George Hotz just called your bluff.
The tiny corp founder and notorious hacker didn’t mince words when he publicly challenged Etched’s technical claims. A $700 million funding round at a $21 billion valuation. Photos of hardware. Orders from Jane Street. But zero data that proves the chip actually delivers.
I’ve seen this movie before. In 2022, Terra promised algorithmic stability. In 2024, a dozen AI projects promised trillion-parameter models on consumer GPUs. The pattern is always the same: raise capital on narrative, delay benchmarks, then disappear when the numbers don’t add up.
Etched might be different. But “might” isn’t a risk management strategy.
Context: The Allure of LVI and Sparse MoE
Etched’s core narrative revolves around its Low Voltage Inference (LVI) technology. The chip supposedly runs AI inference at lower voltages, enabling trillion-parameter sparse Mixture of Experts (MoE) models to achieve over 80% of theoretical peak performance. That’s the Model Floating Utilization (MFU) number they’re leaning on.
MFU is a ratio: actual computation divided by theoretical peak. If your chip’s theoretical peak is low, 80% MFU can still be outperformed by a competitor with 60% MFU but a higher absolute peak.
Wesley Yue, a chip designer, raised exactly this point. A high utilization ratio does not guarantee strong absolute performance. It’s like a DeFi protocol advertising 90% capital efficiency while ignoring that its total value locked is $2 million. The metric matters, but only in context.
Etched’s website still says: “Early customer tests have reached leading levels.” Detailed performance data is promised “future release.” That’s not a benchmark. That’s a placeholder.
Core: The Data Gap That Smart Money Can’t Ignore
From my years auditing DeFi protocols, I’ve learned that claims without verifiable data are the first red flag. I once caught a reentrancy vulnerability in a stableswap contract because the team’s “audited” label didn’t match the bytecode. The same principle applies here.
Etched has not publicly disclosed: - Complete FLOPs (floating point operations per second) - Power consumption under load - Third-party benchmarks from independent labs - Latency or throughput for the trillion-parameter MoE use case
What they have disclosed? Customer names. Jane Street received its first complete rack last month and has already begun deployment. The Wall Street Journal and Reuters confirmed the chips are shipped. So the hardware exists. The question is whether it performs.
Consider the math. Sparse MoE models achieve efficiency by activating only a subset of parameters per token. The theoretical peak performance for such models depends heavily on the sparsity ratio, memory bandwidth, and interconnect topology. If Etched’s LVI reduces voltage but also reduces clock speed, the absolute FLOPs could be lower than a standard chip running at higher voltage.
Let’s assume Etched’s chip achieves 80% MFU on a sparse MoE model with 90% sparsity. That means the effective computation is 80% of the theoretical peak. But if the theoretical peak is 10 teraflops, the actual throughput is 8 teraflops. NVIDIA’s H100 can achieve 60% MFU on dense models at 200 teraflops—that’s 120 teraflops. Even with sparse optimization, the H100’s absolute performance dwarfs Etched’s if the peak is an order of magnitude lower.
This is basic arithmetic. Yet the market is pricing Etched at $21 billion without this arithmetic.
Contrarian: The Euphoria Masks a Structural Risk
The contrarian angle here is not that Etched’s chip is fake. It’s that the market is treating a $21 billion valuation as justified by narrative alone, ignoring the due diligence gap.
We’ve seen this in crypto. Projects with $100 million valuations on testnets that never went live. Protocols with “TVL” that was mostly wash trading. The pattern is the same: raise on hype, promise data, and rely on the fear of missing out to keep the price up.
Hotz’s critique is a classic smart money signal. He’s not saying the chip doesn’t exist. He’s saying the data doesn’t exist. And in a capital-intensive industry like AI chips, where production runs cost hundreds of millions, a lack of performance data is a existential risk.
Jane Street’s deployment is a positive signal, but it’s not a third-party benchmark. Jane Street is a quant firm, not a chip testing lab. They may have internal validation, but that’s not public.
Alpha isn’t broadcasted. Neither is real chip performance.
Takeaway: The Next Six Months Will Determine the Truth
Etched has the money, the partners, and the narrative. But the clock is ticking. If they don’t release comprehensive benchmarks—including FLOPs, power, and independent MLPerf results—the market will eventually discount the valuation.
Regulation is coming for AI hardware, just like it came for crypto. The SEC’s scrutiny of SPACs and pre-revenue companies is only increasing. Etched’s investors should demand transparency before the next round.
Audit the code, ignore the influencer. In this case, audit the benchmarks, ignore the hype.
Panic is just inefficient pricing. But for now, the market is pricing Etched on hope. That’s not a thesis. That’s a gamble.