A $21 billion valuation. A $700 million funding round. Zero public benchmarks. This is the signal the market is ignoring.
Etched, the AI chip startup claiming to revolutionize inference with its LVI (Low Voltage Inference) technology, has shipped hardware to Jane Street. The Wall Street Journal and Reuters confirm the racks are deployed. But the data sheet is empty. No FLOPs. No power consumption numbers. No third-party MLPerf results. Just a website stating, “Early customer tests have reached leading levels.”
For a quantitative strategist who has spent years auditing cryptographic hardware and on-chain data, this is not a red flag. It is a red ocean. In the blockchain world, we demand verifiable proofs. Code is law; hype is just noise. The chip industry should be held to the same standard.
Context: The LVI Promise and the MFU Mirage
Etched’s core selling point is LVI—a technique that allows chips to run AI inference at significantly lower voltages. The claim: trillion-parameter sparse Mixture-of-Experts models can achieve over 80% Model Floating Utilization (MFU). MFU measures the ratio of real computation to theoretical peak. If true, this would be a leap forward. Sparse MoE models are notoriously memory-bound; high utilization implies near-perfect hardware-software co-design.
But here is the catch Wesley Yue, a chip designer, rightfully pointed out: MFU is a ratio, not an absolute. A chip with a low theoretical peak can hit 80% utilization and still be outrun by a competitor with 60% utilization but double the raw FLOPs. Without knowing the absolute peak performance, the MFU number is a distraction. It is like claiming a protocol processes 90% of its theoretical TPS without disclosing the base TPS.
George Hotz, founder of tiny corp and the mind behind tinygrad, publicly questioned this. His skepticism is rooted in years of optimizing inference kernels. He saw the lack of data and called it out. The chip industry has a history of “paper launches”—announcements with impressive slides but no silicon. Etched has silicon. Jane Street has it. But that does not confirm the performance narrative.
Core: The On-Chain Evidence Chain (Applied to Hardware)
In my 23 years in the industry, I have learned to treat every claim as a hypothesis until proven by data. When I audited ZK-SNARK implementations in 2017, I did not trust the white papers. I ran the code. I measured gas costs. I found the 12% inefficiency in Groth16 verification. The same principle applies here.
Let me lay out the evidence chain we need to validate Etched’s claims:
- Absolute FLOPs at different precisions: Etched must disclose the TFLOPS for FP16, BF16, and INT8. Without this, MFU is meaningless. If the chip’s peak is 100 TFLOPS, 80% MFU is 80 TFLOPS. If a competitor like NVIDIA’s H100 peaks at 200 TFLOPS with 60% MFU, that is 120 TFLOPS. Etched loses.
- Power consumption under load: LVI’s advantage is lower voltage. But lower voltage often means lower clock speeds. The trade-off is performance per watt, not raw performance. Etched needs to publish watts per chip at full load. If the power draw is 50% of a competitor but the FLOPs are 40%, the efficiency win is real. But if the FLOPs are 20%, the chip is niche.
- Memory bandwidth and latency: Sparse MoE models are memory-bound. Etched’s architecture presumably has high bandwidth memory (HBM). But bandwidth is not enough—latency to access sparse parameters matters. Without benchmarks on real workloads (e.g., Mixtral 8x7B, GPT-4 style MoE), the 80% MFU claim is theoretical.
- Third-party validation: The Wall Street Journal and Reuters confirmed shipments. But journalism is not benchmarking. Jane Street is a quantitative trading firm, not an AI research lab. They may use the chip for inference in financial models, but that is not a proxy for general AI performance. We need independent verification from MLPerf, Stanford’s DAWNBench, or a peer-reviewed paper.
Based on my experience building an on-chain surveillance dashboard for institutional clients, I know that the absence of data is itself a data point. When a startup with a $21 billion valuation refuses to publish benchmarks, it is either because the numbers are not good enough, or they are not ready. Neither inspires confidence.
Contrarian: High Utilization ≠ High Performance
The contrarian angle here is subtle. The crypto community loves a narrative: “This chip achieves 80% MFU, therefore it is the best.” But correlation is not causation. High MFU can be engineered by reducing the theoretical peak. Imagine a chip that only runs one specific workload—a custom ASIC for sparse MoE. It can achieve 90% MFU because the architecture is hyper-specialized. But general-purpose AI requires flexibility. The question is not whether Etched can hit 80% MFU on a narrow test, but whether it can maintain that on diverse models.
Consider the analogy with L2 scaling solutions. Many L2s claim high throughput, but they slice already-scarce liquidity. High transactions per second (TPS) does not mean high value. Similarly, high MFU does not mean high performance if the absolute FLOPs are low. The market is falling for the same trap: focusing on ratios while ignoring the base.
Another blind spot: the funding itself. A $700 million round at $21 billion valuation creates a halo effect. Investors assume the due diligence is done. But venture capital is not technical validation. Many crypto projects raised billions and still failed to deliver. The capital is a signal of market sentiment, not technical correctness. Check the logs, not the tweets.
Takeaway: The Next Signal to Watch
Etched has until the end of Q2 2025 to release third-party benchmarks. If they are confident, they will submit to MLPerf Inference v4.1. If they do not, the silence will speak louder than any press release.
For the quant community, the data is already writing the story. Jane Street’s deployment is a positive sign, but Jane Street is not a typical AI user. They could be using the chip for low-latency trading, not generative AI. We need to see adoption by AI labs: OpenAI, Anthropic, Google DeepMind.
Personally, I will be watching the power consumption numbers. In the crypto winter of 2022, I predicted the stablecoin de-pegging by monitoring oracle dependency. Now, I will monitor Etched’s power efficiency. If the chip delivers 2x the performance per watt of an H100 at 50% the absolute FLOPs, it is a success. If it only delivers 0.8x the performance per watt, the valuation is a bubble.
Until then, the data is incomplete. Treat the $21 billion as a hypothesis, not a fact. Code is law; hype is just noise. The logs will tell the truth.
