Finding the signal in the static of the new wave.
Hook: The Hacker’s Crosshairs
George Hotz didn’t ask for permission. He never does. The man who hacked the iPhone at 17, who built self-driving software in his garage, and who now runs tiny corp, the open-source AI framework tinygrad, fired a salvo that rippled through the chip industry’s echo chamber. "There are many investors, many orders, many hardware photos. But where is the data?" His words hung in the air like smoke after a black powder shot.
I was in Seoul, nursing a cold Americano, scrolling through the stream. The static was deafening. Etched had just closed a $700 million round at a $21 billion valuation. The narrative was perfect: a David vs. Goliath story, a chip startup promising to break the AI inference bottleneck with something called LVI—Low Voltage Inference. The chorus of celebratory tweets was a wall of noise. But Hotz, a man who reverse-engineers systems for fun, was pointing at the emperor’s new clothes. He wasn’t questioning the existence of the clothes. He was questioning the thread count.
Context: The Narrative of the New Silicon
Etched’s pitch is a seductive one. AI inference is exploding. Every chatbot, every image generator, every autonomous agent needs compute. The dominant players—NVIDIA, AMD, Google—are fighting over the data center, but Etched claims to have found a shortcut. Their LVI technology allows chips to run at lower voltages while maintaining high throughput. Their core promise: trillion-parameter sparse Mixture-of-Experts (MoE) models can achieve over 80% of theoretical peak performance on their hardware.
To understand why this matters, I have to tell you about the bear market. No, not crypto’s—the chip industry’s. In 2022, when the AI hype cycle crashed, everyone was looking for survivors. I was writing deep dives on modular blockchains, but I also ran a side project called "The Skeleton Key," dissecting infrastructure resilience. I noticed a pattern: the survivors were the ones that could prove their claims. Celestia had a public testnet. Ethereum had a roadmap. Etched had a website that still reads, "Early customer tests have reached leading levels." The phrase is a placeholder. A promise. A narrative.
But narratives are hollow without data. In crypto, we learned that the hard way with Terra, with FTX, with every project that promised moonshots but delivered footnotes. The chip industry is no different. The question is not whether the chips exist. The Wall Street Journal and Reuters have confirmed that Etched shipped to Jane Street. Last month, Jane Street received its first complete rack and has already begun deployment. The chips are real. The question is whether they are as powerful as advertised.
Core: The Signal in the Static
Let’s talk about MFU—Model Floating Utilization. It’s the ratio of actual computation to theoretical peak. Etched claims over 80% for sparse MoE. Sounds impressive. But here’s the catch: if the chip’s theoretical peak is low, even 80% utilization is a glass ceiling. Wesley Yue, a chip designer, pointed out the obvious: utilization is not a measure of absolute performance. It’s a measure of how well you’re using your own ceiling. If your ceiling is lower than the competitor’s floor, you’re still losing.
Based on my experience auditing smart contract claims—the same kind of due diligence that separates a real protocol from a rug pull—I know that the first thing to look for is the denominator. In crypto, it’s total supply, total value locked, or issuance rate. In chips, it’s peak FLOPs, power consumption, and real-world benchmarks. Etched has not publicly disclosed these. Not fully. Their website says "leading levels." Investors rely on NDAs and demos. But the market is a public ledger of trust.
The narrative here is a classic crypto tale: a high valuation, a charismatic founder, a technological breakthrough that sounds too good to be true. The echo chamber amplifies it. But the signal—the actual data—is buried in the static. I’ve been tracking this pattern since 2020, when I first started analyzing Uniswap’s liquidity mining yields. The APY was high, but the underlying token was inflating. The narrative of "yield farming" masked the reality of subsidized speculation. Etched’s narrative is similar: "80% utilization" sounds like a victory, but if the chip’s peak performance is half of an H100, then 80% utilization is still half the performance.
And the comparison is inevitable. NVIDIA’s H100 dominates the AI inference market. Its H100 achieves around 60-70% MFU on dense models. For sparse MoE, the numbers are lower due to memory bandwidth constraints. Etched claims to have solved this with LVI. But where are the independent benchmarks? Where is the MLPerf submission? Where is the third-party audit?
I’ve been in this industry long enough to know that shipping silicon is a miracle. Etched has done that. But shipping silicon that outperforms the incumbent by a factor of 2, 3, or 10? That’s a different miracle. The crypto analog is a blockchain that claims to process 100,000 TPS without sharding. We’ve heard that story before. The difference is that blockchains are open source; chips are black boxes. The only way to verify is to run your own workloads. Jane Street is doing that. But the rest of us are left with the narrative.
Contrarian: The Unseen Edge
The contrarian angle is that the skepticism is a feature, not a bug. In crypto, the best projects are the ones that thrive under scrutiny. They invite audits. They publish benchmarks. They let the community verify. Etched is acting like a traditional silicon startup: secretive, under NDA, with a promise to reveal later. That’s normal for the chip industry. But the crypto world operates on transparency. The irony is that Etched’s investors are betting on a narrative that the industry will break from its old habits.
What if the chips are actually good? What if LVI is a genuine breakthrough? The risk is that the market is underpricing the probability. The contrarian take is that Hotz, for all his brilliance, might be wrong. He’s a hacker, not a chip architect. He’s arguing from first principles, but first principles don’t always account for physical breakthroughs. Low voltage inference is a known R&D direction. If Etched has a patentable moat, the $21 billion valuation might be a discount.
But here’s the rub: in crypto, we’ve seen this movie before. The project that promises to be the next big thing, raises billions, ships a product, but fails to deliver on the hype. The difference is that chips are harder to fake than tokens. You can’t inflate a chip’s hashrate. You can’t manipulate its benchmark scores if third parties test it. The biggest question now is not whether the chips exist, but whether they are as powerful as advertised. The market will find out. Jane Street’s deployment is a signal. If they reveal performance data, the narrative shifts. If they stay quiet, the static grows.
Takeaway: The Next Narrative
The next narrative in the AI-crypto convergence is verifiable compute. Projects like Render, Akash, and Gensyn are built on the premise that you can trust the hardware running your models. Etched is a test case. If they cannot prove their performance, the entire sector will suffer. If they can, it will accelerate the adoption of specialized AI chips for decentralized networks.
I’ll be watching the data. The signal is still buried. But the noise is getting louder. And in the static of the new wave, the only truth is the one you can verify.
Finding the signal in the static of the new wave.