AI developer tools · What shipped

Cerebras announced CS-4, its fourth-generation AI accelerator system, delivering up to 30x faster inference than GPU systems in production. The system uses three Wafer Scale Engine 3 Turbo processors and achieves over 1,000 tokens per second on models exceeding 10 trillion parameters through 2-microsecond wafer-to-wafer communication latency. CS-4 delivers 10x more token throughput per watt than its predecessor (CS-3) and supports disaggregated inference, allowing flexible pairing with complementary prefill platforms like AMD Helios and AWS Trainium. The new Nexus rack-scale platform features modular design with 50% fewer components, enabling faster manufacturing and deployment from days to hours, plus high-density power delivery and doubled I/O bandwidth. Shipments begin this quarter.

AMD has achieved a 4x increase in AI energy efficiency from 2024 to mid-2026, surpassing its projected 3x interim target while progressing toward a 20x rack-scale improvement goal by 2030. The company measures progress using performance-per-watt metrics across representative AMD configurations, with 2026 results combining measured product data and modeled estimates. AMD's approach targets full-stack optimization across compute silicon, process technology, memory, interconnect bandwidth, software, and system design rather than component-level improvements alone. The company projects that by 2030, approximately two future racks could deliver the same compute capability as 570 racks from 2024, reducing electricity consumption and carbon intensity by 20x and 28x respectively, or alternatively enable 20x more compute with the same energy use. High-bandwidth memory, larger caches, memory-to-compute integration, and high-speed interconnects like those in AMD's 72-GPU MI455X Helios rack are central to the efficiency gains. AMD's ROCm software stack and algorithmic advances are expected to compound hardware gains, potentially supporting up to 100x overall improvement in AI energy efficiency by 2030.