A Japanese IT giant’s chief researcher just declared Nvidia’s AI compute bubble will burst within three years, citing a missing mathematical theory that could slash compute demand by millions of times. I have seen this story before. In crypto, we called it the EOS supercomputer narrative, the Filecoin storage hype, and the Bitcoin mining rig mania. The same pattern: hardware scarcity masks a deeper structural fragility. Let me audit this thesis through a blockchain lens, because the code here rhymes harder than most realize.
Wang Jiangge, principal researcher at NTT Data, argues that current large models lack an efficient mathematical description tool, forcing an absurdly high compute requirement. His analogy: Newton needed three parameters to describe an apple falling, but neural networks need billions of images to learn the same concept. This is a category error. As a blockchain engineer who spent 18 months modeling modular vs. monolithic architectures, I can tell you that the bottleneck is not the lack of a math tool – it is the fundamental tension between statistical learning and von Neumann architecture. We saw the same in crypto: scaling transactions on Ethereum required more validators, more hardware, more energy. The response was not a new consensus math, but a pragmatic shift to rollups and sharding. The AI industry is approaching a similar inflection point, but the path is not a magical theory – it is incremental efficiency gains that compound.
Liquidity is a mirror, not a foundation. The market is currently pricing Nvidia as if its monopoly will last forever. But the crypto analog teaches us otherwise. When Bitcoin ASICs were scarce in 2017, Bitmain controlled the narrative. Within two years, oversupply and competition crushed margins. Wang’s prediction of a three-year collapse is not backed by any verifiable evidence – he offers no derivation of the “millions of times” reduction, no timeline for commercialization of a new math. But the act of making such a precise forecast is itself a signal. He is betting on a paradigm shift that has no precedent in the history of computing. The most likely outcome is not a sudden collapse, but a gradual erosion of margins as customers build their own chips – just as Ethereum’s shift to PoS slashed power consumption by 99.9% without a new math, just a consensus redesign.
The contrarian angle: decoupling is a myth. The article claims storage chips (like ChangXin Memory Technologies) will thrive regardless of AI compute bust. This is the same logic that pumped Filecoin and Arweave last cycle. But storage is a cyclical commodity, not a long-term growth asset. In 2022, when crypto crashed, storage demand from AI training data did not save the sector – it crashed alongside. The real overlooked beneficiary of an AI compute bubble burst is not storage, but zero-knowledge proof infrastructure. Because ZK proofs can verify computation without re-executing it, slashing the validation cost – exactly the kind of efficiency gain Wang dreams of, but already proven in production (zkSync, StarkNet). In my 2026 strategy, I allocated $5 million into decentralized compute markets (Render, Akash) not because I believe in AI hype, but because the demand for verifiable computation will outlive any single hardware monopoly.
History does not repeat, but it rhymes in code. The NTT Data warning is a useful emotional thermometer for the AI sector, but it is not a trading signal. When a traditional IT giant starts publicly shorting the market leader, it usually means the cycle is late, not over. The real question is not whether the bubble will burst, but which layer of the stack will survive the correction. In crypto, the answer was always infrastructure that provides fundamental utility – not speculative tokens. For AI, the same applies: the compute layer is overvalued, the data availability layer is underappreciated, and the verification layer (ZK) is still nascent. I will not bet against Nvidia in the next 12 months, but I will position for the liquidity rotation that will follow the inevitable margin compression.
The algorithm does not care about your conviction. Whether Wang’s three-year timeline is accurate or not, the market will eventually price in the reality that compute is not infinite. The lesson for crypto investors is clear: do not chase the candle of hardware mania. Study the gravity of protocol design. The most durable assets are those that can function with thin margins – like Bitcoin’s proof-of-work, which will survive any ASIC price collapse because its security budget adjusts dynamically. For AI, that means betting on software stacks that can run on any hardware, not on the hardware itself. I am watching the ZK-Rollup builders and the decentralized compute networks, waiting for the moment when the market panics and sells the picks and shovels. That is when I will buy.
