Startup Success Rates Are So Low Because the Rewards Are So Large
From Anneal’s post last year
If your org is shaped like a Y-combinator company, you can spend dozens of hours absorbing high-quality, expert-crafted content.… How likely is org building success, in this premier reference class
An AI safety lab is not the same as a Y-combinator company.
It is. WAY. FUCKING. HARDER.
The implication is that you should expect AI safety labs for have an even higher failure rate. Is that fair?
Consider:
In the stat Anneal links, “success” means becoming a “unicorn”, i.e. achieving a billion dollar valuation.
The standard YC deal is that they take 7% of your company for $125k.
7% of a billion dollar company is $70m.[1]
Now imagine that the success rate for YC was much higher, say 50%. In that case, they would be paying $125k to get $70m * 0.5 - $35m in expectation, or a 280x return.
What would you do if you had a slot machine that gave you 280x returns in expectation? You would put way more money into it!
Unfortunately the deal doesn’t last. As you increase the supply of capital seeking startups, two things happen:
Valuations go up, meaning you have to pay more for the same amount of equity
Company quality goes down since you have to lower the bar to fund more companies.
So returns diminish. But at 28x, this remains fantastic deal!
Here’s what I’m getting at: the success rate of startups is not some fixed quantity determined by “how difficult it is to build an org”, it is a dynamic rate set by an equilibrium. The success rate is as low as it can be while still delivering returns good enough for LPs to continue pumping money into YC.
Of course, that’s only have the equation. The other side is the willingness of startup founders to start companies. Similarly, imagine if half the time you started a startup, you ended up running a billion dollar company. It would be very popular! Overtime, that popularity drives more people to start companies who wouldn’t have otherwise, and the success rate drops. Again, nothing to do with the inherent difficulty, this is just things falling into equilibrium with potential founder’s next-best alternatives (working for someone else or retiring).
Some of what I’m saying here is just theoretical Econ 101 stuff that you shouldn’t take too literally. But it really is historically true that YC’s class size has grown 33x from 8 companies in the first batch to 260 companies in the latest batch. And that their standard deal has grown from $20k for 6% to $125k for 7%.
The mistake Anneal and many others make is to confuse a dynamic system for a static one. The low success rate of startups does not reflect the inherent difficult of building organizations, rather it reflects a willingness (from the standpoint of LPs who put money into companies and founders who start them) to throw more bodies and capital into this system until the expected value equilibrates with those people’s next best alternatives.
There are other considerations here, LPs might not like illiquidity, founders might not like risk, but at a very rough first approximation, this is the kind of dynamic you should have in mind when making inferences about analogous challenges.
- ^
I’m ignoring various considerations, including:
YC also puts in $375,000k on an MFN Safe (meaning that it converts at the best terms that you offer future investors)
YC’s investment will get diluted over time, very roughly maybe around 30% per round over 5 rounds between seed and becoming a unicorn, so down to .7^5 = 17% of their original ownership.
Some YC companies are going to be worth much much more than a single billion, so it doesn’t really make sense to use what is actually the floor of the population to calculate an expected value.
But these cut in opposite directions and don’t substantially impact the point of my post.
Somewhat relevant blog post by @NunoSempere: https://nunosempere.com/blog/2024/09/10/chance-your-startup-will-succeed/
https://www.benkuhn.net/outliers/
(Unrelated to the post: I’m glad you’re back, ADS. I wasn’t the only one who wondered)
This is true, but the reason success rate falls as more people found companies is that you’re drawing in founders creating marginally-less-good companies, who fail in part because creating successful high value companies is hard.
You can make it easier or harder by varying a lot of factors, which is why different cities and countries vary so widely in startup-companies-per-capita and unicorns-per-capita. Once you set those parameters, yes, I’d expect failure rates to fall into a stable equilibrium.
Also: some companies that are failures from the POV of startup investor needs and goals end up being perfectly fine small and medium size companies that can generate stable revenue but will never become unicorns.
Right yeah, that’s exactly it.
And we agree on that, yes.
I do agree that equilibrium dynamics mean that the failure rate is not fixed, and that this is underappreciated.
But, if the investors judge that the payoff of AI safety lab success would be above average, they should be willing to invest more and accept above average failures rates too. Failure rate increases with additional investment, and investment should increase until expected returns even out across options for further investment.
The harder a problem is, the higher the expected failure rate at any given rate of investment at any given point in time, and the investment won’t happen until and unless the expected return makes it pencil out.
Side note: As I understand it, the real equilibrium ROI for capital in general has been 4-5% on average for as far back as we have records. I find this suspicious, as I have seen no good studies on why. But my headcanon has been that this gives a doubling time about equal to the human generation time. AKA parents and societies save until they’ve saved enough to hand their wealth, undiluted, down to their children, on average.