Etched.ai
Etched designs and manufactures specialized AI inference hardware, including custom silicon and rack-scale clusters, optimized for high-throughput and low-latency workloads. Its A0 silicon and proprietary architecture enable running trillion-parameter models at high FLOPs utilization without thermal throttling, addressing the performance bottlenecks of general-purpose AI chips. The company is currently fulfilling over $1 billion in customer contracts for its first rack-scale product.
- Artificial Intelligence
- Hardware
- Semiconductor
Funding
Raised to date
$1.7BRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
across 6 rounds
+29$700M · Series D
Announced Aug 2026 · from article date
Disclosed cumulative: $1.6B
Founders
Product
Problem
General-purpose AI chips face fundamental performance trade-offs: they either throttle under sustained high-FLOPs workloads, reducing inference throughput, or rely on memory architectures that cannot achieve the low latency required for interactive applications. This forces AI companies to choose between raw compute density and responsiveness, limiting the efficiency and user experience of large-scale model deployment.
Solution
Etched builds a new category of AI hardware—frontier inference clusters—designed from the transistor up to overcome these bottlenecks. Its A0 silicon uses Low Voltage Inference (LVI) to run math blocks at under half the voltage of typical AI chips, enabling sustained throughput at 80%+ of peak FLOPs without thermal throttling. For latency-sensitive workloads, Etched's Cluster Scale Memory (CSM) combines HBM and SRAM in a hybrid design with a proprietary ultra-low-latency interconnect, creating a shared memory pool that delivers SRAM-level decode speeds without sacrificing capacity. The company co-designs its entire stack—from power delivery and advanced packaging to tiling and scheduling algorithms—in partnership with leading AI companies and cloud providers, and has validated its racks in representative data center deployments.
Target Audience
Primary customers are leading AI companies, cloud providers, and hyperscalers that need high-throughput, low-latency inference infrastructure for large-scale model deployment.
Features
- A0 silicon fabricated on TSMC N4P process, designed for high-throughput inference of trillion-parameter sparse MoE models
- Low Voltage Inference (LVI) architecture that runs math blocks at under half the voltage of conventional AI chips, achieving multiple times the FLOPs density without thermal throttling
- Cluster Scale Memory (CSM) with a proprietary ultra-low-latency, high-bandwidth interconnect for a shared memory pool across the scale-up domain
- HBM/SRAM hybrid memory design that balances capacity and latency, avoiding the cost, yield, and thermal trade-offs of SRAM-only or 3D DRAM chips
- Vertically integrated production including a Taiwan factory, data center, test house, and NPI prototyping lab for 24/7 engineering cycles