Cerebras Systems rang the bell at NASDAQ on Friday and walked away with $5.55 billion in fresh capital, pricing 215.23 million shares at $185 apiece before closing the day at $311—a massive win for a company that was valued at just $23 billion back in February during its Series H funding round. The oversubscription hit 25X, which tells you everything you need to know about investor appetite for AI infrastructure plays right now. With all shares and warrants factored in, Cerebras commands roughly $95 billion in market cap on paper. CEO Andrew Feldman's 4.5% stake is worth $3.2 billion; CTO Sean Li's 2.4% stake comes in at $1.7 billion. Not bad for a company that spent years convincing the industry that building chips the size of dinner plates was actually a good idea.

The IPO Timing Wasn't Accidental

Let's be real: you couldn't pick a better window to go public. The hyperscalers—Amazon Web Services, Google Cloud, Microsoft Azure, and Meta Platforms—are collectively projecting $695 billion to $725 billion in capital expenses for 2026 alone. That's not chump change. That's strategic weapon territory, both economically and militarily. Cerebras had about $1.3 billion in cash before the IPO, plus another $1 billion in working capital from its massive $20 billion deal with OpenAI to deploy 750 megawatts of CS waferscale systems by 2028 (with an additional 3 gigawatts coming online in 2029 and 2030). Add the IPO proceeds and Cerebras is sitting on $8.9 billion in cash and equivalents—enough runway to build out those systems while actually innovating instead of just surviving.

The WSE-4 Problem: Going Vertical

Here's where things get interesting from a hardware perspective. The WSE-3 chips are hitting a wall—not because the compute is bad, but because the SRAM-to-compute ratio is fundamentally wrong for low-latency inference workloads. When Cerebras and Groq started duking it out over inference performance, both companies had to gang multiple machines together not for raw compute but purely to get enough SRAM in memory close to the processing cores. At first it was three CS-3 machines, then four, then Cerebras stopped disclosing numbers entirely when they published benchmarks. That's a tell. The core issue: you can't make 300mm wafers bigger (450mm efforts died over a decade ago), and transistors aren't getting dense fast enough to solve this problem in two dimensions. The math doesn't work. If you shrink the process and cut back on compute to jack up SRAM, you're looking at maybe 3X to 4X more memory on a flat wafer—which is technically brutal. But if AMD and Intel have taught us anything with their CPUs and GPUs, it's that going vertical with stacked memory actually works. Stacked SRAM on top of the WSE-4 base could solve this overnight.

What We Want to See in Hardware

Beyond stacking DRAM directly onto the wafer die, optical interconnects using co-packaged optics could be game-changing for Cerebras's architecture. Picture this: optical links running from the wafer to shared DRAM memory trays, dramatically expanding what they call MemoryX capacity while giving the memory its own network fabric—similar to how GPUs handle scale-up memory today. The same optical tech could boost SwarmX clustering bandwidth between WSE devices by orders of magnitude. We're also keeping an eye on potential deals with AWS (low-latency inference boxes complementing Trainium) and whether Anthropic inks something before their own IPO to show they have the iron for serious inference work.

Key Takeaways

  • Cerebras raised $5.55B in its IPO at $185/share, closing at $311—$95B fully diluted valuation
  • The company now has $8.9B cash on hand from a combination of existing funds and new capital
  • WSE-3 inference performance requires bundling multiple systems due to SRAM limitations
  • 450mm wafer scaling failed; 3D stacking is the only viable path forward for WSE-4
  • Optical interconnects could unlock next-gen MemoryX and SwarmX architectures

The Bottom Line

Cerebras just pulled off the kind of IPO that makes VCs look like geniuses, but the real test starts now. With Groq nipping at their heels on inference latency and Nvidia looming in every direction, Feldman needs to deliver a WSE-4 with stacked memory before someone else solves this problem cheaper. The money's there. The talent better be. If 3D wafer-scale doesn't ship by late 2027, this valuation story falls apart fast.