Apr 29, 2024 · 50m · news
Arthur Mensch: Open vs Closed - Who Wins and Mistral's Position | E1146 · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this 20VC episode, Harry Stebbings interviews Arthur Mensch, co-founder and CEO of Mistral AI, detailing how the European startup leverages algorithmic efficiency and an open-core strategy to successfully compete with multi-billion-dollar US tech giants. Mensch shares personal insights, organizational principles, and strategic viewpoints on the changing economics, bottlenecks, and geopolitical dynamics of the global AI landscape.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 23.1% of the talking time here. How this is scored →
Harry as informed peer 3.9 Guest teaching 3.3 Guest disagreement 1.8 Harry pushing back 3.2
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 34:23 Rejecting seed-round compute scaling critique
Arthur forcefully rejects Harry's suggestion that not scaling compute faster was an error, pointing out the first-principles reality that no company in 2023 could raise a $2 billion seed round.
Hardest push from Harry ▶ 34:23 Challenging compute scaling executionHarry directly challenges Arthur, citing investor feedback from Lightspeed and asking whether failing to scale compute earlier was a operational oversight.
Biggest teaching moment ▶ 22:23 Explaining algorithmic efficiency vs hardware cost dropsArthur corrects Harry's assumption that hardware price drops drive AI progress, educating him that algorithmic efficiency improvements provided a 100x gains boost over three years.
Harry holds his own ▶ 38:35 Exposing European VC reliance on US LPsHarry demonstrates inside domain expertise on venture capital structures, exposing that top European funds are largely capitalized by US institutional limited partners.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Episode Opening Sizzle Reel & Guest Montage | 2 | 2 | 1 | 2 | Harry asks personal background questions and gently probes Arthur on whether uncoupling teams creates silos or inefficiencies. Arthur explains the nuance of sharing infrastructure while avoiding meeting bloat. | |
| The Decision to Resign and Found Mistral | 2 | 3 | 1 | 1 | Harry asks why Mistral 7B achieved such rapid popularity among developers. Arthur educates on model compression slack and targeting the hardware capacity of Macbooks and gaming GPUs. | |
| Frontier of Efficiency: Mixtral and Compute Multipliers | 4 | 4 | 2 | 3 | Harry brings up Sarah Guo's question regarding whether scale ultimately trumps efficiency, pushing Arthur on marginal improvements. Arthur reframes compute as a tool to compress models and details compute multipliers. | |
| The End State of Models: Customization Over Commoditization | 5 | 4 | 2 | 3 | Harry leverages insights from Sam Altman and Brad Lightcap to ask about commoditization and quality limits. Arthur argues that raw models are starting points, with true defensibility coming from customization platforms and specialized data. | |
| Generalized Infrastructure vs. Vertically Integrated Applications | 5 | 4 | 2 | 5 | Harry presses directly on how Mistral captures value if domain-specific value accrues at the application layer. Arthur explains that specializing models requires tight integration with pre-training platform tools. | |
| Enabling Developer Freedom Through Open Source AI | 3 | 3 | 1 | 2 | Harry asks what AI developers actually care about beyond Twitter benchmark comparisons. Arthur explains developer priorities around cost, portability, fine-tuning, and enterprise data security. | |
| The Business Case for Branding, Trust, and Community | 5 | 4 | 3 | 4 | Harry asks a direct financial question about when marginal revenues will exceed marginal costs in LLMs. Arthur banteringly deflects by telling Harry that as an investor he should know, before analyzing margins across Nvidia, cloud providers, and app developers. | |
| The High Defensibility of Foundational AI Companies | 4 | 5 | 2 | 3 | Harry asks if rapid compute cost reductions lower barriers for new foundational model entrants. Arthur corrects the premise by showing algorithmic efficiency gains (100x over 3 years) far outpace hardware price drops. | |
| Hardware Dependencies and Strategic Partnerships | 6 | 3 | 2 | 5 | Harry bluntly asks whether big tech investments like Amazon in Anthropic are just round-tripped capital for compute. Arthur concedes that it looks like round-tripping while explaining the strategic necessity for both parties. | |
| Merging Science and Sales Cultures in AI Startups | 5 | 3 | 2 | 4 | Harry voices concern that incumbent enterprise distribution will crush startup model quality advantages. Arthur explains that open source releases act as a strategic shortcut to bypass traditional enterprise sales gatekeepers. | |
| The Practical Readiness of Open Source for Enterprise | 4 | 4 | 2 | 3 | Harry expresses skepticism about European enterprise adoption speed, noting many lack basic tools like Slack. Arthur counters that executive mandate for AI in Europe is strong and the lag behind the US is at most one year. | |
