Hook
A $700 million funding round. A $21 billion valuation. Hardware shipped to Jane Street, confirmed by The Wall Street Journal and Reuters. And yet, after months of hype, the single most important piece of evidence for a chip company—performance data—remains absent. Etched’s website still reads: “Early customer tests have reached leading levels.” That is not a specification. That is a placeholder.
The market is treating this as a success story. I see a liquidity signal that demands forensic scrutiny. When a company raises at a valuation that implies a $200 billion exit, but refuses to publish FLOPs, power consumption, or third-party benchmarks, the burden of proof shifts. The question is not whether the chips exist. The question is whether the architecture delivers what the valuation assumes.
Context
Etched is building an AI inference chip built around a core innovation called LVI—low-voltage inference. The claim is that by operating at lower voltages, the chip can run trillion-parameter sparse mixture-of-experts (MoE) models at over 80% of its theoretical peak performance. That is a bold number. For context, most production AI chips operate in the 30-60% Model Floating Utilization (MFU) range. An 80% MFU would be a structural breakthrough, not an incremental improvement.
The company’s founding team has deep technical roots. But the recent scrutiny comes from an unexpected source: George Hotz, founder of the tiny corp and creator of the tinygrad deep learning framework. Hotz, a hacker known for his early work on iPhone jailbreaks and self-driving systems, publicly questioned the lack of public validation. His critique is not a personal attack—it is a systems-level observation. If you have a chip that achieves 80% MFU, why wouldn’t you publish the benchmark? The answer, in his view, is either because the benchmark is unflattering, or because the metric itself is misleading.
Chip designer Wesley Yue added a critical nuance: a high MFU ratio does not automatically imply strong absolute performance. MFU measures the ratio of actual computation to the chip’s theoretical peak. If the peak itself is low, even 80% utilization can be outperformed by a less efficient chip with a higher peak. The real question is not the ratio—it is the raw throughput under real workloads.
Core
Let me break this down with the same rigor I apply to liquidity cascade analysis. In chip performance, there are three interdependent variables: peak FLOPs (theoretical maximum), power envelope (thermal design power), and actual throughput under a given model architecture. Etched has disclosed none of these. The only claim is a utilization rate.
Consider a simple thought experiment. Chip A has a theoretical peak of 1000 TFLOPS and achieves 80% MFU—800 TFLOPS real. Chip B has a peak of 2000 TFLOPS and achieves 40% MFU—800 TFLOPS real. Identical throughput. But Chip B’s lower MFU is not a weakness; it reflects a different architectural trade-off. The MFU metric, in isolation, tells you nothing about power efficiency, latency, or cost per inference.
The danger here is a trap I’ve seen in crypto protocol audits: vanity metrics. In DeFi, it was total value locked (TVL) as a proxy for security. Here, it is MFU as a proxy for performance. TVL can be inflated with stablecoins that never move. MFU can be inflated by choosing a model that fits the chip’s strengths while ignoring the workloads that matter to customers.
Etched’s core selling point is LVI technology. Low-voltage inference is a well-understood engineering principle: reducing voltage reduces power consumption per operation, but increases the risk of timing errors. The question is whether Etched’s implementation can maintain correct arithmetic at voltages that would cause other chips to fail. If they have solved that, they have a genuine moat. If they have not, the 80% figure is an artifact of a carefully selected test case.
The absence of third-party benchmarks is the most telling signal. In the chip industry, companies like NVIDIA and AMD publish extensive benchmark suites—MLPerf, SPEC, and custom workloads. Etched has not submitted to any independent evaluation. The company promises future data, but in a capital-intensive industry where $700 million has been deployed, “future” is not a timeline. It is a deferral.
I recall a similar pattern from the 2022 DeFi liquidity forensic I did on Terra/Luna. The protocol claimed algorithmic stability, but the underlying data—the relationship between LUNA price and UST supply—was never modeled under stress. The market accepted the narrative until the mechanics broke. The same principle applies here: if the performance data is not public, it is because the data would not survive independent scrutiny.
Contrarian
The prevailing narrative is that Etched is under scrutiny because the technology is unproven. I take a different view. The skepticism is healthy, but the real risk is not that the chips are fake—it is that the market is over-indexing on the wrong metric. The focus on MFU distracts from the more important question: does the chip solve a real deployment bottleneck for its target customers?
Jane Street has already received a complete rack and begun deployment. That is a signal. Jane Street is not a venture capital firm; it is a quantitative trading firm with a reputation for ruthless efficiency. They would not deploy a chip that underperforms. Their willingness to adopt suggests that Etched’s technology works for at least one use case. The question is whether that use case is representative of the broader AI inference market.
The contrarian angle is this: the lack of public benchmarks may be a strategic decision, not a sign of weakness. Etched is targeting a specific niche—sparse MoE models at scale. If they have optimized their architecture for that niche, publishing generic benchmarks could invite unfair comparisons to general-purpose chips like the H100. The chip may be a specialist, not a generalist. In a world where AI workloads are fragmenting, specialization is a legitimate strategy.
But that argument cuts both ways. If the chip is a specialist, the $21 billion valuation implies a market size that does not exist yet. Sparse MoE models are a growing trend, but they are not the majority of inference workloads. The valuation assumes that the market will shift toward Etched’s architecture. That is a bet on a future that has not been validated.
Takeaway
The biggest question is not whether Etched’s chips are real. They are. The question is whether the performance is real. And until the company publishes complete benchmarks—FLOPs, power, latency under standard workloads—the only honest answer is: we don’t know.
Liquidity doesn’t lie. The $700 million is a vote of confidence from investors who have access to data that the public does not. But in the chip industry, as in crypto, capital allocation is not a substitute for verification. The market is pricing Etched as if the breakthrough is proven. The technical community is asking for proof.
Trust is compiled, not given. Etched needs to compile that trust by opening its benchmarks. Until then, the $21 billion valuation is a floating number, untethered from the one thing that matters: the actual performance under load.
Standardize or be standardized. The chip industry has a long history of companies that promised architectural revolutions and delivered only marketing. Etched has the funding, the talent, and the early customer adoption. Now it needs the data. The market is watching. The clock is ticking.