Hook
On August 18, 2024, a senior researcher from NTT Data—Japan’s largest IT services conglomerate—publicly declared that Nvidia’s AI compute monopoly is a bubble destined to burst within three years. The trigger? A “new mathematical theory” that would slash compute demand by a factor of a million. The immediate beneficiary? Storage chip makers. The subtext? A seismic shift in the AI value chain from GPU hardware to data infrastructure. As a Web3 research partner who has spent a decade dissecting narratives from ICO whitepapers to DeFi yield farms, I’ve learned one thing: when an incumbent insider starts selling the “paradigm shift” story, it’s either brilliant foresight or a desperate hedge. This time, the evidence leans heavily toward the latter.
Context
Dr. Wang Jiange, chief researcher at NTT Data, laid out his thesis in a widely circulated interview: current large language models are “black boxes” lacking efficient mathematical descriptions, forcing compute waste. He compared this to Newton’s three-parameter description of an apple falling—implying that AI’s data hunger is a symptom of missing theory, not a fundamental need. His conclusion: within three years, a breakthrough will reduce AI compute requirements by millions of times, popping Nvidia’s valuation bubble. Storage players like Montage Technology and ChangXin Memory (CXMT) would emerge as winners, insulated from the compute cycle.
I’ve seen this narrative before. In 2017, ICOs promised that “blockchain would replace all centralized systems”—a sweeping claim that collapsed under its own weight. In 2021, NFT “blue chips” were deemed eternal until the market realized royalties were a feature, not a business model. The pattern is clear: the more specific the prediction (three years, million-fold reduction), the higher the probability it’s a marketing tool rather than a falsifiable hypothesis. Yet, Dr. Wang’s position at NTT Data—a company that competes with Nvidia in the broader IT service ecosystem—adds a layer of institutional motive worth unpacking.
Core
1. The Category Error at the Heart of the Thesis
Dr. Wang’s core analogy—that Newton described gravity with three parameters while AI needs billions of images—is a classic category error. Newton’s parameters describe a closed-form physical law. AI’s task is not to describe a single phenomenon but to generalize across language, vision, and reasoning in unseen contexts. The complexity of representation learning is fundamentally different from the complexity of a physical law. As I wrote in my 2022 report on scaling laws: “The number of parameters is not a measure of inefficiency; it’s a measure of the manifold’s intrinsic dimensionality.”
2. Scaling Laws Are Not a Faith—They’re a Measurement
Since 2020, every major AI lab has validated the power-law relationship between compute, data, and model capability. Even the shift to “smaller models + inference-time compute” (e.g., DeepSeek R1, OpenAI o-series) does not reduce total compute demand—it reallocates it. The total industry compute budget continues to grow. The claim that a mathematical breakthrough could reduce demand by “a million times” lacks any derivation. Comparing it to quantum chemistry’s leap from exponential to polynomial complexity is misleading because nature itself is quantum. Language and common sense may not obey a simple mathematical law—that’s an untested hypothesis, not a fact.
3. The Power Bottleneck Is Real, but It Doesn’t Support the Million-Fold Claim
Dr. Wang correctly identifies power as a constraint. AI data centers could consume over 1,000 TWh annually by 2026—roughly Japan’s total electricity use. Grid constraints in Northern Virginia and Silicon Valley are real. However, this bottleneck will slow demand growth, not cause a cliff. In fact, constrained supply often inflates GPU prices and valuations, the opposite of a bubble burst. The “million-fold” reduction would require not just a new theory but a complete rewrite of von Neumann architecture—a feat that has no precedent in the 80-year history of computing.
4. The Storage Thesis Is Overly Simplistic
Dr. Wang’s bet on storage chips (Montage, CXMT) assumes data growth is independent of compute cycles. But storage is a highly cyclical industry: DRAM prices fell 50% in 2022-2023, and HBM—a key AI memory type—is directly tied to GPU demand. If AI compute collapses, HBM demand collapses with it. The “storage as safe haven” argument ignores that storage is a commodity with low margins and high volatility. I’ve audited supply chains for DeFi projects; the same principle applies: “sell picks and shovels” only works when the gold rush is real. If the gold is fake, the shovel market crashes too.
5. The Three-Year Window: A Marketing Device
Precise timing predictions in technology are almost always wrong. Dr. Wang’s three-year window aligns with typical institutional investment cycles (3-5 years) and maximizes media attention. It also conveniently avoids the near-term (6-18 months) where counter-evidence could emerge. This is a classic “narrative arbitrage” play: say something bold enough to get quoted, knowing that if it doesn’t happen, the audience will have forgotten.
Contrarian
What If the Bubble Bursts Differently?
Let me offer a contrarian angle that Dr. Wang’s analysis misses: the real risk to Nvidia is not a mathematical revolution but a gradual erosion of its moat through custom silicon (Microsoft Maia, Google TPU, Amazon Trainium, Meta MTIA). This is already happening. Over the next 3-5 years, hyperscalers will shift 30-50% of their training workloads to in-house chips, compressing Nvidia’s margins from 75%+ to around 60%. This is a “slow bleed,” not a “bubble burst.” Storage will benefit from the general increase in data, but not enough to offset the sector-wide capex slowdown when hyperscalers optimize their spending.
The Blockchain Angle
From a Web3 perspective, a collapse in GPU pricing would be a boon for decentralized compute networks like Render Network, Akash, and io.net. Lower GPU costs mean lower barriers to entry for decentralized AI inference. However, storage chains like Filecoin and Arweave would face a double-edged sword: more data from AI applications, but also more competition from centralized storage that becomes cheaper as hardware prices drop. The net effect is nuanced.
The “Narrative Trap”
Dr. Wang’s thesis is itself a narrative product. It fits the “incumbent warns of disruption” archetype that has historically been wrong more often than right (think of IBM dismissing the PC, or Kodak warning about digital cameras while trying to protect film). The three-year window gives it a shelf life. As a narrative hunter, I see this as a signal that the market is saturated with bearish takes, which is often a contrarian buy signal for the underlying asset (Nvidia) in the short term.
Takeaway
Don’t bet against Nvidia’s moat based on a theoretical breakthrough that has no empirical basis. Do watch for the gradual erosion from custom silicon and the organic slowdown from power constraints. The architecture of trust is built, not inherited—and Nvidia’s trust in its CUDA ecosystem is still intact. The storage play is a hedge, not a home run. The real alpha lies in identifying which AI infrastructure narratives will survive a normal cyclical correction, not a fantasy apocalypse.