Etched AI chip startup has secured $300 million in Series C funding at a $10 billion pre-money valuation, backed by a massive $1 billion pre-order book that puts its low-voltage inference hardware on a direct collision course with Nvidia Blackwell systems this summer.
The San Jose-based company is moving its vertically integrated rack infrastructure beyond lab demonstrations to commercial deployment. The AI hardware funding round was led by Sequoia, with A16Z, Jane Street, SK Hynix, Diffusion Capital, and existing investors participating. President Robert Wachen noted that the buyer base consists of the largest AI companies and AI clusters, driven by demand for coding, long-context agents, long-horizon agents, and other inference-token workloads. That demand is not yet deployed customer capacity, making the upcoming summer shipments a critical measurement test rather than just another fundraising milestone.
Low-Voltage Inference Architecture
The core of the Etched AI chip strategy relies on low-voltage inference, or LVI. The company runs the transistors in its math engines at under half the voltage used by other AI chips. This allows the hardware to avoid thermal throttling and sustain 80% or higher utilization for trillion-parameter mixture-of-experts (MoE) models, diffusion models, and state-space models.
Wachen noted that current AI chip efficiency sits at a dismal 0.2 to 0.4 FLOPS delivered for every FLOPS purchased. By applying low-voltage inference broadly, Etched projects it could double or triple inference capacity. However, SRAM remains outside the low-voltage regime because standard SRAM cells cannot handle sub-threshold operation today.
The earlier product pitch for Etched was narrower: chips that hard-wire the transformer architecture directly into silicon, aiming for token throughput an order of magnitude above Nvidia Blackwell-based systems. The newer product case covers a broader range of inference patterns. The design now targets higher FLOPS and bandwidth specifically for multi-trillion-parameter MoE models, diffusion models, and state-space models. This evolution reflects the rapid shift in AI architectures away from standard transformers toward more complex, specialized inference workloads.
Supply Chain and Technical Hurdles
Scaling low-voltage operation is notoriously difficult. Historically, low-voltage operation has been restricted to small chips or crypto-mining ASICs due to high current, current spikes, and slower clock speeds. Etched is tackling this physical challenge across multiple layers, spanning fabrication, ASIC design, custom packaging, board design, cooling, and mechanical engineering.
The chip is built on TSMC N4P using a full-reticle-sized design with six stacks of HBM. This manufacturing path deliberately separates the Etched supply chain from Nvidia’s next-generation Rubin GPUs, giving the startup an independent hardware pipeline.
Cluster memory represents the other critical system-level piece of the Etched architecture. The company uses a low-latency shared memory pool across a whole cluster in a single scale-up domain. This is paired with a custom high-bandwidth interconnect designed for rapid access to SRAM and HBM across multiple chips. By separating the prefill and decode stages across different racks in their San Jose lab, Etched aims to optimize inference workflows that split context loading from token generation, a common bottleneck in current AI infrastructure.
What Happens Next
Etched has grown its team beyond 450 people and established a newer Milpitas, California research and development facility. This site includes a 10 MW data centre, a lab, and a quick-turn SMT line to accelerate hardware iteration. Customer shipments are scheduled to begin this summer, shifting the company’s focus from AI hardware funding milestones to installation and measurement tests.
The startup has not named specific rack customers or published benchmark methodology. EE Times observed racks operating in prefill and decode configurations during a lab visit, but Etched declined to make benchmark figures public. The $1 billion pre-order book remains a strong demand signal, not deployed customer capacity.
The summer shipments will determine whether the $1 billion pre-order book turns into deployed inference capacity or remains a speculative bet. Etched has deliberately chosen a supply path distinct from Nvidia’s Rubin GPUs, insulating its TSMC N4P hardware from the same supply chain constraints choking the broader AI industry. However, the technical caveats are physical, not just architectural. Controlling high current and current spikes at low voltages across a full-reticle-sized design is an engineering feat that has never been deployed at this scale. The entire AI industry will be watching the data. If Etched succeeds and delivers public performance figures proving an order of magnitude token throughput over Nvidia Blackwell, it could redefine AI compute economics and force a rethink of data center power consumption. If the physics fail to scale at the rack level, the low-voltage inference model will serve as a cautionary tale for vertical integration in silicon, and the pre-orders will evaporate.
— Derek Kimura, technology desk, AXO News