Nvidia Bubble Bursting: How China’s DeepSeek Made It?
Since Chinese AI company DeepSeek released its next-generation large language model, tremors have rippled through U.S. tech stocks. In just 17 days, NVIDIA’s stock price plummeted by 23%, erasing over $830 billion in market value. A swift backlash erupted across American industries, with DeepSeek even facing cyberattacks peaking at 120 million requests per second.
While DeepSeek has become a scapegoat for those burned by U.S. market losses, it is merely the trigger. The real cause of the turmoil lies in two dangerous delusions that have long propped up America’s AI strategy—delusions that U.S. media, to sustain the illusion of tech sector prosperity, have either ignored or actively reinforced.
Delusion 1: Computing Power Is Everything
Western media has relentlessly pushed the narrative that AI performance scales linearly with compute investment. NVIDIA CEO Jensen Huang famously declared, the more computing power you have, the more power you have. This mantra fueled Wall Street’s obsession with the compute supply chain, from GPU hardware to cloud services, with investors convinced that monopolizing advanced chips would guarantee dominance over global AI development. Yet this logic is fundamentally flawed: Computing power is merely the “fuel” for AI, not the “engine” that determines success.
Ironically, NVIDIA itself has long recognized the diminishing returns of raw computing. Its investments in CUDA software ecosystems, dynamic parallelism, and model compression algorithms all aim to achieve “higher performance with less compute.” For instance, mixed-precision training cut AI training costs by 40% for certain tasks, while the NeMo framework boosted inference speeds by 30% through parameter sharing. But these breakthroughs never became central to Wall Street’s narrative—after all, acknowledging the fragility of the “computing power myth” would destabilize NVIDIA’s $2.8 trillion valuation.
The delusion shattered when DeepSeek revealed its model was trained at 30% of OpenAI’s cost, primarily using Huawei’s Ascend 910B GPUs. Though the 910B delivers 80% of the A100’s raw compute, DeepSeek’s MoE architecture and dynamic sparse training techniques doubled its per-unit efficiency. This exposed a harsh truth: algorithmic innovation can bridge hardware gaps, rendering NVIDIA’s compute hegemony replaceable. Had U.S. investors understood this earlier, DeepSeek might have been seen as a competitor rather than a threat. But Wall Street’s years of “compute worship” left them unprepared—a single counterexample collapsed the entire narrative.
Delusion 2: Performance Benchmarks Mask an Application Desert
America’s AI sector faces another paradox: soaring performance metrics have failed to translate into real-world productivity gains.
DeepSeek’s emergence was no accident. Its parent company specializes in quantitative trading, forcing its AI models to survive China’s “hellish” stock market—a daily battleground with 210 million retail investors. To thrive, models must process vast unstructured data (policy documents, social media sentiment) and execute decisions within 0.01 seconds. This pressure cooker environment forged Chinese AI’s pragmatic edge: DeepSeek’s financial inference accuracy hits 87%, dwarfing GPT-4’s 72%.
The divergence stems from technical priorities. OpenAI’s closed ecosystem and costly APIs (6x pricier than DeepSeek’s) exclude SMEs and researchers. In contrast, DeepSeek’s open-source framework and model distillation enables local deployment of 1-billion-parameter models for under $300/month, with customization for niche applications. While U.S. academics beg for GPT-4 access, over 200 Chinese universities have integrated DeepSeek into research platforms.
Worse, America’s obsession with “performance theater” has spawned a vicious cycle. To please investors, firms inflate parameters (GPT-5 may hit 10 trillion) while ignoring safety, efficiency, and ethics. When AI aces benchmarks but fails a Chinese high school math exam, it’s clear: chasing scores divorces technology from reality.
Conclusion: Ecosystem vs. Echo Chamber
DeepSeek’s rise reflects a fundamental divide in AI development. China’s path—rooted in industrial integration—has nurtured technologies like Huawei’s Ascend GPUs, refined by smart manufacturing and autonomous driving. DeepSeek itself evolved through 210 million users’ “stress tests,” a crowdsourced evolution.
Meanwhile, America’s closed-source approach, while protecting short-term profits, stifles innovation. GitHub data shows 42,000 vertical applications built on DeepSeek’s open models across 40+ industries, versus fewer than 8,000 GPT-based tools (mostly for marketing and entertainment). When technology cannot benefit society broadly, “performance leadership” becomes a paper crown.
China’s DeepSeek offers a new path for global AI development, presenting an alternative to the “computing power is everything” approach. It reflects the philosophical insights and deep exploration of AI from Eastern civilization. History shows technological revolutions are won not by singular metrics but by ecosystem vitality. If the U.S. clings to its “compute supremacy” delusion, it will sink deeper into a high-cost, low-return quagmire. As NVIDIA’s crash proves: when illusions fade, reality bites harder than anyone expects.
Editor: Charriot Zhai
