Cerebras kicks off IPO roadshow, targets $115-$125

Cerebras Systems began its initial public offering roadshow on Monday, planning to price shares between $115 and $125. The company reported $510 million in revenue and $1.38 in earnings per share for the year ended Dec. 31. A year earlier, Cerebras posted a loss of $9.90 per share and $290.3 million in revenue. Morgan Stanley, Citigroup, Barclays and UBS are managing the planned sale. The company withdrew an earlier offering in October.

The filing comes amid a shift in the artificial intelligence industry from building larger models toward running models in production. Training models requires massive compute throughput, while running models for tasks such as chat or search, known as inference, requires faster memory access and different data handling.

Many customers split workloads across different processors. In a common setup, high-throughput GPUs handle the heavy computation phase called prefill, while chips with faster memory handle the decode stage that returns results quickly. Nvidia’s acquisition of Groq for $20 billion and efforts by cloud providers to pair custom processors with other accelerators reflect that approach.

Amazon Web Services combines its Trainium chips for prefill with Cerebras’ wafer-scale engines for decode tasks. Intel has outlined plans to mix GPUs with processors from SambaNova for similar purposes.

Cerebras designs a wafer-scale chip with large, fast on-chip SRAM and wide data paths. These features suit decode workloads where low latency and fast memory access matter. Smaller companies have focused on decode because SRAM stores less data but delivers quicker access, making multiple chips or a very large chip effective for low-latency inference.

Some firms disagree with splitting workloads across multiple accelerators. Tenstorrent unveiled its Galaxy Blackhole system and CEO Jim Keller criticized the industry trend: “Every company in the industry is pairing up to build the accelerator accelerator accelerator. CPUs run code. GPUs accelerate CPUs. TPUs accelerate GPUs. LPUs accelerate TPUs. And so on. This leads to complex solutions which are unlikely to be compatible with changes in AI models and uses. At Tenstorrent, we thought something more general and simpler would work.”

New hardware approaches have appeared. Lumai announced a hybrid photonic-electrical chip family and projected its Iris Tetra systems could deliver an exaOPS of inference performance while using about 10 kilowatts by 2029. The company plans to use those chips first for batch-processing tasks and later for prefill work.

Cerebras delayed a planned public offering last October and returned to the market after the company posted higher revenue and a swing to profitability in the most recent year.