Etched, the AI chip startup building custom silicon for inference, published a technical update this week describing a new architecture that keeps its processor’s math circuits running below half the voltage common in AI chips sold today. The company says the change lets it sustain more than 80 percent of peak FLOPs on trillion-parameter mixture-of-experts models without the clock-speed throttling that limits rival hardware. Etched frames the design as an answer to a problem every AI chipmaker now shares: heat, not raw compute, is what stops a chip from running at full speed.
That framing rests on a physical relationship the company does not spell out but that determines whether the claim matters. Power draw in a chip’s circuits scales roughly with the square of voltage, so a voltage cut close to half translates into a power reduction well beyond half. Less power drawn means less heat generated, and less heat means a chip can hold its clock speed instead of downshifting under load, which is the throttling problem Etched says its architecture avoids.
The company says most AI chips today lose more than half their peak FLOPs to this exact mechanism: as utilization climbs, chips pull more power, run hotter, and throttle their clocks to stay alive. Etched calls its low-voltage design LVI and says it required rebuilding the chip’s math arrays, power delivery network, voltage regulator architecture, and cooling plates together rather than treating voltage as fixed. The company paired LVI with a second architecture change, a hybrid memory system it calls Cluster Scale Memory, meant to close the latency gap between HBM-based chips and pure SRAM designs across a rack.
None of this has been verified outside Etched. The company’s post is a self-published account of its own hardware, tested in what it describes as representative data center deployments and simulator runs against production traffic patterns, with no third-party benchmark or independent silicon evaluation cited. A claim about how a math block behaves in isolation is not the same as a claim about how a finished chip performs end to end once it is racked, networked, and running real customer workloads at scale. Etched puts the arrival of its debut racks in the coming months; that is the point at which the voltage claim gets tested outside the company’s own labs.
AI Insiders covered Etched on July 1, when the story was $1 billion in customer chip orders and a $5 billion valuation. This is a separate development: an architecture claim about how the silicon itself is engineered, not a financing or order-book update. The distinction matters because voltage and thermal efficiency, not order volume, are what determine whether Etched’s chips can actually deliver the throughput its customers are paying for.
The underlying strategic logic explains why a voltage number is worth attention at all. Transistor density used to be the industry’s main lever. Power and heat are now the binding constraint on data-center AI, which is why TSMC is pouring capital into advanced packaging and why hyperscalers are redesigning data centers around liquid cooling and denser power delivery. Etched’s low-voltage claim is aimed at that same bottleneck, not a new one.
For hardware buyers evaluating 2026 inference contracts, the number to ask Etched for next is not FLOPs density but measured power draw and sustained throughput on a shipped rack, under a real customer workload, verified by someone other than Etched.
Published by Etched on July 23, 2026.
