A $700 million raise at a $21 billion valuation. A chip that claims to run trillion-parameter sparse MoE models at over 80% of theoretical peak. And a cargo rack of hardware already landed at Jane Street’s data center. But for all the money and momentum, Etched is missing one thing: public proof. The ledger does not lie, but it rewards patience. And right now, the market is being asked to accept a single number—Model Floating Utilization (MFU)—as a stand-in for actual performance. That’s a dangerous shortcut in a sector where speed runs require foresight, not just reaction.
From the noise of 2017 to the signal of today, the AI chip race has become the new ICO mania. Capital flows in, hype builds, and everyone looks for the next Nvidia. Etched’s LVI technology, which runs AI inference at lower voltages, is the current darling. But George Hotz, founder of tiny corp and the hacker behind tinygrad, has publicly questioned the lack of raw data. No FLOPs, no power consumption benchmarks, no third-party audits. Just a utilization ratio that sounds impressive but could be meaningless if the raw compute ceiling is low. Based on my experience auditing chip performance claims during the 2020 DeFi yield wars, I’ve learned that a high ratio on a weak base is just noise. The market needs to separate signal from marketing.
Core: MFU is a distraction, not a destination.
Wesley Yue, a chip designer with a track record in tensor processing, raised the critical point: MFU measures the percentage of theoretical peak achieved, not the absolute FLOPs delivered. If Etched’s chip has a peak of 300 TFLOPS and runs at 80% MFU, it delivers 240 TFLOPS. A competitor with 500 TFLOPS peak at 60% MFU delivers 300 TFLOPS. The higher utilization doesn’t win. The market needs to know the raw upper bound. Etched has not disclosed this. Their website still says “early customer tests have reached leading levels,” with detailed data promised later. That’s the same playbook we saw in 2017 ICOs: “audit coming soon.” The sophistication of the investors—Jane Street, WSJ, Reuters—doesn’t change the fact that the technical community is being asked to trust, not verify.
Let’s break down the LVI technology. Running at lower voltage reduces power consumption and heat, which is crucial for inference at scale. But voltage reduction also limits clock speed and, therefore, peak FLOPs. The question is whether the trade-off is worth it. For sparse MoE models, where only a fraction of parameters are activated per token, a lower-voltage chip might hit higher utilization because the memory bandwidth becomes the bottleneck, not compute. Etched’s claim of over 80% MFU on trillion-parameter sparse MoE implies they’ve optimized the memory-to-compute ratio. But without independent testing, we don’t know if that optimization is real or a result of selective benchmark conditions.
Contrarian: The lack of data might be a strategic play, not a flaw.
Here’s the angle most analysts miss: Etched may be deliberately withholding benchmarks to manage expectations and avoid early comparison with Nvidia’s Blackwell or AMD’s MI300. If they release numbers now, the market will immediately compare them to established players. If those numbers are good but not earth-shattering, the $21 billion valuation could crater. By keeping data under wraps, Etched buys time to build a narrative of “early customer success” and lock in more pre-orders. Jane Street receiving a rack last month is a real signal—they’re a sophisticated firm that would not deploy hardware without internal validation. But internal validation is not public validation. The market is left to speculate on Jane Street’s due diligence, which is a risky game.
Additionally, the crypto angle is unavoidable. Decentralized AI compute networks—like Render, Akash, and IO.net—are hungry for inference chips. Etched’s low-voltage, high-utilization chip could be a perfect fit for edge nodes in these networks, where power and cooling are constrained. If Etched ships to crypto miners and AI compute providers, their LVI technology could become the de facto standard for decentralized inference. But that’s a long-term bet. Short-term, the market is pricing in a near-term monopoly on sparse MoE inference. That’s a fragile assumption.
Takeaway: Watch for the shipping date, not the MFU.
Speed runs require foresight, not just reaction. The next signal for Etched is not a benchmark release—it’s the volume and velocity of their customer deployments. If Jane Street scales up and others follow, the chip’s real-world performance will emerge organically. Until then, the $21 billion valuation is a bet on LVI’s potential, not its proven superiority. The ledger does not lie, but it rewards patience. In a sideways market, chop is for positioning—and right now, the smart money is waiting for the data, not the headlines.