| Staying Relevant Against Better-Funded US Competitors | 6 | 5 | 3 | 6 | Harry brings up specific investor commentary regarding Mistral's lower capital raise and asks directly if failing to scale compute faster was a mistake. Arthur pushes back forcefully on first principles, noting nobody could raise $2 billion on a seed round in 2023. | |
| Governance, Founder Control, and Geopolitical Boundaries | 6 | 4 | 3 | 5 | Harry candidly claims the European VC ecosystem is a dirty secret largely funded by US institutions or underperforming government backers. Arthur defends the young European ecosystem, emphasizing that building generational venture ecosystems requires time. | |
| The Limits of European Growth Funds and Capital Supply | 5 | 4 | 3 | 5 | Harry strongly disagrees with Arthur's optimism regarding European growth capital supply, asserting new growth funds will not emerge soon. Arthur lightheartedly jokes that Harry is unusually pessimistic even compared to a Parisian. | |
| Scaling as CEO and Protecting Team Focus | 2 | 2 | 1 | 1 | Harry asks Arthur what advice he would give his past self before becoming CEO. Arthur reflects on staging product development before go-to-market and recognizing the need for fast US expansion. | |
| Continuous Fundraising vs. Commercial Revenue Streams | 4 | 3 | 2 | 4 | Harry opens the segment by asking directly if Arthur feels Mistral has enough cash. Arthur explains that frontier AI companies are perpetually fundraising because research investments intentionally outpace revenue growth. | |
| Quick Fire Round: Global Warming, Fatherhood, and AI Future | 2 | 2 | 1 | 1 | Harry asks if fears of AI job replacement are grossly exaggerated. Arthur argues jobs will be displaced and elevated to higher abstraction levels rather than simply eliminated. | |
| Envisioning Mistral's Success in Ten Years | 1 | 1 | 0 | 0 | Harry asks Arthur where Mistral will be in ten years. Arthur envisions relevant open and commercial models paired with an end-to-end developer platform before Harry wraps up. |
Statements from this episode (54)
Mensch: Mistral operates on 1.5k GPUs, a fraction of competitors' compute
“We are still bottlenecked by compute for sure, but that's because we don't have many of it. We have 1.5 k 800, which is a few percent of our competitors.”
Mensch: A team of five is faster than fifty unless uncoupled
“A team of five is faster than a team of 50. Except if you organize the team of 50 to be 10 teams of five that are sufficiently uncoupled.”
Mensch: Mistral AI operates with a team of just 25 people
“Although the team is only 25 people, so that's actually not super challenging.”
Mensch: DeepMind's initial Gemini development was too slow before recovering
“Gemini was a bit too slow, and I think they recovered sufficiently well since.”
Mensch: 7B parameter models run efficiently on smartphones and MacBooks
“And seven B is the size that allows to run Efficiently a model on your Macbook or on your smartphone.”
Mensch: Pre-Mistral 7B models lacked utility for real applications
“So there was already seven B models before, but they weren't good enough to do interesting applications.”
Mensch: Mistral targets model efficiency to reach top performance per compute cost
“And so that's why we continued of targeting very efficient models with the Mixtral HX seven B and more recently, Mixtral HX 22 B ensuring that for a certain costs and for a certain size, we were reaching the top performance of the market.”
Mensch: Compute scale alone hits a hard ceiling without high-quality data
“Scale isn't the only recipe, the only ingredients to the recipe you need to scale, but you also need to have proper data. Otherwise you reach some data quality limit.”
Mensch: Significant efficiency gains remain possible for given AI model sizes
“I believe there is. I believe we can make models that are much better for a certain size.”
Mensch: AI models will become merely starting points for developers
“And so I think that's the end state is models are effectively going to be A starting point for any AI application developer. They need to be surrounded by tools, by lifecycle management platform basically, and that's the one thing that we started to build.”
Mensch: AI application differentiation comes from data and user feedback
“Like general purpose models are a bit undifferentiated, but the differentiation that you need to create for your application comes from the data you put into it, the user feedback that you gather and the intelligence that you have to figure out what the applic…”
Mensch: Compute is no longer the primary bottleneck for text models
“There's obviously compute, but given the amount of data you have we have at hand compute is already running into is no longer the bottleneck. The bottleneck is more the data at that point. You should look at text to text models.”
