Audit trail incomplete. Red flag raised.
Etched just closed a $700 million round at a $21 billion valuation. The chip startup claims its LVI technology enables trillion-parameter sparse MoE models to hit over 80% of theoretical peak performance. But the data room is empty. No third-party benchmarks. No public FLOPs. No power consumption numbers. The website says “early customer tests have reached leading levels.” That’s not a specification. That’s a marketing slide.
George Hotz — founder of tiny corp, creator of tinygrad, and a hacker who doesn’t mince words — publicly questioned the claims. He pointed out that while there are investors, orders, and hardware photos, there is no data to validate performance. In crypto terms, this is a token launch with a whitepaper but no GitHub repo. The market is pricing in a future that hasn’t been proven yet.
Context: The LVI Promise and the Missing Metrics
Etched’s core selling point is Low Voltage Inference (LVI). The idea is elegant: run AI inference at lower voltages to drastically reduce power consumption while maintaining throughput. For sparse mixture-of-experts models with trillion parameters, this could be a game-changer. But here’s the rub — Model Floating Utilization (MFU) measures the ratio of actual computation to theoretical peak. If the chip’s theoretical peak itself is low, an 80% MFU could still be slower than a competitor’s chip running at 50% MFU. Chip designer Wesley Yue nailed it: “High utilization does not imply high absolute performance.”
In crypto, we see this all the time. A DeFi protocol claims a 99% uptime, but the underlying chain only processes 10 TPS. The metric is misleading without context. Etched is pulling the same trick — they’re showing efficiency without showing the raw power.
Core: What We Actually Know vs. What We’re Told
Let’s break down the verified facts. The Wall Street Journal and Reuters have confirmed that Etched chips have shipped. Jane Street received its first complete rack last month and has already begun deployment. So the hardware exists. The question is whether it performs as advertised.
But here’s where my crypto audit experience kicks in. When a project ships a product but refuses to publish standardized benchmarks, it’s a red flag. Think of the 2020 0x Protocol v2 exploit I flagged — the code was live, but the reentrancy vulnerability was hidden in the exchange logic. The team didn’t disclose it until it was exploited. Etched is doing the same dance: they show the hardware, they show the customers, but they hide the performance data.
Quantitative ROI Analysis
Assume Etched’s chip delivers 80% MFU on a trillion-parameter MoE model. Now compare to NVIDIA’s H100, which runs at roughly 50-60% MFU on similar workloads. If both chips have the same theoretical peak FLOPs, Etched wins. But if Etched’s peak FLOPs is half of H100’s, then 80% MFU equals 40% of H100’s absolute performance. That’s a disaster. Without knowing the peak FLOPs, we can’t calculate ROI.
Here’s the table that matters:
| Metric | Etched (Claimed) | NVIDIA H100 (Known) | Implication | |--------|------------------|---------------------|-------------| | MFU (sparse MoE) | >80% | ~50-60% | Etched looks better on paper | | Peak FLOPs (FP16) | Undisclosed | 989 TFLOPS | Unknown if Etched is even close | | Power Consumption | Undisclosed | 700W | Unknown efficiency per watt | | Third-Party Benchmark | None | MLPerf, etc. | No independent verification |
Based on my analysis of crypto infrastructure projects, the lack of disclosure is a deliberate strategy. It buys time for the hype cycle to inflate the valuation before the numbers are released. Remember the Luna/UST collapse? The team kept saying “the peg is stable” while the data showed liquidity drying up. I watched the on-chain metrics in real-time. The same pattern is emerging here: the story is strong, but the data is weak.
Contrarian: The Chips Exist, But the Hype May Be a Distraction
Here’s the counter-intuitive angle. Etched’s hardware is real. Jane Street deploying is a strong signal. But the crypto-native mindset teaches us that real products can still be overvalued. Consider the Arbitrum airdrop farming strategy I led in 2023 — we optimized gas-efficient bridging, calculated ROI, and executed. The strategy worked because we had data. Etched’s investors are farming hype without data. They’re betting on potential, not proof.
More importantly, the AI chip market is becoming a winner-take-most game. NVIDIA’s CUDA ecosystem is a moat that cannot be crossed by a single hardware breakthrough. Etched’s LVI technology, even if real, faces adoption friction. Developers need to rewrite model code to leverage sparse MoE patterns. That’s like asking DeFi users to migrate from Uniswap V3 to a new DEX just for a lower fee — it’s possible, but unlikely at scale.
Liquidity drying up. Watch the spread.
Takeaway: Wait for the Benchmark, Not the Hype
Every crypto bull market generates unicorns that promise to “revolutionize” the industry. Most of them disappear when the data comes out. Etched’s $21 billion valuation is a bet on the future, not a reflection of current reality. The smart money will wait for third-party benchmarks, FLOPs disclosure, and independent power measurements. Until then, treat the claims like a crypto whitepaper without a GitHub repo — interesting, but uninvestable.
Forward-looking thought: The real test will come when Etched submits to MLPerf or another standardized benchmark. If they don’t, the market will eventually discount the hype. I’ve seen this cycle before. The question is: will you be holding the bag when the data drops?