A $21 billion valuation. A $700 million raise. And zero public benchmarks.
That's Etched's equation. A chip startup that promises to be the next leap in AI inference โ running trillion-parameter sparse MoE models at over 80% theoretical peak โ yet refuses to show the receipts. The industry is whispering. George Hotz, founder of tiny corp and creator of the open-source deep learning framework tinygrad, didn't whisper. He shouted: "Investors, orders, hardware photos โ but where's the data?"
It's a familiar pattern. In crypto, we call it a "narrative before proof." In hardware, it's called a $21 billion bet.
s fragmented logic. The chip industry, like crypto, runs on stories. But stories without data are just fiction with a funding round.
Context: The Hype Cycle of Hardware
Etched's core pitch is LVI โ Low Voltage Inference. A technology that lets chips run AI inference at lower voltages, dramatically reducing power consumption while maintaining high throughput. For the hyperscalers and AI labs drowning in energy costs, that's the holy grail. The company claims its chips can achieve over 80% Model Floating Utilization (MFU) on sparse MoE architectures โ a metric that measures how much of the chip's theoretical peak is actually used during computation.
But here's the catch: MFU is a ratio. If your chip's peak performance is mediocre, 80% of mediocre is still mediocre. Wesley Yue, a chip designer, raised this exact point: "A high utilization ratio does not necessarily indicate strong absolute performance." It's like boasting that your car uses 90% of its engine power โ but the engine is a lawnmower motor.
Etched's website still reads: "Early customer tests have reached leading levels." Leading levels. Not specific numbers. Not FLOPs. Not power consumption. Not third-party benchmarks. Just... leading levels.
In crypto, we'd call that a "roadmap with no GitHub." In hardware, it's a $21 billion valuation.
Core: The Technical Gap โ What We Know vs. What We Need
Let's dissect the claims. Etched says its LVI technology allows sparse MoE models โ think Mixture of Experts, where only a subset of parameters are activated per token โ to hit >80% MFU. That's impressive if true. Sparse MoE is notoriously hard to optimize because of the irregular memory access patterns. Most chips see MFU drop to 30-40% on these workloads.
But here's the problem: we don't know the peak FLOPs. If the chip's theoretical peak is 100 TFLOPS, 80% MFU gives 80 TFLOPS. If NVIDIA's H100 peaks at 2000 TFLOPS (with sparsity), 80% MFU on H100 gives 1600 TFLOPS. Even at 50% MFU, H100 crushes Etched's 80%.
Etched hasn't disclosed its peak FLOPs. Not even a range. They've shown photos of racks โ Jane Street received its first complete rack last month and has begun deployment. Both The Wall Street Journal and Reuters have confirmed shipments. The chips exist. They're real. But the performance data remains locked inside a black box.
This is where my background kicks in. During the Prague Protocol Audit in 2017, I learned that a smart contract can look perfect โ clean code, audited, deployed โ until you run the edge case. The integer overflow in EtheriumGold's swap function was invisible until you tested at scale. Same principle here: hardware can be shipped, racked, and powered on, but the real question is whether it delivers on its promised inference speed under real-world conditions.
Based on my audit experience, I'd never trust a protocol that releases only a single metric (MFU) without the full context. And I'd never trust a chip company that releases no metrics at all.
Contrarian: The Silence Might Be Strategic โ But It's Also a Red Flag
Here's the contrarian angle: maybe Etched is holding back data because they're protecting a competitive advantage. In the chip industry, revealing detailed benchmarks can give competitors โ especially giants like NVIDIA, AMD, or Google โ a roadmap to replicate your architecture. If LVI is truly revolutionary, keeping it under wraps until widespread deployment makes sense.
But that logic only holds if you're already shipping to major customers. Jane Street is deploying. Early customers are testing. Yet no third-party validation. In crypto, that's like having a DeFi protocol with $10 billion TVL but no public audit. Would you trust it?
Another possibility: the 80% MFU claim is technically true but misleading. Perhaps it's measured on a specific synthetic workload designed to maximize the metric โ not on real-world inference tasks. We've seen this in crypto: projects claiming "100,000 TPS" on a single-node testnet. It's not a lie. It's a carefully chosen framing.
But here's the real blind spot: the market is treating Etched like a narrative asset, not a hardware company. The $700 million raise at $21 billion valuation implies a 30x multiple on potential future revenue โ typical for a unicorn with proven traction. Etched has proven chips exist, but not their performance. The valuation is based on belief, not data.
Sound familiar? It's the same dynamic that drove SOL to $200 in 2021 and LUNA to $120. Belief precedes data. Then data catches up โ or doesn't.
Takeaway: The Next Narrative โ Verification as a Service
What does Etched's silence mean for the broader AI-crypto convergence? It signals a massive opportunity for on-chain hardware verification. If chips are the new scarce resource (compute, memory, bandwidth), then trustless attestation of their performance becomes a crypto-native use case.
Imagine a decentralized network where chips self-report benchmarks via zero-knowledge proofs โ proving they ran a specific inference workload at a specific FLOP rate without revealing the model weights. That's the intersection of AI and crypto that Etched's opacity is inadvertently highlighting.
The question isn't whether Etched's chips are real. They are. Jane Street has them. The question is whether they're as powerful as advertised. And until we see independent benchmarks, we're all just betting on a narrative.
In crypto, we learned that narratives collapse when data arrives. In hardware, the same rule applies. The only difference is the time horizon โ and the size of the bags.