Hook: The Signal from an Insider Skeptic
When NTT Data's Chief Researcher, Professor Wang Jiange, publicly declared that Nvidia's AI bubble would burst within three years—citing a missing mathematical framework that would slash compute demand by millions of times—the market took notice. The article, published via Phoenix Finance around August 2024, represents a high-profile warning from a traditional IT services giant. But as a battle-tested trader who has survived ICO audits, DeFi arbitrage, and the LUNA collapse, I’ve learned one thing: ledgers don’t lie. The data tells a different story, and the assumptions behind this prediction are dangerously simplistic.
Context: NTT Data’s Institutional Position
NTT Data is Japan’s largest IT services firm, a system integrator with deep interests in cloud, storage, and legacy infrastructure. Its core business does not revolve around GPU compute leasing—unlike cloud hyperscalers. When a senior researcher from such a firm publishes a three-year doomsday scenario for Nvidia, it’s rarely pure altruism. It’s a strategic narrative shift: redirect the AI narrative from compute to storage and basic research, where NTT Data’s comparative advantage lies. The article’s timing—mid-2024, when Nvidia’s market cap flirted with $5 trillion—amplifies its impact. Yet, the core thesis—that a new mathematical theory will reduce compute demand by millions of times—lacks empirical grounding. My own experience auditing smart contracts in 2017 taught me that precise claims require precise evidence. This one has none.
Core: The Flawed Logic of Mathematical Salvation
Professor Wang’s argument hinges on a category error. He compares the three parameters needed to describe an apple falling under Newtonian physics to the billions of images used to train a large language model. The implication: that a better mathematical framework could compress the latter into a few parameters. But this confuses the complexity of describing a physical phenomenon with the complexity of learning universal representations. The task of an LLM isn’t to describe a single event—it’s to generate coherent language, images, and reasoning across countless unseen contexts. The scaling laws empirically validated over the past five years show that model capability improves predictably with parameter count, data, and compute. OpenAI, Anthropic, and Google have all confirmed this. Even the recent shift toward small models plus inference-time compute (e.g., DeepSeek R1) only redistributes compute, not reduces total demand.
Furthermore, the “new mathematical framework” remains a theoretical fantasy. Academic efforts like state-space models, linear attention, and hypergraph networks optimize within the existing paradigm. They don’t rewrite the language of intelligence. A reduction of compute demand by “millions of times” would require a breakthrough comparable to quantum mechanics for chemistry. But language and common sense may not obey a simple hidden law. The probability of such a breakthrough within three years, based on my assessment of technological history, is below 5%. I’ve seen too many claims of paradigm shifts fizzle out. Risk is not a variable, it is a constant—and betting on a miracle is a losing strategy.
Contrarian: Who Really Wins When the Bubble Bursts?
The article recommends storage chips (e.g., Montage Technology, CXMT) as safe havens, arguing data volumes will only grow. But this ignores the cyclical nature of storage. The 2022-2023 storage crash wiped out billions in market cap. Moreover, if AI compute demand collapses, HBM memory—a storage product tied directly to AI servers—will suffer. The broader storage sector is not immune to a tech downturn. The real beneficiaries of an AI compute crash would be alternative compute providers (AMD, custom ASICs, cloud providers with self-designed chips) and AI application layers that gain from lower inference costs. Also, the power infrastructure sector would face stranded assets—despite the article’s omission. Yield is the tax on your ignorance: chasing storage as a “safe” play without understanding its cyclicality is a recipe for losses.
Takeaway: Positioning for the Chop
Sideways markets reward precision, not prophecies. The NTT Data article is a sentiment indicator, not a trading signal. It reflects growing skepticism among legacy IT players, but it doesn’t invalidate Nvidia’s moat—CUDA, NVLink, and developer ecosystem. The most likely scenario is a gradual margin compression from 75%+ to 60%+ over 2-3 years, driven by competition and supply normalization, not a collapse. My advice: do not short Nvidia purely based on a three-year prediction. Instead, build a multi-dimensional AI portfolio that balances compute, storage, application, and power. Survival precedes profit in every cycle. The blockchain remembers what you forget, but the market remembers what you overbet.