Cerebras plans to enhance its high-performance inference-as-a-service platform, with new datacenters in America, Canada, and France. While G42 remains a major customer, Cerebras will maintain control over its Oklahoma City and Montreal sites. The company's wafer-scale chips are designed to deliver significant performance advantages, achieving up to 125 petaFLOPS at FP16, and are positioned as a key differentiator in high-throughput inference for large models.
Key takeaways:
- Cerebras Systems resolved CFIUS concerns by amending its agreement with G42, limiting the UAE-based firm to non-voting shares, clearing the way for its planned IPO.
- G42 accounted for over 87% of Cerebras' revenues in the first half of 2024, but Cerebras aims to diversify its customer base with a high-performance inference-as-a-service platform.
- Cerebras plans to deploy over a thousand wafer-scale accelerators across six new datacenters in America, Canada, and France by the end of 2025, with most sites operated in partnership with G42.
- Cerebras' systems are designed to achieve up to 125 petaFLOPS at FP16, significantly outperforming conventional GPU-based providers in model serving speeds.