Cerebras, AWS and new data centers
Cerebras is partnering with cloud and data‑center players — including an AWS collaboration for accelerated inference on Bedrock and projects with CoreWeave/BCE — as the company pushes wafer‑scale options into the cloud reported reported. That expands deployment models beyond traditional GPU farms.
Cerebras’s wafer‑scale WSE‑3 — the silicon at the heart of its CS‑3 systems — contains about 4 trillion transistors, roughly 900,000 AI cores, 44 GB of on‑chip SRAM and is rated at 125 petaFLOPS peak AI performance. (cerebras.ai) AWS described a “disaggregated inference” design that splits LLM work into Trainium‑handled prefill and CS‑3‑handled decode stages, connected over Elastic Fabric Adapter (EFA) networking to reduce end‑to‑end latency. (press.aboutamazon.com) The AWS press release promises inference “an order of magnitude faster” with the Trainium+CS‑3 pairing, (press.aboutamazon.com) while multiple outlets have published more specific early estimates — for example reporting roughly 5× token throughput gains in cloud inference comparisons. (winbuzzer.com) Bell (BCE) announced a 300‑megawatt AI data‑centre project near Regina, Saskatchewan, with construction starting this spring and the first stage expected online in the first half of 2027. (bce.ca) Public filings and reports allocate roughly 160 MW to Cerebras and 140 MW to CoreWeave at the Saskatchewan site, with CoreWeave’s compute slated to run on Nvidia GPUs — a mix that explicitly reserves large blocks of sovereign capacity for wafer‑scale and GPU‑based stacks. (ca.investing.com) AWS said Bedrock access to the Trainium+CS‑3 stack will roll out “in the coming months,” and noted that later this year it will offer leading open‑source foundation models and Amazon Nova on Cerebras hardware through Bedrock. (press.aboutamazon.com)
