Hook
A $700 million funding round. A $21 billion valuation. A chip that claims to run trillion-parameter sparse MoE models at over 80% theoretical peak. Yet the only data Etched has publicly released is a single photo of a server rack and a promise that 'early customer tests have reached leading levels.'
George Hotz, founder of tiny corp and the mind behind the open-source deep learning framework tinygrad, publicly dissected the gap: 'Many investors, many orders, many hardware photos. But where is the data?'
Trust is a vulnerability we audit, not a virtue. Etched’s narrative is built on a foundation of silence, and silence in the blockchain is louder than the hack. Let’s run the forensic analysis.
Context
Etched is a chip startup specializing in AI inference accelerators. Their flagship technology is LVI (Low Voltage Inference), which allows chips to operate at reduced voltages while maintaining high throughput. The company claims this enables unprecedented efficiency for sparse mixture-of-experts (MoE) models—the architecture behind many modern large language models.
The funding round, led by a consortium of institutional investors including Jane Street, closed at a $21 billion valuation—a figure that places Etched among the most valuable private AI hardware companies in the world. Jane Street received its first complete rack last month and has already begun deployment. Both The Wall Street Journal and Reuters have confirmed that chips have shipped.
But the industry is not buying the narrative. Wesley Yue, a chip designer with a track record of calling out vaporware, raised a critical point: high utilization ratio does not equal strong absolute performance. Model Floating Utilization (MFU) measures the ratio of actual computation to theoretical peak. If the chip’s peak performance is lower than competitors, even 80% MFU is a castle built on sand.
Core
Let’s deconstruct the numbers. Etched claims their LVI technology allows trillion-parameter sparse MoE to achieve 'over 80% of theoretical peak performance.' The phrase is deliberately vague. Theoretical peak of what? The chip’s own FLOPs ceiling? Or the theoretical peak of the algorithm? Most likely the former, which is a standard marketing trick.
During my audit of the 0x protocol back in 2018, I learned that elegant claims become dangerous when they lack external verification. The same principle applies here. MFU is a ratio, not an absolute. If a chip’s peak is 100 TFLOPS, 80% utilization yields 80 TFLOPS. If a competitor chip achieves 200 TFLOPS at 70% utilization, that’s 140 TFLOPS—75% more raw compute. The utilization ratio becomes a distraction.
Etched has not disclosed complete FLOPs, power consumption, or any third-party benchmark results. Their website reads: 'Early customer tests have reached leading levels.' Leading compared to what? The industry has a well-established benchmark suite—MLPerf, SPEC, or even the simpler GPT-3 inference latency tests. None of those are referenced.
In my 2020 DeFi summer analysis of Compound and Aave, I modeled interest rate curves and found that theoretical soundness crumbled under oracle manipulation. Here, the theoretical soundness of LVI technology hinges on chip architecture that no one outside Etched has independently tested. The closest we have is Jane Street’s deployment. But Jane Street is a market maker, not a chip verification lab. Their use case is likely low-latency trading, not large-scale AI inference. The chip they received might be a specialized variant, not the general-purpose MoE accelerator marketed.
Logic dissolves when code meets human greed. The $21 billion valuation is a bet on future performance, not a reward for proven delivery. The lack of public data is not a sign of secret superiority—it is a red flag that the company is hiding its weaknesses.
Contrarian
But let’s pause. The bulls have a point. Etched has shipped hardware to a real, respected institution. Jane Street is not a crypto startup with a whitepaper; they are a quant trading powerhouse that has deployed custom hardware for years. If they accepted the rack, they likely performed internal validation. The fact that the chips are not in a landfill but in a live trading environment suggests that Etched’s technology works at some level.
Furthermore, the chip industry is notoriously secretive. NVIDIA doesn’t publish full architecture details either. The difference is that NVIDIA has a decades-long track record and independent benchmarks from MLPerf. Etched is a new entrant. The silence could be a strategic choice to avoid revealing trade secrets to competitors like Cerebras or Groq.
However, the valuation gap is the problem. A $21 billion valuation demands transparency. NVIDIA’s market cap is built on thousands of benchmarks, customer testimonials, and a decade of beating expectations. Etched is asking the market to trust them based on a single rack delivery and a press release. Trust is a vulnerability we audit, not a virtue. The market is currently over-indexing on the funding round and ignoring the data vacuum.
Takeaway
The question is not whether the chips exist—they do. The question is whether they are as powerful as advertised. Etched’s silence on FLOPs, power, and benchmarks is a ticking clock. Every summer has a winter of truth. When the next generation of chips from NVIDIA, AMD, or custom ASICs arrives, Etched’s 80% utilization will need to be measured against absolute performance. If the numbers don’t hold, the $21 billion valuation will be remembered as the biggest hardware bluff since Theranos.
Interoperability is the illusion of safety. Here, the only interoperability is between marketing and capital. The data will come. The only question is whether it will confirm the narrative or expose the gap.