I didn't need to see the chip to know something was off. The numbers were too clean, the claims too precise. Etched, a chip startup that just closed a $700 million round at a $21 billion valuation, is selling a story of engineering perfection. But after spending the last five years auditing smart contracts and DeFi protocols, I've learned one thing: perfection is a lie. It's a red flag wrapped in a press release.
Context: The Chip That Promises Everything
Etched's core pitch is LVI (Low Voltage Inference) technology. The company claims its chips can run AI inference at significantly lower voltages, enabling trillion-parameter sparse MoE models to hit over 80% of theoretical peak performance. That's a big number. In the AI chip world, 80% Model Floating Utilization (MFU) is a holy grail. NVIDIA's H100, for context, peaks around 60-70% in real-world inference workloads. Etched says it's beaten that by a margin so wide it should make every competitor nervous.
But here's the problem: no one outside Etched has seen the data. Not the complete FLOPs figures. Not the power consumption per watt. Not a single third-party benchmark. The company's website still reads, 'Early customer tests have reached leading levels,' with a promise of detailed performance data coming 'soon.' In crypto, we call this a vaporware roadmap. In chips, it's called an unverified claim.
Core: The Forensic Teardown
Let's start with the numbers we do have. Etched claims an 80%+ MFU for sparse MoE models. MFU is a ratio: actual computation divided by theoretical peak. If the theoretical peak is low, a high MFU doesn't mean much. Wesley Yue, a chip designer, pointed this out: a chip with a low peak FLOPs could hit 90% MFU and still lose to a competitor with a lower MFU but higher raw throughput. Etched hasn't disclosed its peak FLOPs. So the 80% claim is floating in a vacuum.
The bottleneck wasn't the voltage. It was the lack of verifiable performance data. In my audit work, I've seen this pattern before. A project announces a breakthrough. They show hardware photos. They claim early customer adoption. But when you ask for the raw data—the transaction logs, the gas usage, the execution traces—they deflect. Etched has done exactly that. The Wall Street Journal and Reuters confirmed that chips have been shipped. Jane Street received a full rack last month. But shipping hardware is not the same as proving performance.
Let's break down the MFU claim. Sparse MoE models activate only a subset of parameters per token. This reduces computation but increases memory bandwidth demands. Etched's LVI technology supposedly lowers the voltage, which reduces power consumption, allowing higher clock speeds or more parallelism. But low voltage introduces timing errors. You need error correction or redundancy. Etched hasn't shown how it handles this. The absence of this detail is a red flag. In crypto, when a DeFi protocol hides its liquidation mechanism, it's because the mechanism is flawed. Here, Etched hides its error correction. Same logic.
You don't need to be a chip engineer to smell the problem. You just need to know how to read a data sheet. Etched's data sheet is blank. Compare this to NVIDIA or AMD, which publish detailed benchmarks, power curves, and thermal limits. Even startups like Cerebras publish whitepapers with full mathematical proofs. Etched has a website and a funding round. The asymmetry is telling.
Contrarian: What Bulls Got Right
To be fair, the skeptics might be wrong. Etched has shipped hardware. Jane Street is a credible quantitative trading firm—they don't deploy vaporware. The fact that Jane Street received a rack and is testing it means the chips exist. The question is whether they perform as advertised. If Etched's LVI technology works, it could be a genuine breakthrough. The sparse MoE model is the future of large language models, and a chip that efficiently handles them at low voltage would be worth billions.
But the bulls are betting on trust. They're extrapolating from the funding round and the names involved. In crypto, we saw this with Terra. The team had a working product, real users, and a $40 billion market cap. But the underlying mechanism was fragile. The collapse was a matter of when, not if. Etched might be different. But until they release third-party benchmarks, the comparison is valid.
Takeaway: The Valuation Is a Bet, Not a Fact
A $21 billion valuation for a company that hasn't published a single peer-reviewed benchmark is a bet on the team's reputation, not on the technology. The market is treating Etched's claims as proven, but the evidence is missing. In my years of analyzing on-chain data, I've learned that the absence of data is data. It tells you that the project is either hiding something or not ready for scrutiny. Either way, the risk is asymmetrical.
Etched's chips might be the future. Or they might be the next Theranos. The difference is that Theranos had a compelling story and a charismatic founder. Etched has a compelling story and a $700 million check. But the story is not the data. And in this industry, data is the only thing that survives the bear market.