Hook
A 40-page internal memo from NTT Data’s chief researcher, Wang Jiange, has leaked into the public domain. The core thesis: Nvidia’s 75%+ gross margins and 90% market share are built on a mathematical illusion. He predicts a 1,000,000x reduction in compute demand within three years due to a missing mathematical framework for black-box AI models. For the crypto mining sector, this is not mere noise. If GPU supply suddenly floods the secondary market, the cost base for every proof-of-work network—from Bitcoin to Ethereum Classic—shifts overnight. Data doesn’t lie: the current GPU rental market already shows a 30% drop in spot prices for H100s since Q1 2025. The question is whether this is a cyclical correction or the beginning of a structural collapse.
Context
Wang Jiange is a chief researcher at NTT Data, one of Japan’s largest IT services firms. His analysis, originally published via Phoenix Finance on August 18 (likely 2024), argues that current AI scaling laws are unsustainable because they lack a “physics-level” mathematical description of intelligence. He compares training a large language model to describing an apple falling with billions of images—a category error in his view. The article positions Nvidia as the key victim, while betting on memory chip companies (e.g., Montage Technology, CXMT) as long-term beneficiaries. For crypto miners, this is relevant because Nvidia GPUs are the backbone of most GPU-mineable coins (ETH is gone, but ETC, RVN, FLUX, and others remain). A collapse in AI demand would directly impact GPU availability, pricing, and mining profitability. Additionally, NTT Data’s stance reflects a traditional IT giant’s frustration with the AI hardware monopoly—similar to how ASIC manufacturers once disrupted GPU mining.
Core: Facts and Immediate Impact
Let’s break down Wang’s claims with on-chain and market data. First, the 1,000,000x reduction thesis: based on my audit experience during the Ethereum Classic 51% attack aftermath, I know that theoretical breakthroughs in cryptography once reduced specific attack complexities by orders of magnitude—but those were within well-defined mathematical boundaries. Applying such a reduction to general intelligence is a leap. Over the past 5 years, scaling laws have held empirically: doubling model parameters consistently improves benchmark scores by a predictable margin. Even if Wang were right, the transition would take years, not three. Meanwhile, the crypto mining market is already feeling the heat. The 2024-2025 GPU supply glut, driven by AI demand normalization, has pushed used H100 prices from $30,000 to under $20,000. This is a 33% decline—hardly 1,000,000x, but significant for miners who bought at peak. If AI demand truly collapses, GPU prices could fall to cost-plus levels, making mining profitable for those with low electricity costs. However, the flip side is that many miners also run AI workloads (e.g., renting out GPUs for inference). A dual-use market means a downturn in one sector can ripple into the other. Verify the hash, ignore the hype: the real risk is not a sudden “mathematical revolution,” but a gradual erosion of Nvidia’s pricing power as hyperscalers (Microsoft, Google, Amazon) ramp up custom chips. This is already happening—Microsoft’s Maia chip, Google’s TPU v6, and Amazon’s Trainium 2 are eating into Nvidia’s share. For miners, the key metric to watch is the GPU days-to-ROI, which has stretched from 12 months to 18 months in 2025. On-chain metrics > Twitter polls: the number of active GPU miners on Ethereum Classic has dropped 15% year-over-year, while the network hashrate has remained stable due to newer ASICs. This suggests that GPU miners are being squeezed out, not by AI, but by ASIC competition. Wang’s storage thesis, however, is intriguing. He claims that memory chips (DRAM, NAND) will benefit regardless of AI architecture shifts. In crypto, this maps directly to Filecoin, Arweave, and Chia—networks that require massive storage capacity. If AI data generation continues to grow, the demand for decentralized storage could rise. But note: Filecoin’s storage utilization rate is only 15% as of May 2025. The correlation between AI data storage and crypto storage is not one-to-one. Data doesn’t lie: the average Filecoin deal size has grown 22% in the last quarter, but the price of FIL has not followed. This divergence suggests that the market is pricing in a different narrative—perhaps the risk that AI bubble burst could reduce data generation, hitting storage demand instead.
Contrarian Angle: The Blind Spots
Most analysts covering Wang’s memo focus on the “Nvidia is overvalued” angle. The contrarian take is that the crypto mining sector might actually suffer from a GPU glut, not benefit. Here’s why: a flood of cheap GPUs sounds great for miners, but it also reduces the barrier to entry, leading to increased network hashrate and difficulty. This is exactly what happened after the 2018 crypto bear market—GPU prices plummeted, but hashrate surged, compressing margins. Additionally, if AI bubble bursts, the broader tech sell-off could drag down crypto prices, making mining revenue fiat-denominated less attractive. The storage thesis has a similar blind spot: memory chips are a cyclical industry. In 2023, DRAM prices fell 40% due to oversupply, and SK Hynix’s revenue dropped 30%. If AI demand slows, memory demand will follow, albeit with a lag. Wang’s claim that storage companies are “unaffected” is a fallacy. My analysis of DeFi Summer liquidity pool stress tests taught me that correlated risk is often underestimated. In 2020, when Uniswap liquidity pools were stressed, the ripple effects hit stablecoins and lending protocols. Similarly, a tech sector downturn would hit all semiconductor stocks, including memory. The actual contrarian opportunity might be in GPU-as-a-service platforms (e.g., Render Network, Akash) that can dynamically shift between AI and crypto workloads. If AI demand falls, these networks can pivot to rendering or scientific computing, maintaining utilization. Wang’s memo ignores this flexibility.
Takeaway: What to Watch Next
The next 12 months are critical. Monitor the following metrics: (1) Nvidia’s data center revenue growth rate—if it dips below 50% year-over-year, the narrative shifts. (2) GPU rental spot prices on platforms like Vast.ai and Lambda Labs—a sustained decline below $1.50 per hour for H100 signals oversupply. (3) Filecoin’s storage utilization—if it breaks 30%, the storage thesis gains credibility. For miners, the best hedge is not to sell all GPUs, but to lock in low electricity rates and diversify into dual-use assets. The question is not whether the bubble will burst, but whether the market has already priced in a gradual cool-down. Verify the hash, ignore the hype.