NTT Data’s chief researcher Wang Jiange dropped a bomb on August 18 (year unstated, but the context screams 2024–2025): Nvidia’s dominance is a bubble that will burst within three years, driven by a missing mathematical framework that could slash compute demand by millions of times. The market yawned. I didn’t.
As a Nansen-certified analyst who has spent years tracing wallet flows through DeFi, NFT, and now AI-adjacent crypto plays, I’ve learned one thing: Hashes don’t lie. Wallets do. Wang’s thesis is not a prediction—it’s a data point. And when a senior insider at NTT Data, a traditional IT giant that built its empire on system integration rather than GPU leasing, publicly short-sells the narrative, the signal is worth decoding.
Context: The Data Methodology Behind the Warning
Wang’s core argument: current large language models are “black boxes” lacking efficient mathematical descriptions, forcing compute consumption far beyond physical necessity. He draws a parallel to Newton’s three parameters describing an apple falling—a category error. The task of a foundation model is not to describe a single phenomenon but to generalize across language, vision, and reasoning. Scaling laws (model size ↔ data ↔ compute) have held empirically for five years; even the shift to “small models + inference-time compute” (DeepSeek R1, OpenAI o-series) didn’t reduce total compute, it just redistributed it.
But Wang’s real contribution is framing the bottleneck: electricity, not just chips. AI data centers may consume over 1,000 TWh by 2026—roughly Japan’s entire electricity usage. This is a physical constraint no mathematical tool can instantly erase. Yet his claim of a “million-fold” reduction lacks any derivation. From my experience auditing 2017 ICO token distributions, I know that when a prediction is too precise (three years, million-fold), it’s often a rhetorical device, not a forecast.
Core: The On-Chain Evidence Chain
Let’s follow the liquidity, not the narrative. On-chain data from several angles supports the “bubble” label but contradicts the “immediate collapse” timeline.
1. GPU-Linked Token Flows:
Tokens that track GPU compute—like Render Network (RNDR) and Akash (AKT)—showed a clear correlation with Nvidia’s stock price from 2023 to early 2025. But in Q2 2025, a divergence emerged: Nvidia’s stock continued climbing while these tokens’ on-chain volume stagnated and wallet accumulation slowed. This suggests that the marginal buyer of AI-infrastructure tokens (often retail forced by hype) has become exhausted. Institutional flows into Nvidia, however, remained strong (BlackRock IBIT ETF inflows correlated with OTC desk sales, a pattern I documented in 2024). Insider moves in silence. Watch the gas.
2. Storage-Play Token Anomalies:
Wang recommends storage chips (Montage, CXMT) as the “deterministic winner.” But on-chain data for storage-related tokens (e.g., Filecoin, Arweave) tells a different story: Filecoin’s storage utilization rate barely broke 30% despite AI hype. The narrative that “AI data will fill all storage” is already priced in. When I traced the 2021 BAYC insider wallet cluster, I saw the same pattern—early adopters front-run the narrative. Today, storage token wallets show a high concentration of long-term holders who bought in 2020–2021, not new capital flowing in.
3. Electricity Futures and Mining Derivatives:
Bitcoin mining has historically been a proxy for compute demand. The hash price (revenue per TH/s) has been declining since April 2025, even as AI compute demand supposedly grows. Why? Because miners are pivoting to AI inference, but the marginal revenue per GPU is dropping. The 2022 Terra-Luna collapse taught me to watch stablecoin reserves on Curve; today, I watch the spread between GPU rental prices (vast.ai, AWS spot) and Nvidia’s forward earnings. That spread is narrowing. Fragmented yields, fragmented trust.
Contrarian Angle: Correlation ≠ Causation
Wang’s thesis suffers from a classic logical flaw: assuming that because Nvidia’s margins are high, they must be unsustainable. In reality, the CUDA ecosystem (4M+ developers), NVLink, and supply chain moats are not easily disrupted by a new math theory. The real threat is customer self-chip development (Microsoft Maia, Google TPU, Amazon Trainium). But that’s a gradual erosion, not a “million-fold” collapse.
Moreover, if compute demand truly drops by a million-fold, storage is not immune. Most AI data is generated by compute; less compute means less data, which means less storage demand. The “storage as safe haven” narrative ignores the fact that HBM (high-bandwidth memory) is a key component of AI servers—if AI server demand tanks, HBM prices crash. Wang’s recommendation of Montage and CXMT (Chinese storage players) may have more to do with geopolitical hedging than pure market logic.
On-chain truth > Twitter narrative. The data shows that the “bubble” is real but its bursting is likely a slow bleed—margin compression, not a sudden collapse. The real contrarian play is not to short Nvidia and buy storage, but to monitor the divergence between on-chain compute token volume and Nvidia’s price. That divergence is the early warning signal.
Takeaway: The Next-Week Signal
Next week, watch the OTC desk volumes for Nvidia (Coinbase Prime) and the GPU rental price trends on decentralized compute networks. If the divergence between token volume and stock price continues, the market is acknowledging the risk before the news. The million-fold math revolution is a fairy tale; the real revolution is a slow, data-driven reckoning. The institutions are already hedging. Are you?