Etched Unveils Frontier Inference Clusters With 80% Peak FLOPs at Half Voltage

AI data center operators seeking to reduce power costs and increase inference throughput now have a new hardware option from Etched. The startup has emerged from stealth with a complete rack-scale solution designed specifically for large language model inference. This matters because current AI chips often struggle to maintain peak performance without overheating or consuming excessive electricity. Etched claims its new architecture solves these thermal and efficiency bottlenecks for high-density workloads.

Startup ships first rack-scale AI hardware built on TSMC N4P silicon

The core of this system is the Low-Voltage Inference (LVI) processor, which Etched built from the ground up. The company is co-designing the silicon, the physical racks, and the accompanying software stack to ensure they work together seamlessly. This integrated approach targets the specific demands of frontier AI models that require massive computational resources. Etched has recruited over 400 engineers from major competitors like NVIDIA and TSMC to develop this platform.

Specifications

• Process Technology: TSMC N4P
• Processor Architecture: Low-Voltage Inference (LVI)
• Memory Architecture: Cluster Scale Memory (CSM) – HBM/SRAM hybrid
• Performance Metric: 80% peak FLOPs at half voltage
• Funding Raised: $800M

Etched manufactured its first A0 silicon using TSMC's N4P process technology. The hardware features a Cluster Scale Memory (CSM) architecture that combines HBM and SRAM into a shared pool. This design aims to provide lower latency and high-bandwidth interconnects for faster memory access. Early customer tests indicate the LVI processor can run trillion-parameter sparse Mixture-of-Experts models at 80 percent of peak FLOPs without thermal throttling.

The company reports achieving state-of-the-art throughput, latency, and power efficiency on inference workloads during these early tests. Etched has secured more than $1 billion in customer contracts and raised $800 million in funding to support production. The startup has already built its first racks following the successful tapeout of its initial silicon. These early results suggest the platform is ready for immediate deployment in commercial environments.

Etched will begin shipping its first rack-scale products globally this summer. The company has confirmed it is coming out of stealth mode with a fully functional hardware and software stack. This launch marks a significant entry into the AI infrastructure market with a focus on inference efficiency. Buyers can expect the first units to arrive in the coming months as the company scales its manufacturing.