Etched announced that it has come out of stealth after building its first racks following a successful A0 tapeout. The company also disclosed that it has raised $800 million and secured more than $1 billion in customer contracts.
Etched said early customer tests show its inference systems achieving state-of-the-art throughput, latency, and power efficiency on inference workloads. The company’s first racks are expected to ship this summer.
Etched has built a team of more than 400 engineers from companies including NVIDIA, Google TPUs, Broadcom, SK Hynix, and TSMC. The company is backed by Jane Street, HRT, Two Sigma, and Jump, with strategic investment from VentureTech Alliance.
The company’s Series B round was led by Stripes, with participation from Ribbit Capital, Radical Ventures, Positive Sum, Primary, and Argo.
Etched said its inference systems are designed to improve performance across frontier models, including many-trillion-parameter mixture-of-experts models, long-context workloads, and agentic workloads.
As part of its launch, Etched highlighted two technical breakthroughs: Low-Voltage Inference and Cluster-Scale Memory.
Low-Voltage Inference is designed for high-throughput workloads. Etched said current AI chips often struggle to scale FLOPs without thermal throttling because higher FLOPs utilization increases power draw and causes chips to reduce clock speed. According to the company, this can result in sustained inference throughput that is less than half of peak FLOPs.
Etched said its new architecture runs the chip’s math blocks at less than half the voltage of most AI chips, enabling multiple times the FLOPs density of current AI chips.
Cluster-Scale Memory is designed for low-latency workloads. Etched said today’s AI chips using high-bandwidth memory cannot achieve SRAM-level decode speeds because of memory subsystem and interconnect bottlenecks, while SRAM-only chips have lower FLOPs density and memory capacity.
The company said its architecture creates a shared low-latency memory pool across the entire scale-up domain. Etched uses a proprietary ultra-low-latency, high-bandwidth interconnect to enable faster memory access across chips.
Etched’s HBM/SRAM hybrid design is intended to address both memory capacity and memory-to-memory latency, enabling high throughput and interactivity at the same time. The company said Cluster-Scale Memory improves latency while avoiding the cost, reliability, yield, thermal, and compute tradeoffs associated with SRAM-only chips, 3D DRAM chips, or optics.
Etched said it is scaling production as quickly as possible. The company has built a 2-megawatt data center in its office and opened a Taiwan factory for 24/7 engineering. Etched expects to share additional updates about performance and its roadmap this summer.
