Etched's $21B Black Box: Why the Chip Industry's 'Trust Me' Model Is a Security Vulnerability

The arithmetic is simple. $700 million raised. $21 billion valuation. Chips shipped to Jane Street. Yet the industry has not seen a single verifiable benchmark. This is not a startup; it is a cryptographic black box where the output is market cap and the input is narrative. George Hotz, the hacker behind tinygrad, publicly called out the absence of performance data. Wesley Yue, a chip designer, pointed out that Model Floating Utilization (MFU) is a ratio, not an absolute. The truth is: without FLOPs, power consumption, and third-party audits, every claim about Etched's LVI technology is an unverified smart contract running on a live mainnet with billions of dollars at stake. Let me rewind the execution trace. Etched is building a chip specifically optimized for AI inference, with a core selling point called LVI (Low Voltage Inference). The claim is that their hardware can run trillion-parameter sparse Mixture-of-Experts (MoE) models at over 80% of theoretical peak performance. The term 'theoretical peak' is the first red flag. In my years auditing smart contracts, I have seen this phrase used as a lure. It is the equivalent of a whitepaper stating 'the protocol will handle 10,000 TPS' without specifying the number of validators or the network latency. Theoretical peak is a mathematical maximum under ideal conditions—no memory bottlenecks, no IO latency, no thermal throttling. Achieving 80% of that is impressive, but only if the theoretical peak itself is competitive. Consider the context. The chip industry is currently in a frenzy driven by the AI gold rush. Nvidia’s H100 and B200 dominate the market, but they are general-purpose accelerators. Etched aims to build a specialized chip that, by design, cannot run general-purpose workloads. It is an ASIC for transformer inference. This is a high-risk, high-reward bet. If the chip is as fast as claimed, it could undercut Nvidia on cost per inference. But if the performance is inflated, the valuation collapses. The problem is that the market is making a bet based on photos of hardware and quotes from Journalists, not on open-source benchmarks. Hotz's skepticism is not just noise; it is a formal verification of the claims. He asked: where is the data? The chip industry has a culture of publishing technical papers and benchmarks. AMD, Intel, Nvidia all release detailed specs. Etched’s website still says 'Early customer tests have reached leading levels' with promises of future data. That is a non-answer. In my work on smart contract audits, I have a rule: if a function's behavior is not documented, assume it is a bug. The same applies to hardware. If the performance is not benchmarked, assume it is worse than advertised. Let me dive deeper into the technical claim. Sparse MoE models are notoriously difficult to accelerate because they require dynamic routing of tokens to different experts. The LVI technology supposedly allows the chip to run at lower voltages, reducing power consumption without sacrificing clock speed. That is a plausible engineering trade-off: lower voltage reduces power quadratically, but also reduces the maximum stable frequency. The 80% MFU claim implies that the chip is efficiently utilizing its compute units, but the question is: what is the peak FLOPs? If the chip has a peak of 100 TFLOPS, 80% is 80 TFLOPS. If a competitor’s chip has a peak of 200 TFLOPS but only 50% utilization, it still achieves 100 TFLOPS. Etched’s 80% is meaningless without the peak. Based on my experience analyzing smart contract gas costs, I know that efficiency ratios are often used to mask absolute deficits. In DeFi, a protocol might claim '99% gas efficiency' compared to a naive implementation, but the absolute gas cost is still higher than a simpler alternative. The same trick applies here. Etched needs to publish the raw FLOPs for the specific models they target. Additionally, they must publish power consumption. A chip that uses 700W to achieve 80% utilization is less impressive than one that uses 300W for the same work. The market is flying blind. Now, the contrarian angle. The obsession with benchmarks might be a blind spot in itself. The AI chip market is not purely about raw performance. It is about the ecosystem. Nvidia benefits from CUDA, a software stack that has been optimized for years. Etched’s chip is a new architecture. Even if the hardware is 2x faster, the software toolchain may be immature, creating a hidden tax. However, the market is currently ignoring this and focusing on the hardware claims. The real blind spot is that the chip industry's due diligence process is broken. Investors are relying on the reputation of founders and the presence of hardware photos as evidence. This is no different from the crypto market where a team with a white paper and a testnet attracted billions. The market is repeating the same pattern: trust the narrative, verify later. There is no evidence that Etched has fabricated its claims. The Wall Street Journal and Reuters confirmed that chips have shipped. Jane Street received a full rack and is deploying it. That is real. But shipping a chip is not the same as shipping a chip that performs as advertised. Many hardware startups have shipped early versions that underperformed and were later fixed. The question is whether the current valuation reflects the potential or the reality. Let me apply a mental model from smart contract security: the 'worst-case execution path'. In an audit, we assume the attacker will find the most vulnerable code path. For Etched, the worst case is that the LVI technology works only for a narrow set of models or that the 80% MFU is achieved only under specific conditions (e.g., batch size, sequence length, sparsity pattern). If the chip is optimized for a single configuration, its real-world performance could be much lower. The company has not released the evaluation conditions. This is a classic case of 'unspoken assumptions' becoming a vulnerability. A bug is just an unspoken assumption made visible. Etched's assumption is that the market will take their word for it. The bug is that the market is starting to question that assumption. The chip industry needs to adopt a standard for verifiable performance claims, similar to how DeFi protocols now undergo formal verification. Until then, every chip startup is a smart contract with a backdoor labeled 'trust us'. The takeaway is not that Etched is a fraud. It is that the valuation is a function of faith, not data. For the industry to mature, we need open-source benchmark suites, audited by third parties, with reproducible results. The same way we demand that a smart contract's bytecode matches the source, we must demand that a chip's benchmarks match its advertising. The stack overflows, but the theory holds. Etched's theory is compelling, but the stack currently has a memory leak of unverified claims. Until the benchmarks are published, any investment in Etched is a bet on the founders' ability to deliver, not on the technology. Compiling truth from the noise of the blockchain. In this case, the noise is the hype. The signal will be a single, trustworthy benchmark. Let's wait for that before we assign a $21 billion value.