Cerebras Systems, Inc. "A Fast Inference Pure Play With Backlog Building;..."

研报英文原文证据摘录

Cerebras Systems, Inc. "A Fast Inference Pure Play With Backlog Building;..."

Cerebras Systems, Inc. UBS Research

UBS Research THESIS MAP Thesisa guideMapto our thinking and what´s where in this report

Pivotal Questions Q: Does CBRS' technology offer a durable differentiation within the AI inference market?

Yes. CBRS targets ultra-fast inference, where SRAM tightly coupled to compute drives much higher

token speeds. HBM systems hit a “memory wall” (~1K tokens/sec). CBRS’ wafer-scale design avoids

inter-chip latency, enabling superior performance vs. smaller SRAM players (e.g., Groq). This

approach is hard to replicate given yield, interconnect, and thermal challenges. That said, we expect

ultra-fast inference to be a distinct but niche segment as it excels in single-model, low-concurrency

use cases (premium tier), while GPUs remains the scalable, multi-tenant workhorse.

Q: Do we see meaningful upside to the financial model with the new potential partnership?

Yes. An AWS–CBRS disaggregated inference solution (Trainium for prefill, WSE for decode) mirrors

NVDA’s GPU+Groq solution that targets faster than GPU inferencing at lower than SRAM-only

system cost - and we think will be increasingly more popular in the industry. If scaled, we see

meaningful upside: potential WSE deployments alongside new Trainium racks (est. ~30K in CY27)

and even legacy systems. While we currently model ~$2B of hardware revenues in CY29, these

deployments could translate into a $4-10B+ annual hardware opportunity. Longer term, MSFT could

also adopt similar architectures as MAIA scales.

Q: Does Cerebras cloud + hardware business model drive a meaningfully different

profitability profile vs hardware peers and warrant a different valuation approach?

No.