Most people think a $21 billion valuation means the chip works. It doesn’t.
I’ve seen this playbook before. In 2021, a DeFi protocol raised $200 million at a $2 billion valuation. The founder promised a zero-slippage AMM. The code was closed source. The testnet had a “trust us” narrative. I ran my own analysis—forked the repo, simulated the math. The curve was broken. The floor didn’t hold. The project collapsed within six months, but not before retail apes got burned.
Etched is the same story, different hardware. $700 million raised. $21 billion valuation. Orders from Jane Street. Hardware photos. But no third-party benchmarks. No FLOPs numbers. No power consumption data. The only data point is a promise: “Early customer tests have reached leading levels.” That’s not a data point. That’s a placeholder.
Context: The Etched Narrative
Etched builds an AI inference chip with a proprietary technology called LVI—Low Voltage Inference. The claim is that LVI allows the chip to run trillion-parameter sparse Mixture-of-Experts (MoE) models at over 80% of theoretical peak performance. That’s an impressive number—if it’s true. But the definition of “peak performance” is a moving target. Peak performance is the theoretical maximum FLOPs the chip can achieve under ideal conditions. Real-world performance is always lower. Etched says their chip achieves 80% of that theoretical peak. Sounds good. But what if the theoretical peak itself is low?
Chip designer Wesley Yue raised the exact question. Model Floating Utilization (MFU) is a ratio. A high MFU doesn’t mean absolute performance is high. It just means the chip is using its own capacity efficiently. If your chip’s peak is 10 TFLOPS, 80% MFU gives you 8 TFLOPS. If a competitor’s chip peaks at 100 TFLOPS and achieves 50% MFU, they get 50 TFLOPS. The ratio is a distraction. The absolute number is what matters.
Etched has not published any absolute FLOPs figures. They have not released standard benchmarks like MLPerf. They have not provided power consumption per inference. The website says: “Early customer tests have reached leading levels.” Leading compared to what? A 2020 GPU? A 2024 ASIC? The term is empty.
Core: The Mechanics of Verification
Let’s break down what a chip buyer actually needs to validate. There are three pillars: theoretical peak performance, real-world performance under load, and power efficiency. All three must be independently verified.
First, theoretical peak. This is a simple arithmetic calculation—number of multiply-accumulate units times clock speed times operations per cycle. For AI chips, it’s usually measured in TFLOPS (tera floating-point operations per second). A company can publish this number. It’s easy to compute. Etched hasn’t.
Second, real-world performance. This is measured by running a known model (like GPT-3 or Llama) and measuring throughput in tokens per second. The industry standard is MLPerf, a benchmark suite managed by MLCommons. Etched has not submitted any MLPerf results. They claim “early customer tests” but customers are not impartial. Jane Street received a rack last month and is deploying it. Jane Street is a proprietary trading firm—they care about latency, not about publishing benchmarks. They might be using the chip for a specific, narrow use case. That doesn’t tell us if the chip is general-purpose.
Third, power efficiency. Chip performance is meaningless without power. A 10 TFLOPS chip that draws 1000 watts is worse than a 5 TFLOPS chip that draws 200 watts. Power efficiency is the key metric for data center operators. Etched hasn’t shared TDP (thermal design power) or watts per inference.
I’ve been through this drill before. In 2020, I audited a DeFi yield aggregator that claimed “200% APY with zero risk.” The code had a hidden reentrancy vulnerability. The yield was real—until it wasn’t. The floor didn’t hold. The protocol drained in a single transaction. The same logic applies here. If Etched is real, they should be able to provide a simple benchmark: run a standard model, measure throughput, measure power, compare to NVIDIA H100. That’s a 30-minute test. They haven’t done it.
The Contrarian Angle: Why Retail Is Wrong and Smart Money Is Waiting
Retail investors see $700 million raised and $21 billion valuation and think “this must be real.” That’s emotional, not analytical. Smart money—the VCs, the hedge funds—they’re not buying the chip yet. They’re buying the story. The valuation is based on a future where LVI chips dominate. But the future is uncertain. Jane Street is deploying a rack—that’s a pilot, not a commitment. They could have bought one rack for testing. That’s a $100,000 order, not a $100 million order. The WSJ and Reuters confirmed the chips exist. That’s table stakes. Existence doesn’t equal performance.
Volatility is a tax on the unprepared. Etched’s volatility is the uncertainty of their claims. If the chip is as good as advertised, the stock (if it trades) will moon. If it’s not, the valuation will collapse. The market doesn’t know which is true. That’s the definition of a binary event. Retail traders are buying options on a binary outcome without knowing the underlying probability. That’s a losing trade in the long run.
George Hotz, the hacker and founder of tiny corp, called out the lack of data. He’s known for being blunt and technically rigorous. When he says “there are many investors, orders, and hardware photos, but a lack of data to validate performance,” I listen. He’s the type of person who would simulate the chip’s architecture in software to verify the claims. He hasn’t published a detailed analysis—but the fact that he’s skeptical means the burden of proof is on Etched.
Takeaway: The Only Metric That Matters
Capital efficiency is the only metric that matters. Etched’s capital efficiency is currently zero—they’ve raised $700 million but produced no data that allows a buyer to calculate ROI. The chip might be a breakthrough. It might be a dud. The market doesn’t know. The floor didn’t hold for many projects that had similar hype—Theranos, for example. The difference is that Theranos’s technology was fake. Etched’s technology might be real, but the lack of transparency is a red flag.
Smart money doesn’t chase narratives, it engineers them. If I were a fund manager, I would not allocate capital to an Etched-related token (if one existed) until I saw independent benchmarks. The timing is everything. The first reliable third-party benchmark will trigger a massive re-rating—either up or down. That’s the point of maximum risk and reward. Until then, the $21 billion valuation is a mirage.
The floor didn’t hold. The question is: will it hold for Etched? The answer depends on whether they can deliver data, not chips. Chips are easy to make. Data is hard to fake. We’ll see.
