Hook
On a quiet August afternoon, a single report from NTT Data’s chief researcher, Professor Wang Jiange, rippled through trading desks from Boston to Shanghai. His thesis was stark: Nvidia’s monopoly on AI compute is a bubble, and it will burst within three years. The trigger? A new mathematical theory that could slash compute demand by a factor of a million. As a token fund manager who has spent years auditing the gap between code and narrative, I know a familiar pattern when I see one. This is not a technical prediction—it is a cultural signal, a warning that the market’s belief in infinite scaling has reached its saturation point. Tracing the static in the protocol’s genesis block, I see echoes of the 2017 ICO frenzy and the 2021 NFT mania. The same psychological forces are at play, and the same question lingers: when does the narrative break?
Context
Professor Wang’s argument rests on a simple but seductive analogy. Newtonian mechanics described the fall of an apple with three parameters. Today’s large language models require billions of images and trillions of parameters. If only we had a better mathematical tool, the reasoning goes, we could compress that complexity by orders of magnitude. The conclusion is that Nvidia’s 75%+ gross margins and 90%+ market share rest on a fragile assumption—that the current scaling law will persist. Wang’s employer, NTT Data, is a traditional IT integrator, not a GPU hyperscaler. Its business model benefits from a narrative shift away from compute and toward storage and infrastructure. The report is not just analysis; it is a power play in the ongoing war for the AI value chain.
To understand the stakes, we must look at the numbers. As of mid-2025, Nvidia’s market cap hovers around $5 trillion, with a P/E ratio north of 50. The company’s revenue is still growing at triple digits, but the market is already pricing in years of future growth. Meanwhile, the physical world is pushing back. Data center power demand is projected to exceed 1,000 terawatt-hours by 2026, straining grids in Virginia, Silicon Valley, and beyond. Transformer delivery lead times are 2–3 years. The infrastructure is hitting a wall, and the market is betting that the wall will be broken by more chips, not by a paradigm shift. The question is whether the wall, or the math, will break first.
Core
Let me separate the wheat from the chaff using the tools I learned auditing smart contracts in 2017. Wang’s core claim—that a new mathematical theory could reduce compute demand by a factor of a million—is a category error. He conflates the complexity of describing a physical phenomenon with the complexity of learning universal representations. The apple’s fall is a closed system; a language model must handle open-ended, unconstrained contexts. The history of AI shows that scaling laws have held with remarkable consistency across five orders of magnitude. Even the shift toward “small models with inference-time compute” (e.g., DeepSeek R1, OpenAI o-series) has not reduced total compute demand; it has merely shifted the workload from training to inference.
Based on my experience analyzing the Ethereum infrastructure in 2017, I learned that security is a silent promise kept between nodes. Similarly, the efficiency of a neural network is not a promise that can be kept by a mathematical trick. The state-space models (SSM, Mamba), linear attention, and hypergraph networks are real, and they are improving the efficiency of Transformers by factors of 2–10x, not 1,000,000x. The claim of a million-fold reduction is not supported by any published, reproducible result. It is a narrative device, not a technical forecast.
But Wang is not entirely wrong about the bubble. The profits of Nvidia are built on a fragile equilibrium: a handful of hyperscalers (Microsoft, Google, Meta, Amazon, Oracle) account for over 50% of revenue, and each is building its own custom chip. Maia, TPU, Trainium, MTIA—the seeds of Nvidia’s erosion are already planted. The real question is not whether the bubble will burst, but whether it will deflate gradually or collapse catastrophically. The token market has taught me that yields do not vanish; they merely change form. The same is true for compute margins. The margins will migrate from GPU hardware to the software stack, to the data center, and eventually to the application layer. The image is not the asset; the belief is. The belief in Nvidia’s eternal dominance is the asset, and it is already being traded on thin ice.
Contrarian
Here is the contrarian view that the market is missing. Wang’s prediction of a “clean recession” in AI compute—where bad projects die and good ones survive—is a comforting fantasy. If compute demand were to collapse by a factor of a million, the impact would not be limited to Nvidia. It would cascade through the entire semiconductor supply chain: TSMC’s CoWoS packaging, HBM memory manufacturers (SK Hynix, Samsung), power equipment vendors (Vertiv, Eaton), and even the storage companies Wang recommends. Storage, after all, is a cyclical industry. In 2023, DRAM prices fell 50% during a glut. If AI training data growth slows, storage demand will follow. The claim that “storage is immune to the compute cycle” is a dangerous oversimplification.
Moreover, the narrative of a “savior math theory” distracts from the real risk: the alignment problem. Even if a new mathematical framework emerged tomorrow, it would not solve the challenge of aligning AI goals with human values. That is a philosophical and ethical question, not a mathematical one. The AI safety community is already underfunded; if the hype cycle collapses, the first budgets to be cut will be safety research. The bubble burst could leave us with a less safe, not more safe, AI ecosystem.
From a token investment perspective, the most interesting contrarian play is not storage or compute. It is the layer-2 infrastructure for decentralized AI—networks that allow small models to run on edge devices, secured by token incentives. Projects like Akash Network, Render Network, and Bittensor are building the “decentralized sequencing” of AI inference. But as I have argued before, the current L2 sequencers are basically single centralized nodes. “Decentralized sequencing” has been a PowerPoint for two years. The same skepticism applies to decentralized AI. The math is not there yet, but the narrative is already being priced in.
Takeaway
Yields do not vanish; they merely change form. The Nvidia bubble, if it bursts, will not be the end of AI. It will be the beginning of a new phase where the value migrates from hardware to data, to algorithms, and to the humans who govern them. For the token market, the lesson is clear: stop chasing the narrative of infinite compute, and start asking where the real scarcity lies. Is it in storage? In data? In trust? The next crypto bull run will be built on the answer to that question. Stability is the quiet architecture of trust. And trust, unlike compute, cannot be scaled by a factor of a million.
