So I am in Taiwan, mid-celebration the New Lunar Year, in a Taipei joint having a few with a startup fellow. Discussion inevitably slips towards the newly dropped DeepSeek model and its associated development costs, which, to say the least, opened the Americans eyes to the pre-conception that good tech can only emerge from Silicon Valley.

That convo triggered a topic idea - namely how much money running the local European AI hero Mistral actually takes and what its revenue side looks like. Is that a compelling case for investors further betting on the company?

I worked out some assumptions and numbers below.

 

 

---

 

Mistral's economics

EXPENSES

The breakdown:

1. Salaries - €17 million

Assuming they have 100 employees, and considering an average annual salary of ​€170,000​ (AI talent is expensive), the total would be €17 million.

2. Infrastructure costs - €15 million

Hard to estimate this one, particularly since Mistral has an equity for resources deal with Microsoft - let's just assume some €10-20 million in annual spent.

3. R&D - €5 million

That's on top of the infra for running the ops, required for the AI training, data acquisition and further prototyping of new stuff - costs that can go somewhere to €5 million a year.

4. Marketing - €3 million

This is basically customer acquisition, which should go higher in the short run, given that Mistral is yet to build a significant American presence.

5. Overheads - €4 million

10% on top of the total expenses.

Total expenses: €44 million

 

REVENUE

There's a few possible revenue streams, below a generic breakdown per a standard customer: 

1. Subscription to a customised model: $250k/year

It's highly dependable of the customer's size, likely ranging between $50k and $500k annually for large companies, depending on usage and customisation. 

2. Related customisations/implementation fees: $200k per year.

Subject to complexity, assume 30-60 days at $300/hour, this can add to some additional $200k per customer per year.

3. API-based access: $350k per year.

Assuming for a mid-sized customer a minimum usage of 1 billion tokens/month, $0.03 per 1000 tokens, we would get some $360k per year ($0.03*12B tokens). The larger the customer, the more tokens consumption and lower unit economics - fwiw, I know of a fairly large European company paying one cent per 1000 tokens for a seven figure yearly contract with OpenAI.

This price competitiveness should particularly reflect on Mistral's revenue line when approaching the American market, which is already dominated by local tech companies. 

TOTAL ARR per one customer: $800k/year

So how much does that translate in yearly numbers? Let's model it:

• 10 customers: $8 million ARR • 50 customers: $40 million ARR • 100 customers: $80 million ARR

Does Mistral have 100 paying customers? Not impossible but keep in mind that the US-presence is still minimal, at least from the intel that I have - so those customers should be mostly European, maybe some Asian, which business-wise have a lower dynamism than their American counterparts. Not to mention that the pool size is rather small, again compared to a sizeable North American market.

Speaking of which, Anthropic is said to be close to $1 billion in 2024 ARR (up from $200 million ARR) while OpenAI is at $4 billion turnover for 2024 (up from $1 billion in 2023). Competing with those guys requires aggressiveness and creativity - besides, building trust and market share in the USA while maintaining European leadership will stretch resources. Funnily enough, resources is not what the French lack.

And, of course, the revenue line can be approached in a more granular way, as one can segment a bit the customer pipelines. 

The ideal case is signing up large accounts such as Fortune 500 companies or governmental agencies, which can commend higher usage and customisation needs, ending up in $1M+ key accounts.

The market reality of it though is that Mistral is a second tier vendor in spite of having a decent product available in the market. 

That should reflect in the ability and implicit costs of producing sales from high paying customers. As such, it is likely that the French sales approach should be more creative aiming for different segments not as competitive, whereas i) the customers' LTVs are lower and ii)) the cost of doing business is at par with other competitors. 

Bottom line, if we're playing a bit this segmentation game, we get lower numbers than the above rough estimation:

• 10 big customers ($1M+) - $10 million
• 30 mid-sized customers ($500k) - $15 million
• 100 small customers ($100k) - $10 million

Total Revenue: $35 million

 

Other notes

• Putting the two together, it's safe to assume that Mistral is a $50 million business, which is no small feat for an European startup to pull off in 24 months of existence. 
• Mistral also notably raised almost $1 billion in capital - and while getting to this point is not easy, you can also argue the lack of not being enough aggressive with its growth. The capital has to be put to work not to sit in the bank, and that's one of Mistral's biggest challenges, publicly shared in the investors circle. 
• Do they have a product problem? I don't think so - definitely not the best, albeit the market says it is at par with the leading ones. Besides, general LLMs are commodities, tech improvements are marginal at this level already, and strategy-wise the push should be around finding moats acting as long-term competitive advantages - hence the R&D budget's importance.
• Do they have a marketing problem? While there's not too much intel other than anecdotes, I would imagine the French need much more exposure to non-bubble circles - AI market aggregators like ​these guys​ don't even include Mistral in their props. 
• Do they have a sales problem? I think they do, only from judging that they're at less than 100 million in sales backed by 1 billion in raised capital - as I said before, that money has to be put to work, since it has a higher cost than the cash in the bank, which at this level should be at 24-36 months of burn. 
• Not least, having enough resources and getting less than a 10% market share with a comparable product re-inforces the sales problem, in the context of the market leader Open AI still growing at double digits on yearly basis.

 

---

 

 

Did you find this useful? 🤔

Great    •    Good   •    Meh

 

---

 

Questions or comments? Let us know!