Cerebras Systems – Perfect Timing

Cerebras times the market perfectly. 

• Such is the demand for anything AI, that, despite some serious question marks, Cerebras Systems has been able to price its IPO above the already increased range and still have a good prospect of a rally when the shares open for trading.
• Cerebras Systems is a designer of very large AI chips, where the idea is that having all the transistors on the same piece of silicon allows inference to run both much faster and cost-effectively.
• The problem with making larger chips is that the probability of a defect increases exponentially, but this is where Cerebras has its core IP, as it has worked out how to work around defects and still run optimally.
• This makes Cerebras a good choice for anyone who is running work loads that need very fast execution and for those who don’t care about third-party developers or industry standards.
• This is because Cerebras does not work with Nvidia’s CUDA platform, meaning that it is optimal for those who are creating an end-to-end, in-house AI system.
• Hence, I see this is applicable to defence, intelligence and sovereign AI applications, and to be fair to Cerebras, there is plenty here to keep it busy for a while.
• However, the company is asking for a valuation of 80x 2025 EV/Sales and 420x 2025 PER, which is pretty heady by any stretch of the imagination.
• Against this, it is growing very quickly, and so I would expect these numbers to come down materially when the estimates for 2026 become more widely known.
• Demand for AI equipment suppliers is white hot at the moment, and with capex from the hyperscalers and everyone else still growing quickly, this looks set to continue into 2027 at least.
• Cerebras has two major risk factors:
  • First, customer concentration: where as of December 2025, one customer (MBZUAI (see here)), accounted for 77.9% of outstanding invoices.
  • MBZUAI is effectively the AI development lab of the UAE, which develops models more oriented to the Arabic language and the global south.
  • It has recently released K2 Think, which has 70bn parameters and performs reasonably well on benchmarks and my own non-scientific testing.
  • Cerebras has recently signed a 750MW deal with OpenAI (see here), which could translate into revenues of around $20bn and go a very long way to reducing the single client risk that Cerebras has long had.
  • It also gives Cerebras a mark of credibility that it can now use to win other clients and diversify the client base further.
  • However, as OpenAI’s deal with Nvidia shows, these deals can be very flimsy, and until revenue starts showing up in the income statement, I am a little concerned that it won’t amount to anything.
  • Either way, Cerebras can still use OpenAI as a logo to win other business, which is what it will now need to do once the IPO is done and dusted.
  • Second, Nvidia obsolescence: which is by far the largest concern.
  • Now that Nvidia designs and sells the entire AI system from chip to complete data centre design, it claims that it can network many thousands of GPUs together so that they operate as one in a similar way that Cerebras does.
  • This is unlikely to be as effective as or as fast as the Cerebras design, but the fact that Nvidia is 2 more generations ahead of Cerebras could make a large difference.
  • Furthermore, it has just acquired Groq to address the fast inference market, and its CUDA platform is very popular with developers.
  • Hence, I think that Cerebras occupies a niche in the AI inference market, but Nvidia is expanding its reach to cover part of that niche, meaning that when supply and demand are better balanced, competition is likely to heat up.
• The net result is that for the moment, Cerebras is likely to sell all it can make as demand for AI silicon remains much greater than supply due to the race to develop superintelligent machines.
• This race is likely to continue into 2027 at least, meaning that sentiment towards the sector should remain strong, giving Cerebras plenty of demand for its silicon.
• I would still prefer the memory sector, given that it is the biggest constraint and that its players have the lowest valuations, but I think for now, Cerebras will see its share price increase from the IPO price.