Mensch: Vertical AI models will be built by application makers
“And actually these vertical models are not going to be out there. They're going to be built by the application makers because the only way you can Make a low latency model that is super good at a specific task is to get rid of the general purpose aspect becaus…”
Mensch: Mistral focus is enabling developers to easily fine-tune models
“It's a very hard job to make a specialized model. So it's actually very tied to the way you create a pre-trained model. And so bringing the tools that allow to do it in a foolproof way. So allowing developers to create a customized model that are performing ve…”
Mensch: The price per unit of AI intelligence will definitely decrease
“The price around the model, the dollar per intelligence unit, let's say, is definitely going to reduce.”
Mensch: Developer freedom is the best path to ubiquitous AI
“Bringing freedom to developers and AI application makers is I think the best way in, in distributing generative AI as widely as possible, which is our objective as a company, making AI ubiquitous bringing frontier AI into everyone's head.”
Mensch: Developers prefer customizing open-source models over proprietary APIs
“This open source part was, I believe, a good enabler for the community and made people realize that they could build very interesting technology by modifying the models themselves instead of depending on the APIs of a couple of providers.”
Mensch: Current AI fine-tuning approaches are too low-level
“Like the fine tuning aspect that has been like the go-to solution is probably a little too low level from What we should be doing.”
Mensch: Developers rely on community vouching because evaluating every model is impossible
“People use certain models because they are known to be good. You can't afford to evaluate everything out there. And so having some form of community vouching is super important.”
Mensch: Open-source distribution creates critical trust and brand equity in AI
“Brand is important because trust is important in that domain and open source brings trust in terms that Provide some trusted brand.”
Mensch: Nvidia captures top AI margins while LLM margins lag traditional software
“NVIDIA is, at that point. The cloud providers are pretty much at cost LLM providers, we're not at cost hopefully, but the margin that, Are known to be lower than the typical software margins.”
Mensch: Foundational AI model margins will not drop to zero
“I don't think there's any way in which the marginal cost and the margin of the most important part of that technology, which is really the foundational layer becomes zero because otherwise there's definitely going to be I guess fairness problem.”
Mensch: Foundational AI model market retains high defensibility
“There's a few barriers that are pretty hard to face that you need to you need to accrue sufficient, well, to raise sufficient capital to have enough compute and be relevant. You need to have people that knows how to train models, which is still a scarce resour…”
Mensch: NVIDIA roadmap delivers 30% compute cost reduction every two years
“The cost of compute reduces over time, just based on hardware costs. It reduces around 30% every two years if you follow NVIDIA roadmap. For the same amount of flops.”
Mensch: AI model training achieved 100x algorithmic improvement over three years
“So if you look at the way we train models from three years ago, and the way we train model today, I think we have probably made something around a hundred times algorithmic improvement. So that's the, that's probably where most of the gains were actually made …”
Mensch: Cloud investments in AI startups resemble round-tripping
“It looks like around flipping. Yes. I don't know about that deal particularly, but it makes sense from both perspectives.”
Mensch: Open-source AI shifts market value to platform and customization layers
“It moves the value a little higher than the model itself. It moves the value to the platform and customization part which is really I guess something that we're expecting and it's, it accelerates that process.”
Mensch: Mistral will remain an open-source leader while monetizing commercial models
“Like we still intend to be a leader in the open source part and to have some unique assets that we can license. And to have some unique platform that developers can use.”
Mensch: Exposing AI researchers to business teams improves model performance
“Ensuring that the science team has some relatively direct exposure to the product and to the business team is actually important to make them understand what, where the model is failing and how it could be improved significantly.”
Mensch: Open-source models provide a shortcut to enterprise software distribution
“A shortcut to distribution is to create demand through open source models.”
Mensch: Lack of developer tooling limits enterprise open-source AI adoption
“I think they're still lacking some products around like managing correctly load balancing customizing the models because you can do it with DIY solutions, but if you want to make it robust enough and scalable enough, it's actually not easy. And if you want to …”
Mensch: Everyone will adopt core-business generative AI in five years
“So to be not thinking about generative AI as a way to As a way of increasing productivity in a world processing, but rather as a way to change completely the way you operate your core business, which usually involve taking models and customizing them pretty he…”
Mensch: European enterprise AI adoption lags US by maximum one year
“So there's some delay compared to the U S market for sure. But it's not, I wouldn't say it's very significant It's one year maximum in terms of delay.”
