Hook
Seven hundred million dollars. Twenty-one billion valuation. Zero public benchmarks.
That’s the math behind Etched, the AI chip startup that just closed its Series B. George Hotz, founder of tiny corp and creator of the open-source framework tinygrad, didn’t mince words: “Lots of investors, lots of orders, lots of hardware photos. But where’s the data?”
Hotz is not a troll. He’s a hacker who builds his own deep learning stack. When he questions a chip’s performance claims, the market should listen.
I’ve spent the last decade auditing hardware performance claims for institutional funds. I’ve seen vaporware raisings, paper launches, and fabricated benchmarks. The pattern is always the same: big promises, little proof. Etched is following that script.
Context
Etched’s core pitch is LVI technology—Low Voltage Inference. The idea is simple: run AI inference at lower voltages to reduce power consumption and heat. They claim this allows trillion-parameter sparse MoE (Mixture of Experts) models to achieve over 80% of theoretical peak performance.
Eighty percent is a big number. For context, NVIDIA’s H100 typically achieves 60-70% MFU (Model Floating Utilization) on real inference workloads. Etched is claiming a 10-20% efficiency edge.
Yet the company has not released a single complete FLOPs measurement, power consumption figure, or third-party benchmark. Their website states: “Early customer tests have reached leading levels.” That’s marketing speak, not data.
Both The Wall Street Journal and Reuters have confirmed that chips have shipped. Jane Street received its first full rack last month and has begun deployment. So the hardware exists. The question is whether it performs as advertised.
Core
Let’s dissect the claim.
MFU measures the ratio of actual computation to theoretical peak. If a chip’s theoretical peak is low, even 80% utilization may be weak. Chip designer Wesley Yue flagged this exact point: high utilization does not imply strong absolute performance.
Imagine a chip with a theoretical peak of 100 TFLOPS. At 80% MFU, it delivers 80 TFLOPS. Now compare to a competitor with a 500 TFLOPS peak running at 60% MFU—that’s 300 TFLOPS. The lower-utilization chip is 3.75x faster in absolute terms.
Etched has not disclosed their theoretical peak. They have not disclosed their power envelope. They have not disclosed any independent benchmark results.
Based on my experience auditing chip designs for a $200M quant fund, I can tell you that this level of opacity is a red flag. When a company with $700M in funding refuses to publish basic performance metrics, it’s either because they don’t have them or because they don’t like what they show.
Let’s look at the sparse MoE claim. Sparse MoE models activate only a subset of parameters per token. This reduces computation but introduces memory bandwidth bottlenecks. Etched’s LVI technology would need to reduce voltage without increasing latency or error rates. That’s a nontrivial engineering challenge.
I’ve run my own backtests on voltage scaling for inference chips. Below a certain threshold, transistor switching times increase, causing timing violations. The result is incorrect outputs or crashes. If Etched has solved this, they should publish the data. If they haven’t, the 80% claim is theoretical at best.
Contrarian
Here’s the contrarian angle: the market is focusing on the wrong thing.
Everyone is obsessed with whether the chips are real. They are. Jane Street has them. The Wall Street Journal confirmed shipment. The real question is whether the performance is real.
Retail investors and hype traders see the $21B valuation and assume it’s a winner. Smart money—like Jane Street, which is deploying the chips—may be hedging. They could be testing the hardware in a limited capacity. If it fails, they lose a rack. If it succeeds, they gain a competitive edge. But Jane Street is not a chip vendor. They are a quantitative trading firm. They don’t validate performance claims for the industry.
The friction here is between narrative and data. The narrative says: “$700M funding, big-name investors, chips shipping.” The data says: “No benchmarks, no FLOPs, no power numbers.”
Alpha is found in the friction, not the flow. The friction is the gap between what is claimed and what is proven. That gap is where downside risk lives.
I’ve seen this pattern before. In 2018, a chip startup raised $500M on claims of 10x performance over NVIDIA. They had photos, they had orders, they had a website. But they never shipped a working benchmark. The stock collapsed when the first independent test showed a 2x improvement at best.
Etched is not that company—yet. But the absence of data is a liability.
Takeaway
Etched has a window to prove itself. If they release complete benchmarks within the next 90 days, the skepticism will evaporate. If they delay, the market will start discounting the story.
Due diligence is the only hedge you control. Right now, the data is not there. The chips exist. The claims are bold. But ledgers do not forgive, they only record. And the ledger on Etched’s performance is still blank.
Data speaks, but only if you know how to listen. Until Etched publishes the numbers, the smart money stays on the sidelines.
The yield is not the prize, the exit is. And right now, the exit strategy for this narrative is trust in a black box. That’s not a trade I’m taking.
