A $21 billion valuation. A $700 million funding round. And a chip that, according to its own marketing, can run trillion-parameter sparse MoE models at over 80% of theoretical peak performance. Yet, when George Hotz—the hacker behind tinygrad—publicly asked for hard data, the response was silence. No FLOPs. No power consumption figures. No third-party benchmarks. Just a promise: "Early customer tests have reached leading levels."
This is not a story about a chip company. This is a story about a systemic failure of trust in the tech industry—a failure that those of us in the blockchain space have seen before, and one that we must learn from if we are to build systems that truly earn the confidence of their users.
Context: The Anatomy of a Unicorn
Etched, a startup aiming to build specialized AI inference chips, recently closed a $700 million round at a $21 billion valuation. Its core selling point is LVI (Low Voltage Inference) technology, which the company claims allows its chips to run AI models at significantly lower voltages while maintaining high throughput. The result, they say, is a chip that can achieve over 80% Model Floating Point Utilization (MFU) on sparse mixture-of-experts architectures with billions of parameters.
But here's the catch: MFU is a ratio of actual computation to theoretical peak. If the theoretical peak is low, even 80% utilization may not impress. Chip designer Wesley Yue raised this exact concern. He pointed out that a high utilization ratio does not necessarily indicate strong absolute performance—it could simply mean the chip is optimized for a narrow set of operations while being weak elsewhere.
To date, Etched has not disclosed complete FLOPs, power consumption, or third-party benchmarks. Both The Wall Street Journal and Reuters have confirmed that chips have been shipped—Jane Street received its first complete rack last month and has begun deployment. But shipping is not the same as proving. The biggest question remains: are the chips as powerful as advertised?
Core: The Code of Conscience, Revisited
I have seen this pattern before. In 2017, during the ICO boom, I spent four months auditing the smart contracts of a popular fundraising platform called EtherTrust. I discovered a reentrancy vulnerability that could have drained $4.2 million in user funds. The team's response was to offer me a private bug bounty—if I kept quiet. I published the technical exposé on Medium instead. The decision cost me a lucrative consulting offer, but it established a principle: trust is earned, not mined.
Etched is facing a similar moment. They have raised an enormous amount of capital, they have shipped hardware, and they have a compelling narrative. But they have not provided the data necessary for independent verification. In the blockchain world, we call this a "trust me" model—and it is the antithesis of everything we stand for. A decentralized system does not rely on promises; it relies on provable, auditable outputs. If a chip's performance cannot be independently verified, then its claims are no more credible than a whitepaper that promises a world computer but delivers only a token.
This is not about whether Etched is lying. There is no evidence of fabrication. But the absence of evidence is not evidence of absence. The burden of proof lies with the claim-maker. When a company asks for a $21 billion valuation, it must also ask itself: what data am I willing to release to justify that number?
Contrarian: The Pragmatism Test
One might argue that Etched's approach is standard for the semiconductor industry. Companies often guard their performance data as trade secrets until they have a competitive advantage to protect. And Jane Street, a sophisticated quantitative trading firm, has already deployed the chips—surely they would not do so if the hardware were not performant?
But this argument misses a crucial point. Jane Street's deployment is a single data point, and it is not a public benchmark. The firm's internal use case may be narrow, and its tolerance for risk may be high. For the broader industry to trust Etched's claims, we need more than a single reference customer. We need open, reproducible testing. We need the kind of transparency that allows developers to assess whether the chip fits their specific workloads.
In the blockchain space, we have learned this lesson the hard way. The collapse of Terra Luna, the failure of FTX—these were not failures of technology alone. They were failures of transparency. When a system cannot be audited, trust becomes a gamble. And gambling is not the foundation upon which we should build the future of AI.
Takeaway: Soul in the Machine
Etched's story is not yet over. They may release benchmarks tomorrow that silence all critics. But the episode serves as a reminder that in a bull market—whether for AI or crypto—euphoria can mask technical flaws. We must remain vigilant. We must demand data, not just narratives. We must hold ourselves and our industry to a higher standard.
Because at the end of the day, the soul of the machine is not in its silicon. It is in the integrity of the claims we make about it. And that integrity must be verifiable, not just promised.
Trust is earned, not mined. And the mining process starts with transparency.