Etched Pulls 400+ Engineers From NVIDIA, TSMC & More to Build a New Frontier Inference Cluster For AI Which Is Already Worth $1B in Demand

Etched, a new startup with a team of 400+ engineers, has announced its latest AI solution, the Frontier Inference Clusters.

Etched Comes Out of Stealth To Unveil Its Frontier Inference Clusters, Achieves Successful A0 Tapeout & On-Track For $1B+ In AI Contracts

What happens when 400+ Engineers from leading firms such as NVIDIA, Google, Broadcom, TSMC, SK Hynix, and more come together? Well, Etched happens.

We're coming out of stealth.

We've built our first racks after a successful A0 tapeout, $1B+ in customer contracts, and $800m raised.

Early customer tests show us achieving SOTA throughput, latency, and power efficiency on inference workloads.

Our first racks ship this summer. pic.twitter.com/FLccrkLTza

— Etched (@Etched) June 30, 2026

This is a new AI startup who are co-designing chips, racks, software, and advanced manufacturing methods for frontier models. The company promises best-in-class throughput, latency, cost, and power efficiency for both prefill and decode workloads, and they have first silicon proof to back up their claims.

2 of 9

Today, Etched officially announced itself and the successful tapeout of its very first A0 silicon. The tapeout was actually done earlier this year on TSMC's N4P process technology, and they've since been busy validating their first rack-scale product, which has already racked in $1B+ in AI customer demand. The company has so far raised $800M across four unannounced financings, including a strategic investment from VentureTech Alliance, and is working to deepen & expand partnerships with leading semiconductor firms.

VLI Processor Pumps Out 80% Peak Flops At Half The Voltage of Existing AI Chips

So coming to the infrastructure, Etched lays out the plans to tackle frontier models including multi-trillion-parameter MoEs, long-context and agentic AI workloads. For this, the company had to work on a range of new chips, packages, PCBs, cold plates, interconnects, and more.

The first of these is a Low-Voltage Inference (LVI) for high-throughput workloads. This chip comes with a new architecture that allows it to do math at half the voltage of most AI chips. With this approach, Etched solves the sustained performance issues that occur in most AI chips that throttle down as chips draw more power at full-voltage mode, leading to under half the peak FLOPs.

CMS Accelerator Offers HBM/SRAM Combo

The second part is the Cluster Scale Memory (CSM) for low-latency workloads. We've seen the move towards massive SRAM blocks over HBM for faster decode speeds, but SRAM chips don't offer good FLOPs throughput or memory capacity. Etched's solution with CSM is a lower-latency and shared memory pool product that retains high-bandwidth interconnect for faster memory access. This HBM/SRAM hybrid tackles both memory capacity and memory latency while offering lower cost, higher reliability, better yields, and improved thermal characteristics.

All combined, Etched has been able to achieve state-of-the-art throughput, latency, and power efficiency in early customer tests across inference workloads. Etched's VLI processor can run trillion-parameter sparse MoEs at 80% peak FLOPs without thermal throttling.

Currently, Etched is scaling production at an unprecedented pace and has built a 2MW datacenter in its offices, along with the opening of a factory in Taiwan for 24/7 engineering. The company is promising more updates on performance and roadmaps this summer.

Follow Wccftech on Google to get more of our news coverage in your feeds.