The chip industry runs on a simple premise: you can’t fake physics. Yet here we are, staring at a startup called Etched, which just closed a $700 million round at a $21 billion valuation, and the most telling signal isn’t the funding — it’s the absence of data. George Hotz, the hacker who built the open-source deep learning framework tinygrad, publicly called out the company on X: "Where are the benchmarks? Show me the throughput, latency, and power under load." He’s not alone. Multiple chip designers, including Wesley Yue, have pointed to the same gap. Etched’s core claim — that its LVI (Low Voltage Inference) technology can run trillion-parameter sparse MoE models at over 80% of theoretical peak — is a bold statement. But without a single third-party benchmark, without FLOPs numbers, without power consumption figures, that statement is just noise. The contrast is stark: Jane Street received its first full rack last month and has begun deployment. The chips exist. The question is whether they’re as fast as advertised. Logic is binary; performance claims are not. Let’s dig into the technical architecture, the metric sleight-of-hand, and the real risk hiding in plain sight.
To understand why Hotz’s skepticism matters, you need to see the context. Etched is not designing a general-purpose GPU or a CPU. It’s building an ASIC (Application-Specific Integrated Circuit) optimized for inference — specifically for transformer-based models with sparse Mixture of Experts (MoE) architectures. The LVI technology is the centerpiece: by lowering the operating voltage, Etched claims it can reduce power consumption while maintaining high clock speeds, effectively boosting the compute-to-watt ratio. In theory, this is a well-known technique in chip design. Lower voltage reduces dynamic power quadratically (P = C V² f), but it also increases the risk of timing errors and signal integrity issues. Etched’s patent filings suggest they use a combination of error-correcting codes and adaptive voltage scaling to mitigate this. The engineering is plausible. But there’s a gap between "plausible" and "80% of theoretical peak on a trillion-parameter model." The theoretical peak of an ASIC is determined by its clock frequency, number of compute units, and the precision of arithmetic operations. If Etched’s chip has a lower peak FLOPs than an NVIDIA H100 or a Cerebras Wafer-Scale Engine, even 80% utilization won’t save it. This is where Model Floating Utilization (MFU) becomes a dangerous metric. MFU measures the ratio of actual computation to the chip’s theoretical peak. If the peak is low, 80% might still be a weak absolute number. Based on my experience auditing hardware-software interfaces in blockchain nodes, I’ve seen teams use MFU to mask poor absolute performance. It’s a classic trap: "Our chip achieves 80% utilization" sounds impressive until you realize the competitor’s chip achieves 70% on a peak that’s 3x higher. The real question is not utilization — it’s sustained throughput in teraflops per watt.
Let’s break down the numbers we don’t have. Etched’s website still says "Early customer tests have reached leading levels" with a promise to release detailed performance data "in the future." That’s a red flag. In the chip industry, leading companies publish benchmarks early — often before tape-out — to build ecosystem confidence. NVIDIA releases MLPerf results. Intel publishes SPEC scores. Cerebras shares training throughput. Silence is a choice. The Wall Street Journal and Reuters have confirmed that Etched has shipped hardware; Jane Street has a rack. But shipping hardware is not the same as validating performance. I’ve seen too many projects — both in crypto and hardware — where the first units work but fail to scale. The lack of public FLOPs, power consumption, or latency numbers means we cannot independently verify the LVI claim. The most charitable interpretation is that Etched is still optimizing the software stack. The least charitable is that the chip’s real-world performance is significantly below the marketing narrative. The market is betting on the latter: the $21 billion valuation implies a belief that Etched will disrupt the AI inference market. But the data suggests otherwise. If the chip were truly revolutionary, why not release a single benchmark? Why rely on a vague "leading levels" statement? The answer, as always, lies in the incentives. Etched raised $700 million from investors who want to see returns. A public benchmark that shows mediocre results would tank the valuation. So they delay. They hope the deployment at Jane Street will generate organic metrics. But for a technical analyst, the absence of evidence is evidence of absence.
Here’s the contrarian angle: the chip might actually be good, but the performance claims are misleading in a more subtle way. Consider the "trillion-parameter sparse MoE" claim. Sparse MoE models activate only a subset of parameters per token, which reduces the effective compute per inference. Etched’s LVI could theoretically exploit this sparsity to achieve higher utilization. But "over 80% of theoretical peak" is a moving target. Theoretical peak is defined for dense operations; for sparse operations, the peak is lower because memory bandwidth becomes the bottleneck. If Etched is measuring utilization against a sparse-optimized theoretical peak, then 80% is less impressive. The real benchmark should be against a dense model on a standard GPU. Without that comparison, the number is meaningless. Furthermore, the chip industry has a history of "vaporware" — startups that announce stellar specs but fail to deliver. Think of the cryptocurrency mining ASIC boom: many companies took orders, shipped hardware, and then the chips underperformed by 30-40%. Etched’s situation is eerily similar. The funding, the orders, the hardware photos — none of it proves the chip works as advertised. The only way to resolve this is a third-party audit. As a smart contract architect, I’m used to verifying code before trusting it. The same standard applies here. Without a public, reproducible benchmark, Etched’s claims remain unverified.
So what’s the takeaway? The ultimate fate of Etched depends on whether they can produce data within the next six months. The market is in a sideways consolidation phase — capital is scarce, and investors are demanding proof. If Etched delivers a benchmark that shows 2x the throughput of an H100 at half the power, the $21 billion valuation will look prescient. If they don’t, the valuation will collapse. The biggest risk is not that the chips are fake — they clearly exist — but that they are mediocre. A chip that only matches the H100 in performance but costs less to manufacture could still be a successful product, but it wouldn’t justify a $21 billion valuation. The market is pricing in a breakthrough. The next few months will reveal whether that breakthrough is real or just another silicon mirage. Until then, the only honest position is skepticism. Logic is binary; intent is often ambiguous. The data will tell us which is which.