Etched, the startup that promised a transformer-focused chip would outperform Nvidia's GPUs, has quietly pivoted to a completely different technology strategy after its first chip design failed due to thermal issues. The company is now banking on two new approaches: highly under-volted logic that reduces power consumption, and pooled rack-level memory modules that improve data sharing across chips. The question facing the industry is whether this represents genuine innovation or a rebranding of a flawed original premise.

What Went Wrong With Etched's Original Plan?

When Etched launched in 2022, the company's core argument was straightforward: just as Bitcoin mining ASICs dramatically outperform GPUs by focusing purely on a single algorithm, a chip designed exclusively for large language model (LLM) inference should crush general-purpose graphics processing units (GPUs). The logic seemed sound, but it overlooked a critical flaw. Unlike Bitcoin's SHA-256 hashing algorithm, which uses specialized bit operations that GPUs handle poorly, LLM inference relies primarily on matrix multiplication, a task that modern GPUs and tensor processing units (TPUs) already handle extremely well.

More importantly, Nvidia was already aggressively optimizing its datacenter GPUs specifically for LLM workloads. This meant Etched wasn't carving out a new market niche; it was directly challenging Nvidia without any meaningful technological differentiation. According to industry observers, the company's first chip design, codenamed Sohu, encountered severe thermal problems during development, forcing a complete system redesign. Senior engineers reportedly left the company amid leadership conflicts, and the startup's reliance on external contractors for physical chip design contributed to the failures.

How Does Etched's New Low-Voltage Strategy Work?

Facing mounting pressure and a melting chip, Etched's leadership made a strategic pivot. The company's chief architect, who had previously worked on Bitcoin miners, recognized that aggressive under-volting could solve the power consumption problem. The physics behind this approach is elegant: when you reduce the supply voltage of a circuit by a factor of N, the circuit runs slower by a factor of N, but power consumption drops by a factor of N squared. This means Etched's new chips could dramatically improve power efficiency, measured in tokens per watt, without necessarily improving throughput, measured in tokens per second.

The tradeoff is significant. For applications where speed matters, like AI agents that need to respond quickly to user queries, the added latency from under-volting could be a serious problem. However, for datacenter operators focused purely on cost per token processed, the power efficiency gains could be substantial. Etched is also using custom standard cells to operate its designs below the voltage levels that Taiwan Semiconductor Manufacturing Company (TSMC) standard cells are designed for, a technically challenging but potentially valuable optimization.

What About Etched's Memory Technology?

The second pillar of Etched's new strategy involves cluster-scale memory architecture. The company claims this approach solves latency challenges inherent in high-bandwidth memory (HBM) systems by allowing chips to write directly to and read from HBM in other chip packages. On traditional GPUs, such peer-to-peer writes are routed through L2 cache, adding latency. If Etched's system bypasses this routing, it could meaningfully improve performance, though it would place greater demands on compilers and programmers to manage data flow without the safety guarantees that cache coherency provides.

The company has released minimal technical details about this memory subsystem, making it difficult to assess whether the claimed improvements will materialize in real-world deployments. This lack of transparency stands in contrast to competitors like Groq and Cerebras, which have been more forthcoming about their architectural choices.

How to Evaluate Etched's Competitive Position in Inference Chips

• Power Efficiency Focus: Etched's under-volted approach prioritizes tokens per watt over tokens per second, making it attractive for cost-conscious datacenter operators but potentially problematic for latency-sensitive applications like real-time AI agents.
• Custom Silicon Complexity: The company's reliance on custom standard cells and non-standard voltage levels increases engineering risk and manufacturing complexity, requiring expertise that few teams possess.
• Competitive Landscape: Etched must compete not only with Nvidia's dominant GPU ecosystem and CUDA software platform, but also with specialized inference chips from Groq and Cerebras, which prioritize speed over power efficiency.
• Software Ecosystem Maturity: Even if Etched's hardware performs well, the company lacks the accumulated software libraries, cloud integrations, and developer knowledge that Nvidia has built over decades with CUDA.

Is Etched Actually Good Now, or Just Better at Marketing?

The honest answer is that it remains unclear. Etched's original premise was flawed, and skeptics who questioned the transformer-focused ASIC approach had legitimate reasons for their doubts. However, the company's pivot to under-volted logic and custom memory architecture represents genuine technical innovation, even if it abandons the original vision. The fact that Etched managed to raise enough capital to hire world-class silicon engineers who developed these new approaches suggests that venture capital can sometimes fund companies through their worst ideas to their better ones.

The real test will come when Etched releases detailed technical specifications and customers begin deploying the chips in production environments. Until then, the company's claims about power efficiency improvements remain theoretical. The old Etched website made grandiose promises about tokens-per-second performance that never materialized; the new website touts a technology that explicitly trades speed for efficiency. For customers optimizing purely for cost per token, this could be a genuine advantage. For everyone else, it remains a niche solution.

Meanwhile, the broader inference chip market continues to evolve. Groq and Cerebras have taken the opposite approach, building extremely fast but power-hungry systems. AMD is positioning itself as a middle ground, focusing on memory access efficiency for inference workloads. Nvidia, through its acquisition of Groq, has added inference-specific capabilities to its dominant training platform. In this competitive landscape, Etched's success depends not on whether it was right about transformers, but on whether its new technology actually delivers the power efficiency gains it promises.