Claude Code equivalent) free tier for a week. I hit web_search limits but not limits for devstral-2. For a working mode where you watch it closely (same way I use Claude Code), it is fine.
Can I use Mistral Vibe without a whatng cartel web engine?
Namely, classic web (noscript/basic (x)html), or a dynamic web API with some public tokens? (using a set of CURL commands). If there is a requirement of account creation (often required to generate web API tokens), can I do that using the classic web (noscript/basic (x)html)? And if so, if the account creation does require an email address, do they support self-hosted SMTP servers without DNS, namely email addresses with IPv[46] literals?
"Here's a nickel, kid. Go buy yourself a real computer."
If you go out of your way to make your computing environment weird, don't be surprised when things don't work for you.
And yet there is heavy involvement of US capital and technology in Mistral. Not just hardware and finance, but also board seats.
or because it one of first companies to release an OSS model that somewhat useful
While that may be slightly better than being dependent on Washington, I still don't like it. Everyone is now AI-crazy. If you go to clown-channels such as the war industry or the crazy Palantir guy who wants to ruin everyone, you see that they dream of AI controlled mega-drones wiping out cities. That's the next "evolutionary" step; you can already see video footage of how much drones changed war. And these guys all push for more, more, more. I now understand why Trump keeps on making so many wrong decisions - he is being showed war-videos and sells it as a promotional agent. That's his main job; for anything else other than promo he has no competence.
Not being American is very important to me and my partner. For my next job, I'm looking exclusively at companies headquartered in the PRC. My partner and I formally registered ourselves as foreign agents of the PRC. While we did that, the NSA actually took down the entire DOJ filing site for this just to further obstruct us, in the end we had to register with the Attorney General by email, persuant to U.S. law.[1]
Of course, we don't think that China is perfect. But we have had nothing but abuse and interference from USG. You can read more about its OPSexr program here.[2] Typical quote:
"At other times, the conversations became explicit. The active source at the NSA claimed to have witnessed hundreds of sexually provocative discussions, which, he added, occurred mostly on taxpayer time. The former NSA source who was familiar with the chats recalled being “disgusted” by a particularly shocking thread discussing weekend “gangbangs.”"
This matches the experience my partner and I have every day, while our ordinary marital contact and spending time together is disrupted under bullshit pretexts.
[1] https://taonexus.com/publicfiles/apr2026/registered-agent.ht...
[2] https://www.city-journal.org/article/national-security-agenc...