NTT Data’s Warning: The AI Compute Bubble Will Burst in Three Years

The data doesn’t lie. But the narrative does. This morning, a report from NTT Data’s chief researcher, Professor Wang Jiange, hit the wires like a depth charge in the AI ocean. His thesis is stark: Nvidia’s monopoly on AI compute is a bubble, and it will burst within three years. The reason? Our current black-box large language models lack a mathematically efficient framework—they are consuming compute orders of magnitude beyond what is physically necessary. His conclusion: storage chips (e.g., Montage, ChangXin) will be the long-term winners, while GPU makers face a reckoning. For context, this is not a random blogger. NTT Data is Japan’s largest IT services firm, a system integrator with deep roots in cloud, storage, and enterprise infrastructure. Professor Wang’s position gives him a birds-eye view of the AI supply chain. His timing is deliberate: published on August 18 (likely 2024), during the peak of Nvidia’s market cap surge past $3 trillion. The article is a strategic narrative weapon, not a research paper. But the question remains: Is his technical logic sound, or is this a classic case of a legacy player trying to steer the narrative toward its own strengths? Let’s cut through the noise. Wang’s core claim is that we lack a ‘new mathematical description tool’ for intelligence, analogous to how Newton’s laws reduced the complexity of planetary motion. He argues that once this tool emerges, compute demand could drop by a factor of millions. This is a category error. Describing an apple falling requires three parameters, but understanding language, vision, and reasoning in any unseen context is a fundamentally different problem. The physics analogy is seductive but wrong. Scaling laws have held for five years across OpenAI, Google, and Anthropic: more compute, data, and parameters consistently yield better capabilities. Even with the recent shift to inference-time compute (DeepSeek R1, OpenAI o-series), total compute demand is still rising, not falling. My own experience with ICO due diligence in 2017 taught me a hard lesson: code is law, until it isn’t. I spent six weeks auditing a top-10 ICO’s smart contracts, only to have my report ignored because hype trumped security. The market priced in sentiment, not utility. The same dynamic is at play here. Wang’s ‘million-fold reduction’ has no basis in any reproducible experiment. The most promising research—state-space models, linear attention, hypergraph networks—is still optimizing within the machine learning paradigm, not inventing a new language for intelligence. No one has demonstrated a path to a three-year, million-fold efficient algorithm. Volume lies. Liquidity speaks. The real question is not whether AI compute is overvalued, but what happens when the supply-demand imbalance shifts. Nvidia’s gross margins at 75%+ are unsustainable, but the erosion will come from self-designed chips (Microsoft Maia, Google TPU, Amazon Trainium), not from a mathematical revolution. These custom chips are already eating into Nvidia’s market share, and as CoWoS packaging capacity expands in 2025, GPU supply will loosen. The bubble will deflate, not burst. The timeline is three to five years, not three years. Wang’s blind spot is the network effect. Nvidia’s moat is not just the GPU; it’s CUDA, NVLink, InfiniBand, and the developer ecosystem. Replacing that requires a new compute stack, not just a new algorithm. I saw this firsthand during DeFi Summer in 2020, when I ignored the herd chasing 1000% APYs and stuck to my risk model. The bZx hack nearly wiped out my competitors, but I survived. Stability is a narrative too. The same applies to Nvidia: its ecosystem is sticky, and even if a new math tool emerges, it will take years to displace the installed base. From a regulatory perspective, Wang’s call aligns with the growing push for AI compute to be treated as a national resource. The US CHIPS Act and export controls assume Nvidia’s chips are strategically scarce. If compute demand collapses, that narrative weakens. But the more likely scenario is that compute demand continues to grow, but at a slower rate, and the marginal beneficiaries are not just storage but also alternative compute providers—especially in China, where Huawei Ascend is gaining traction. The analysis of Wang’s article by industry experts (which I have parsed) notes that he names ChangXin and Montage, both Chinese companies, suggesting a geopolitical undercurrent. Japan’s IT giants want to reduce dependence on US compute, and storage is a domain where they can compete. Contrarian angle: The real risk is not a mathematical breakthrough but a demand shock from AI application failures. If the current wave of AI agents and chatbots fails to generate sustainable revenue, enterprises will cut their compute budgets. The 2026 AI-agent integration framework I developed for Render revealed that many projects have tokenomics that don’t account for agent transaction fees. If AI agents drain liquidity, the whole ecosystem cracks. That’s the bubble—not the hardware, but the unfunded liabilities in the token economy. Takeaway: Wang’s thesis is a useful narrative stress test, but it’s not an investment thesis. The next three years will see compute supply gradually catch up, forcing Nvidia to lower prices, but the paradigm shift from GPU to storage is a stretch. Storage is cyclical—HBM demand will peak with AI server builds, then fade. The smart play is to diversify across compute, storage, and applications, with a hedge on regulatory clarity. Code is law, until it isn’t. But the market’s reaction to this warning will be a signal: if Nvidia’s stock drops 10% and stays down, the narrative has shifted. If it bounces, the bubble still has room to inflate.