Seven hundred million dollars. A $21 billion valuation. Zero usable benchmarks.
Etched just closed a funding round that would make most unicorns blush. The pitch: a chip that runs AI inference at lower voltages, courtesy of LVI technology. Trillion-parameter sparse MoE models hitting 80% of theoretical peak? That’s the claim. The problem is simple: no data. No FLOPs. No power consumption numbers. No third-party benchmarks. Just hardware photos, customer orders, and a website that says “early customer tests have reached leading levels.”
I’ve seen this movie before. It’s the same script played out in DeFi summer, in NFT minting bots, in the Terra collapse. Hype first, numbers later. And when the numbers finally arrive, they often don’t match the story.
George Hotz, founder of tiny corp and veteran of the AI trenches, publicly questioned Etched’s technical claims. He’s not alone. Chip designer Wesley Yue pointed out the MFU trap: Model Floating Utilization measures how much of a chip’s theoretical peak you’re using. If that peak is low, even 80% utilization is a joke. You’re running at 80% of a weak engine. Meanwhile, competitors might hit 60% of a much stronger engine and still crush you on absolute performance.
This is basic math. Yet the market is eating it up.
Let’s break down the context. Etched is building an ASIC for AI inference, specifically for sparse MoE models. The LVI technology allegedly allows lower voltage operation, reducing power draw and heat. That’s a real engineering challenge. Lower voltage means less noise margin, higher error rates, and more complex error correction. It’s not magic—it’s trade-offs. But the narrative focuses on the upside, ignoring the cost.
The funding round was led by institutional investors. The Wall Street Journal and Reuters confirmed that chips have shipped. Jane Street received a full rack last month. That’s real. But existence is not performance. I’ve audited 0x protocol arbitrage strategies in 2017 where the code worked, but the liquidity was too shallow to execute. The difference between a prototype and a production system is everything.
Core analysis: the MFU trap.
Wesley Yue’s critique is sharp. MFU is a ratio, not an absolute. If Etched’s chip peaks at 100 TFLOPS, and they achieve 80% MFU, that’s 80 TFLOPS. If Nvidia’s H100 peaks at 200 TFLOPS and achieves 60% MFU, that’s 120 TFLOPS. The H100 still wins by 50%. Etched’s 80% is a marketing number, not a competitive advantage. They need to show absolute performance, not relative efficiency. But they haven’t.
Why? Because the data might not be flattering.
Contrarian angle: the retail vs. smart money divide.
Retail sees $700M and assumes it’s real. Smart money knows that funding rounds are often about FOMO, not fundamentals. The same dynamic played out in DeFi: protocols raising millions on TVL numbers that were propped up by wash trading and liquidity mining. Etched’s orders might be real, but are they for production use or evaluation? Jane Street deploying a rack is a signal, but Jane Street is a quantitative trading firm, not a cloud AI provider. They care about latency, not throughput. The chip might be excellent for low-latency inference, but terrible for batch processing.
And that’s the blind spot. Everyone is benchmarking against Nvidia’s general-purpose GPUs. But Etched is an ASIC. It’s optimized for a narrow workload. If that workload doesn’t dominate the market, the chip is a niche product. The scarcity of Layer2 liquidity taught me that fragmentation kills value. A chip that only works well for sparse MoE models is a silo. The market might not reward it.
Takeaway: actionable price levels.
Etched’s valuation is a bet on a technology that hasn’t been proven. The smart money will wait for third-party benchmarks. When they drop, expect volatility. If the numbers beat Nvidia’s GPUs on key workloads, the valuation might double. If they fall short, the stock (if public) or private valuation will crater. The binary outcome is sharper than any DeFi position I’ve ever taken.
Speed is the only moat that doesn’t last. Etched’s real test is whether they can ship silicon that outperforms in the wild. Not in a lab, not in a press release. In a production rack, under real load, with real power limits.
I’ll be watching. So should you.
