Etched Emerges With New AI Inference Hardware and More Than $1 Billion in Customer Demand

AI hardware startup Etched has come out of stealth mode with plans to build large scale inference systems designed for frontier AI models. The company says it has already completed the first tapeout of its custom silicon, raised $800 million, and secured more than $1 billion in customer demand for its upcoming rack-scale products.

Etched is positioning itself as a specialist in AI inference, the stage where trained models generate responses, write code, process documents, and handle real-time requests. While training large models receives much of the attention, inference is becoming one of the largest long-term costs for AI companies. That creates an opportunity for new hardware firms that can improve speed, lower power use, and reduce the cost of serving each request.

The startup says its first systems will begin shipping this summer.

Etched Is Building Hardware Around Inference Rather Than General Compute

Etched’s approach is built around the idea that AI inference needs different hardware priorities from training. Instead of relying only on traditional accelerator designs, the company is developing chips, memory systems, racks, cooling hardware, interconnects, and software as one combined platform.

The company says it has hired more than 400 engineers with backgrounds from companies including NVIDIA, Google, Broadcom, TSMC, and SK Hynix. That is a large team for a young semiconductor startup and suggests Etched is trying to compete at the infrastructure level rather than with a single chip product.

Its first silicon reportedly taped out earlier in 2026 using TSMC’s N4P manufacturing process. Tapeout means the design has been finalized and sent for manufacturing, though the real test will come when production hardware is available in larger quantities.

Low Voltage Design Could Target Better Sustained Performance

One of Etched’s main technologies is called Low Voltage Inference, or LVI. The company says its architecture can run at roughly half the voltage used by many existing AI chips.

Lower voltage can reduce power use and heat, both of which are major concerns in modern AI data centers. AI accelerators often face limits from cooling, power delivery, and thermal throttling when pushed close to peak performance.

Etched claims its LVI processor can maintain around 80 percent of peak floating point performance when running trillion-parameter sparse mixture-of-experts models. That claim will need independent testing once the hardware reaches customers, but the target is clear: sustained inference performance rather than short benchmark bursts.

A New Memory Design Combines HBM and SRAM Ideas

The company is also introducing a Cluster Scale Memory system, or CSM, aimed at low latency AI workloads. This design appears to combine elements of high bandwidth memory and SRAM based approaches.

HBM offers high bandwidth but is expensive and difficult to source. SRAM is much faster for certain tasks but offers lower capacity and can be costly at larger scales. Etched says its memory system aims to balance capacity, latency, bandwidth, cooling, reliability, and manufacturing yield.

That could be important for AI models that need fast access to large amounts of memory during inference, especially long context and agent based workloads.

Etched Faces a Difficult Market Despite Its Early Momentum

The AI accelerator market is crowded and expensive. NVIDIA remains the dominant player, while AMD, Google, Amazon, Microsoft, Broadcom, and several startups are investing heavily in custom inference hardware.

Etched will need to prove that its systems can deliver the promised performance, power efficiency, and cost advantages in real deployments. Building a chip is only one part of the challenge. The company also needs reliable manufacturing, memory supply, software compatibility, customer support, and a strong production ramp.

Still, the company’s early funding, experienced engineering team, and reported customer interest suggest that investors and AI providers are actively looking for alternatives to traditional GPU based infrastructure.