Floor price broken. Truth verified.
When a chip startup raises $700 million at a $21 billion valuation before shipping a single public benchmark, the crypto community knows the drill. We’ve seen this movie before — in 2017 with ICOs promising “world computer” specs, in 2021 with NFT projects claiming revolutionary art protocols, and now in 2025 with AI inference chips. Etched, the darling of the AI hardware wave, is under fire. George Hotz, the hacker behind tinygrad and tiny corp, publicly questioned the company’s technical claims. His attack vector is simple: there are investors, there are orders, there are glossy hardware photos — but zero performance data. For a sector that lives and dies by transparency, that’s a red flag. As someone who audited 12,000 NFT transactions in 2021 to verify floor prices, I know that seeing is not believing. Data is the only truth.
Context: Why Now?
The bull market of 2025 is euphoric. AI agents are executing crypto transactions autonomously. Decentralized compute networks like Render, Akash, and io.net are swarming with demand. Every AI chip startup with a white paper is getting funded. Etched’s core selling point is LVI (Low Voltage Inference) technology, which claims to run AI inference at dramatically lower voltages, enabling trillion-parameter sparse Mixture-of-Experts (MoE) models to achieve over 80% of theoretical peak performance. That’s a big number. But in a market where hype often masks technical debt, the community is right to be skeptical. The context isn’t just Etched — it’s the entire pattern of “we’ll release benchmarks later” that has plagued hardware startups since the dawn of silicon. Remember when Graphcore promised 100x speedups? Or when Cerebras’s wafer-scale chip was supposed to change everything? The pattern is familiar. The difference this time is that Etched has actual customers — Jane Street received a full rack last month — and credible journalists like The Wall Street Journal and Reuters have confirmed chip shipments. The chips exist. The question is whether they perform as advertised.
Core Key Facts and Immediate Impact
Let’s dissect the numbers. Etched claims its LVI technology allows chip utilization of 80% for sparse MoE models. Chip designer Wesley Yue raised a critical point: high MFU (Model Floating Utilization) does not equal strong absolute performance. MFU is the ratio of actual computation to theoretical peak. If the chip’s peak performance is low, even 80% can be weak. Imagine a car that can only go 50 mph, but you drive it at 40 mph — 80% utilization, but still slower than a car that can go 100 mph at 50% utilization. Etched hasn’t disclosed its raw FLOPs, power consumption, or third-party benchmarks. Their website still says, “Early customer tests have reached leading levels,” with detailed performance data “promised for future release.” That’s not good enough for a $21 billion company.
I’ve been covering crypto hardware since the 2018 ASIC wars. The same pattern repeated: Bitmain claiming 14 TH/s for T9+ miners, only for independent tests to reveal 10% lower performance. Trust bridge crossed. Crash imminent. The immediate impact of this scrutiny is twofold. First, it undermines confidence in the AI-crypto infrastructure narrative. If Etched’s chips can’t deliver, then decentralized AI applications that rely on them — from autonomous trading agents to on-chain ML models — face a performance bottleneck. Second, it sets a dangerous precedent: if a $21 billion company can hide benchmarks, what about smaller projects? The market is pricing in hope, not physics.
Trust bridge crossed. Crash imminent.
But here’s where the contrarian angle emerges. The biggest assumption in the criticism is that Etched is hiding bad performance. What if the opposite is true? What if the chips are actually good, but the company is playing it safe by not releasing data until they have a competitive moat? In the world of AI hardware, releasing benchmarks early can arm competitors. If Etched’s LVI technology is truly revolutionary, tipping off Nvidia, AMD, or even Google would be suicidal. The silence might be strategic, not deceptive.
Yet, the crypto community’s trauma from Terra Luna, FTX, and countless rug pulls makes us see smoke where there might be fire. My own experience during the 2022 Terra Luna collapse — where I coordinated 15 journalists to create a Red Flag List of fraudulent recovery tokens — taught me that skepticism is a survival mechanism. But it also taught me that real innovation can be buried under FUD. The Jane Street deployment is a strong signal. Jane Street is not a venture gambler; they are a quantitative trading firm that runs production workloads. If they received a full rack, they’ve done their own internal validation. The fact that they haven’t leaked numbers suggests either NDA restrictions or genuine confidence.
Liquidity gone. Run. — but not yet. The real issue is the lack of a verifiable, open-source benchmark framework. In crypto, we have zero-knowledge proofs and on-chain verification. Why can’t Etched run a third-party audit of its chip performance on a standardized AI inference task, with cryptographic attestation of the results? That would be the ultimate trust bridge. The fact that they haven’t done so is alarming, but not damning. The biggest question now is not whether the chips exist, but whether they are as powerful as advertised. And the answer will determine the future of decentralized AI compute infrastructure.
Takeaway: What to Watch Next
Watch for three things. First, independent benchmarks from third-party labs like MLPerf or SPEC — if Etched submits, we’ll know. Second, the Jane Street deployment results — if they expand their order, that’s a vote of confidence. Third, the regulatory angle: the SEC’s new crypto-AI oversight framework might require hardware disclosures for projects claiming “decentralized inference.” If Etched remains opaque, the market will correct. If they deliver, the bull case for AI-crypto convergence strengthens. Either way, the truth is coming. And when it does, the floor price of trust will be the first to move.