How do you value startups creating foundational LLMs, or related generative AI technologies?
So far in 2024, according to Coatue, AI represents 3% of venture capital deals, and 15% of the capital. These rounds have been at roughly 5x the valuation of non-AI companies, with 6x the round size.
There’s heat, but how do we understand the extent to which it connects with value?
Cohere, for example, was generating around $35M at the time it raised a round at a $4.6B pre-money valuation. This scenario is a good example of why revenue multiples don’t often make sense in private markets, but – for the purpose of a rough comparison – how do investors justify a 131x revenue multiple?
If you look at trends in the recent funding of generative AI related companies (primarily foundation models) you can see roughly three milestones: rounds at ~$2B, a ~$5B, and then at ~$18B (OpenAI is in a category of its own, for the time being).
In this graph of post-money valuations by quarter, derived from Crunchbase data, we can see some consensus around the ~$2B valuation mark, and then a fairly clear convergence towards ~$5B.
While each of these startups fits in the foundation model theme, they all have different target markets, different product strategies, and different technology. We should be seeing a wide range of pricing, without obvious patterns in fundraising data.
It’s a curious pattern, and – if it continues to bear-out – has two clear implications:
- Investors haven’t yet developed a framework for valuing foundation model companies.
- Instead of trying to value these companies, most have become caught in the hype trap.
This means you will see a venture capital firm move out ahead and plant a flag: “we think, at this stage, a foundation model is worth $2B”. This provides a reference point, and a target, for all other foundation models that come after. In theory, this isn’t unreasonable for as long as the comparison makes sense.
However, we can see that foundation model startups with sufficient heat are speedrunning this process:
- Cohere took four years to get to their ~$2B Series C and a further year to hit the ~$5B valuation Series D.
- In contrast, Mistral hit their ~$2B Series A the same year they were founded, and their ~$5B Series B just 6 months later.
In this graph of post-money valuations by the age of the company, there’s a fairly clear picture of how rounds have inflated significantly. The starkest example is at Series B, where AI21 Labs, who raised a $664M Series B after five years, compared to Mistral who were incorporated in 2023 and raised a $6.67B Series B the next year. Two clear groups emerge in this representation: those that incorporated in the 2015-2019 period, before Generative AI hype started to influence valuations, and those created in 2021-2024 who immediately benefitted from the investor momentum.
The issue raised by this trend is that if value is disconnected from revenue for these companies (which it probably should be), and nobody fully understands the implications of the technology (though there are competing visions), why shouldn’t Mistral exceed the $4.05B bar set by Anthropic a year earlier? Who’s to say it isn’t more valuable now?
Indeed, as illustrated by the following chart, the inflation in valuations is basically correlated with the year a company was incorporated. Mistral (2023), the most recent entrant, has the fastest growth. AlphaSense (2008), the oldest, has the slowest. One upset to this is Anthropic (2021) outgrowing Cohere (2019) – though Anthropic has managed to position itself as a more direct competitor to market-leader OpenAI.
So what does this tell us?
The fundraising data for generative AI companies, and particularly those working on foundation models, indicates a significant amount of irrationality. Venture capital firms are chasing momentum into these deals, and letting momentum drive valuations up significantly, more and more quickly.
There’s a few reasons for this, mostly related to the underlying assets:
- The hardware required to train models may get radically cheaper in the near future, with companies like Extropic and Etched building new chip designs that are aimed specifically at reducing costs and driving efficiency in training LLMs – but they’re not here yet. Hardware remains a major cost-center, with recent examples including the OpenAI & Microsoft supercomputer project, costing $100B.
- The second major cost center is talent. There are only so many AI PhD researchers in the world, so the premium they command has risen dramatically with companies like Anthropic and OpenAI offering $700-900k in compensation for those roles.
If you look at the graph above which shows funding, by chronology of incorporation, it’s clear how much AI round sizes are inflating. This is the scale of capital you need to be competitive. And the majority of this will go to zero, so investors will concentrate further on the apparent winners.
These factors do mean that foundation model startups are having to raise ever greater amounts of capital, which will encourage higher valuations – but it doesn’t necessarily justify them.
Aligning Generative AI with Future Potential
Valuation is a reflection of future cash-generating potential, and valuations will have to be reconciled with financial performance when they head for any kind of exit. If you are betting on OpenAI capturing the majority of the market (given their list of partners, including both Microsoft and Apple) that may be reasonable. For the rest? Who knows.
There is a growing likelihood that we will be looking back on the valuations of these companies, a few years from now, much as we view the frothy-peaks of 2021. Investors got carried away with momentum, have burned a jaw-dropping amount of cash in the process, and the wins that emerge will be even further compressed into a handful of portfolios. There are already signs of this happening, with the Pitchbook / NVCA Venture Monitor report highlighting concerns about concentration:
“At a glance, Q2’s venture numbers look promising: With $55.6 billion invested, deal value rose to an eight-quarter high. But $14.6 billion of that capital was concentrated in just two companies—CoreWeave and xAI.”