Disclaimer: This is not financial advice. If you make investment decisions based on the advice of some dude with a blog, you may lose money.
On Monday, Cerebras filed their S-1, an SEC filing indicating their plan to go public. This is a big deal; Cerebras would be the first of the big AI chip startups to go public. Other early players in the AI chip space failed or got acquired, so this could serve as a major proof point that AI chip startups could become legitimate companies. On the other hand, this could just be a hail-mary way to raise additional capital. Cerebras already raised $700M total, with their most recent round being a Series F. Venture capitalists may just not be willing to give them more money at a reasonable valuation.
In an S-1, a company has to disclose a lot of financial information that startups can keep secret. So we now know how much revenue Cerebras is generating, how much money they’re spending, and what their major business risks are. Off the bat, their financials don’t look all that bad. In the last 6 months, they generated $130M in revenue, with about $100M of that coming from hardware sales. Hardware is expensive to build, though, so they only made a gross profit of about $56M. This isn’t enough to cover their $97M in expenses, so they ran a pretty significant operating loss of $41M. With only $90M in cash, this means that they don’t have too much time to continue operating at this steep of a loss without the capital injection provided by this IPO.
From a growth perspective, though, Cerebras is doing very well. Their revenue from the first 6 months of 2024 is over 10x what it was during the first 6 months of 2023. If they keep up that sort of explosive growth in hardware sales, they could quickly become a major player in the AI hardware market and maybe even start turning a profit. From a pure numbers standpoint, I think the S-1 actually looks pretty decent.
So, that answers it, right? Cerebras is going public because they want to raise additional capital to continue operations and maintain their high growth rate until they can start turning a profit. But that still leaves an open question: given that AI chips are so hot right now, why doesn’t Cerebras just raise more venture capital? Being a public company is difficult and comes with all sorts of regulatory and logistical requirements. And there’s clearly still VC interest in late-stage AI chip companies: Groq just raised $640M to build their new chips. But if you look at Cerebras’s S-1 a bit more closely, you can spot the issue that might have scared off another round of venture investment: G42.
Cerebras has massive concentration risk in the UAE.
Group 42 Holdings, Ltd, also known as G42, is a holding company based in Abu Dhabi that does various AI-related things. It’s run by an Emerati prince, it peddles repressive spyware, and it has concerning ties to China. It also accounts for almost 90% of Cerebras’ revenue:
Group 42 Holding Ltd (together with its affiliates, “G42”) accounted for 83% and 87%, respectively, of our total revenue for the year ended December 31, 2023 and six months ended June 30, 2024.
In general, making most of your sales to one single customer is a significant business risk. If G42 and Cerebras have a falling out, that would kill over 80% of their revenue overnight. But this risk is amplified by the fact that G42 is headquartered in the UAE. As AI infrastructure becomes more tightly regulated, Cerebras’ ability to sell to the UAE might start being limited.
The UAE falls into Country Groups D:3 and D:4 in the US Export Control country groups, which means that certain advanced computing technologies require additional licensing to be exported to the UAE. Thus far, this hasn’t hurt Cerebras, but if regulations tighten or if the relationship between the UAE and the US sours, their main source of revenue could be at risk of drying up.
However, If Cerebras lands other large customers, they could eliminate this concentration risk and be a really promising investment. If G42 is buying tons of Cerebras chips, they should be able to make some other big sales too, right? Well, there’s another issue here: G42 has a significant vested interest in Cerebras.
G42 is not a representative customer.
G42 isn’t necessarily buying Cerebras systems because they’re the best systems on the market. Cerebras’ chips are admittedly extremely powerful, but G42 is also an investor in Cerebras. They currently own a relatively small amount of stock in Cerebras -- about 1.5M shares -- but if you look carefully in the S-1, there are a number of agreements between Cerebras and G42 that let G42 buy more Cerebras stock.
G42 has already agreed to buy 22 million non-voting shares of Cerebras for $335M. This would make them the largest stockholder in the company. Plus, if G42 buys more Cerebras hardware, G42 will be able to buy additional Cerebras stock at a 17.5% discount compared to the 30-day average trading price. Essentially, this means that G42 is incentivized to buy Cerebras hardware over other AI chips. Not only do they get stock purchase discounts in return for buying Cerebras systems, but their purchases add to Cerebras’ revenue, and push the stock price up.
This means that G42’s massive purchase orders from Cerebras aren’t evidence of a superior product. Even if Cerebras chips slightly underperform Nvidia in terms of price-to-performance, power efficiency, latency, or some other key metric, G42 would still buy chips from Cerebras.
Ultimately, what Cerebras needs to be an appealing investment is customer traction outside of G42, which they don’t seem to have. Their wafer-scale-chip strategy is based around building ultra-large, ultra-expensive chips for high-end customers -- but it doesn’t seem like the high-end customers are biting. This brings me to what I believe is Cerebras’s biggest risk: that the only customers with the budget to buy their wafer-scale chips are probably just going to build chips in-house instead.
Hyperscalers are building chips in-house.
The biggest AI cloud providers out there aren’t jumping to get access to Cerebras wafer scale engines, Groq LPUs, or SambaNova RDUs. They’re cutting out the middleman and building their own chips in-house. Google famously has their TPUs, Microsoft is developing their Azure Maia chips, and Amazon has the cleverly-named sisters of Trainium and Inferentia. While these cloud providers also offer Nvidia chips, they only do so because of customer demand. I’m sure that if it was up to AWS, GCP, and Azure, they would get rid of GPUs entirely and only offer their in-house AI chips. The last thing they want is another silicon vendor trying to sell them expensive systems.
But at the same time, Cerebras systems are so expensive that few companies other than large cloud providers like Google, Amazon, and Microsoft would consider buying them. Each system is $2-3M, so building a cluster of Cerebras systems gets eye-wateringly expensive very quickly. There are very few groups who can actually afford that price tag. As G42 shows us, it turns out that one of the few ways to afford Cerebras hardware is by running an oppressive monarchy monopolizing oil and gas reserves in the Middle East.