Etched's $21B Valuation: The Data Void Behind the Silicon Hype

I don’t trust a chip until I see the benchmark. That’s the first rule of hardware analysis, and it’s the same rule I apply to on-chain protocols: code is not the product, data is the proof. Etched just closed a $700M round at a $21B valuation. The valuation is real. The investors are real. The hardware photos are real. But the performance data? That’s an empty ledger. Let’s start with the hook: a $21B valuation with no public FLOPs, no power consumption numbers, and no third-party benchmarks. In the crypto world, we call this a “white paper project.” In the chip world, it’s called a leap of faith. George Hotz, founder of tiny corp, publicly called out the gap: “Many investors, many orders, many photos. Where’s the data?” He’s right. The data doesn’t lie, but the absence of data tells a story too. Etched’s core selling point is LVI (Low Voltage Inference) technology—a technique that lets chips run AI inference at lower voltages. The claim is that this enables trillion-parameter sparse MoE (Mixture of Experts) models to achieve over 80% of theoretical peak performance. That sounds impressive. But as chip designer Wesley Yue pointed out, a high utilization ratio doesn’t mean strong absolute performance. Model Floating Utilization (MFU) measures the ratio of actual computation to theoretical peak. If the chip’s peak performance is low, 80% MFU could still be slower than a competitor’s 60% MFU on a higher-peak chip. It’s like claiming 80% fuel efficiency in a go-kart while a Ferrari runs at 60%—the Ferrari still wins the race. Based on my experience tracking DeFi protocol metrics during the 2022 crash, I’ve learned one thing: high utilization without absolute performance is a red flag. In DeFi, a 90% TVL utilization rate on a low-liquidity pool means nothing if the pool is tiny. The same logic applies here. Etched needs to disclose peak FLOPs (floating point operations per second) and power consumption to give context to their 80% MFU claim. Without that, the number is meaningless. Let’s look at the evidence chain. The Wall Street Journal and Reuters confirmed that Etched’s chips have been shipped. Jane Street received its first full rack last month and has already started deployment. That’s a positive signal—the chips exist. But existence is not performance. The crash wasn’t triggered by fake chips; it was triggered by unrealistic expectations. The same could happen here if the benchmarks don’t match the hype. Context: Etched is designing a chip specifically for transformer-based AI inference. The market is crowded: Nvidia dominates with its H100 and B200, AMD is pushing MI300, and startups like Cerebras and Groq offer alternative architectures. Etched’s LVI approach is unique—lower voltage reduces power consumption and heat, potentially enabling higher density in data centers. But the trade-off is lower clock speed, which reduces peak performance. The question is whether the efficiency gain outweighs the performance loss. Core analysis: I’ve modeled the math using public data from similar chips. Assume a hypothetical Etched chip with 100 TFLOPS peak at low voltage, achieving 80% MFU = 80 TFLOPS real. Compare to a Nvidia H100 at 200 TFLOPS peak with 60% MFU = 120 TFLOPS real. Etched loses by 40 TFLOPS. Even at 50% lower power, the absolute throughput is lower. For sparse MoE models, the comparison changes—sparsity can reduce effective computation by 50-90%, making Etched’s efficiency more attractive. But without knowing the exact sparsity pattern and memory bandwidth, we can’t validate the claim. Contrarian angle: The market is obsessed with MFU, but the real metric is “performance per dollar per watt.” Etched’s LVI might shine in power-constrained environments. If a data center can fit 2x the number of Etched chips in the same power budget as Nvidia, the total throughput could be higher even if individual chip performance is lower. That’s a blind spot that most analysts miss. The crash narrative of “Etched is overhyped” might be premature. The real question is whether Etched’s chip can achieve the same or better total throughput per rack. But here’s the immutable ledger: without third-party benchmarks, everything is speculation. The crypto world learned this lesson with Solana’s early days—the network ran fast in demos but fell apart under real load. Etched needs to release a public benchmark suite, ideally on a standard like MLPerf. Until then, the $21B valuation is a bet on the team, not the data. Takeaway: Next week, watch for two signals. First, whether Etched releases any FLOPs or power numbers. Second, whether Jane Street or other customers leak performance data. If the data doesn’t come, the market will reprice the risk. I don’t bet on chips without benchmarks. Neither should you.