Most believe a $21 billion valuation validates technical claims. That belief is incorrect.
George Hotz—founder of tiny corp, creator of the open-source deep learning framework tinygrad—publicly dissected Etched’s narrative. His diagnosis: plenty of investors, orders, and hardware photos. But zero data that confirms performance. In an industry where hype often precedes substance, Hotz’s skepticism is a rare signal worth heeding.
Etched recently closed a $700 million funding round at a $21 billion valuation. The company’s core pitch is LVI (Low Voltage Inference) technology, allowing chips to run AI inference at lower voltages. They claim this enables trillion-parameter sparse MoE models to achieve over 80% of theoretical peak performance. Impressive. But as any veteran of hardware audits knows, theoretical peak is a sandcastle waiting for the tide.
Context: The Valuation-Knowledge Gap
Etched’s website still reads: “Early customer tests have reached leading levels.” Detailed performance data is promised for future release. The Wall Street Journal and Reuters confirm chips have shipped—Jane Street received its first full rack last month and began deployment. So the hardware exists. The question is not existence. It is performance.
Chip designer Wesley Yue raised a critical distinction: Model Floating Utilization (MFU) measures the ratio of actual computation to theoretical peak. If the chip’s peak performance is lower than competitors, even an 80% utilization rate may still lose. Utilization is a relationship, not an absolute. Etched’s 80% claim could be a ratio of a modest denominator.
Core: The Performance Data Void
Let’s examine the gaps. Etched has not publicly disclosed:
- Complete FLOPs (floating point operations per second) for their chip.
- Power consumption under load.
- Third-party benchmarks from independent labs.
- Comparisons to Nvidia’s H100 or B200 on standard inference tasks.
In my years auditing hardware claims—from Bitcoin ASIC efficiency to GPU mining rigs—I’ve learned one rule: if a company with $700 million in funding cannot produce a single independent benchmark, the probability of overpromising is high. Not fraud. Overpromising. The difference is intent, but the result is the same: capital misallocation.
LVI technology is legitimate in principle. Lower voltage reduces power draw and heat, enabling denser compute. But the actual performance depends on architecture, memory bandwidth, and software stack. Etched’s sparse MoE advantage requires that the sparsity pattern aligns with their hardware design. Real-world workloads are messy. Sparse models often have irregular memory access patterns that kill throughput.
Contrarian: The Real Trap Is Not Existence, It’s Expectation
Most media coverage frames Etched’s story as a binary: does the chip work or not? That misses the point. The chip works. Jane Street wouldn’t deploy a non-functional rack. But “works” and “outperforms Nvidia” are worlds apart.
Here’s the contrarian angle: Etched’s $21 billion valuation is less about technical superiority and more about a scarcity narrative. The AI chip market is dominated by Nvidia, with AMD and Intel struggling to gain share. Any new entrant that claims better efficiency captures a premium. Scarcity is a narrative; utility is the anchor. Etched’s utility remains unproven.
Yield is the lure; liquidity is the trap. In this case, the yield is the promise of AI inference efficiency, and the liquidity is the massive capital inflow. If the performance data fails to impress, the liquidity will dry up. Investors who bought the narrative without data will be left holding equity in a chip that is merely “good enough” in a market that demands “exceptional.”
Takeaway: Demand Data, Not Photos
Etched’s story is not unique. It mirrors the pattern of many crypto projects that raised billions on whitepapers and mockups. The same due diligence principles apply: verify claims with independent benchmarks, compare against existing solutions, and question the denominator of utilization ratios.
Hype decays; adoption endures. Etched has shipped hardware—that’s adoption momentum. But until they release complete FLOPs, power consumption, and third-party benchmarks, the $21 billion valuation remains a bet on faith, not facts.
For investors, the takeaway is clear: when a company asks for trust without data, the counter-offer should be skepticism. The market will eventually correct. The question is when.
Efficiency hides risk until the pivot breaks. Etched’s pivot will come when they release performance data. Until then, the only rational position is watchful waiting.