The past seven decades of microchip production have been a race to get smaller. Cerebras Systems, which hopes to break a drought in initial public offerings, is instead going big. Its Frisbee-sized chips are, it says, as fast as it gets at handling super-complex artificial intelligence models. In that sense, size may be on Cerebras’ side. In another, the opposite is true.
Cerebras hopes its jumbo wafers can take on industry champ Nvidia. The idea is that by plonking more memory and processing power on a larger area, data can be moved, stored and crunched faster and with less power consumption. The company’s tests suggest it enables Meta Platforms’ Llama 3.1 model to spit out answers about 20 times faster than rivals.
It is not just its speed that is unusual. The lossmaking firm gets almost all of its revenue from a single customer, Abu Dhabi’s G42. Large prepayments from that and other customers make up the lion’s share of its cash. If G42 makes a sufficiently big order in future it gets more shares at a discounted price, diluting IPO investors.
A further risk could come from irritating a colossus. Nvidia is the go-to for AI, and engineers are used to using its own proprietary programming language. Cerebras warns in its filing that large rivals could pressure their customers to give it the cold shoulder. They might not need to: anyone splurging vast sums on building bots may favour a supplier soundly tested in the field. Switching isn’t simple: using Cerebras chips means also using its other hardware and cooling systems.
Nvidia isn’t the only giant with sway over Cerebras’s future. Both firms’ chips are manufactured by the ubiquitous Taiwan Semiconductor Manufacturing Company. But where Nvidia is a big influential customer of TSMC, Cerebras is a tiny one. Semiconductor supply chains are long and brittle. The company has been stuck with unsold inventory before, and warns it may again.
Being a flea on the ankle of a giant could pay off: a little more business would go a long way. Assume Cerebras can keep up its recent habit of doubling revenue each six months, and it would be in line for $400 million or so this year. Put that on Nvidia’s 25-times multiple, and it is worth $10 billion.
Given the paucity of tech IPOs and the heat around anything AI, there is every chance investors might entertain that kind of valuation. But given the unusual risks, Cerebras would be best keeping its price ambitions smaller than its outsize chips.
john.foley@ft.com