Mensch: AI spending in customer support has moved to core enterprise budgets
“It's moving into core budget for customer support, for instance, where like areas where the application of AI is pretty obvious. It's definitely moving into core budgets.”
Mensch: Telecom and healthcare AI spending will move to core budgets within a year
“It's also at the experimental stage in my many other functions and for Core applications in the industry in, I guess, telecom the telecom industry and in healthcare, this is still In the playground, but I think it's going to evolve in the next year.”
Mensch: AI model quality is correlated with compute, not dependent on it
“Capital is equal compute. Then compute is correlated with quality. It's not completely dependent on it.”
Mensch: Mistral does not need to scale compute at competitor rates
“So, I mean, we're growing our compute like every companies. We are convinced that we don't need to grow at the same rate because there's a lot of barriers that are not compute related that are appearing on the way and that we're already seeing”
Mensch: Physical limits on hiring and infrastructure cap startup scaling speed
“You can only hire that fast. You can only scale your infrastructure to manage more GPUs that fast, and you can only raise capital that fast. So there's some acceleration constraints that are pretty hard to fight, and that are pretty much the first principles o…”
Mensch: Early-stage companies must stay under founder control to preserve vision
“What is important for a young company like us is to be under, under the control of the founders because there's a lot of things to be invented and the vision it can only be carried by them by us.”
Mensch: Startups cannot operate in both the US and China simultaneously
“We don't operate in China because you can't really operate in US and China without being like a very, very large corporation.”
Mensch: European 23-year-olds match Silicon Valley engineers in four months
“On the talent side we can hire 23, 24 years old people that we can onboard in four months and they operate as well as any software engineer in the Valley.”
Stebbings: Top European VC funds are backed primarily by US institutions
“Honestly, in large part, yes. There's government institutions which are backfilling it, but largely backfilling it with bad players who aren't very good. But the best providers in Europe are largely US funded by top US institutions.”
Mensch: Silicon Valley leads in senior AI talent while Europe excels in junior talent
“For senior AI scientists you find them more in the Valley than in France for junior AI scientists. It's there's a wealth of talent in France, in Poland, in the UK And that's that I think one of the strengths of the area.”
Mensch: European VCs could not structure Mistral's large pre-revenue Series A
“European funds were unstructured to do the kind of deal that we were proposing. So we didn't even have a lot of conversation cause they just couldn't get their head around the investment that needed to be made as well as we were a pre-revenue company.”
Stebbings: Europe will not build new growth funds within 3-5 years
“That is not going to happen. We are not going to see many more European growth funds be built in the next few years for sure. Not in the next three to five.”
Mensch admits he is not scaling properly as CEO of Mistral
“I'm not, I don't think we're, I don't think I'm doing it properly. But we are actively trying to find sources of information to learn new things, let's say.”
Mensch: Mistral started go-to-market when it had nothing to sell
“We did start the go to market motion at the time where we had absolutely nothing to sell.”
Mensch: Mistral could not operate solely from Europe without US presence
“We could not operate only from Europe, and that we needed to go to the US very quickly.”
Mensch: Frontier AI investments will exceed revenue for years to come
“For the years to come the investment are going to exceed the revenue by design, because you do need to scale and you do need to stay relevant on the as the frontiers company.”
Mensch: AI research speed will outpace go-to-market development for years
“But I think today and for the years to come, the speed. For developing research should be faster than the speed at which you can develop your go to market.”
Mensch: Mistral was surprised by Cohere's recent strong model releases
“We were surprised by cohere recently. They did came up with new good models. And I think that was a surprise for us.”
Mensch: Founders should not start new foundational AI model companies today
“Is it too late? I wouldn't recommend going into the foundational layer business. I know Sam didn't recommend to do that one year ago, and we did, and it seemed to have so far changed a few things. So I wouldn't be I think it would be arrogant for me to say tha…”
Mensch: Fears of AI job replacement are grossly over-exaggerated
“I think they are. I mean, it depends on who you're speaking to. I think job are going to be displaced for sure. Some will be replaced. Some will open up.”
Mensch: AI abstraction speed is unmatched in human history
“I think what's happening right now is that probably the speed in, in our elevation toward higher abstraction level is probably occurring at an unmatched rate in history. Though that means that the society adaptation is going to be more challenging.”