Summary of Key Points
The recently released Kimi K3 model by a Chinese company has caused a collective decline in AI-related stocks worldwide (with NVIDIA and the semiconductor sector losing trillions in market value). On the surface, it seems the market is worried about the power of Kimi, but the real issue is that the narrative on Wall Street over the past three years—“AI equals spending money on GPUs”—has been shattered. Kimi achieved a top-tier model at only 60% of the cost, prompting capital to question the logic that more computing power (the more GPUs, the better) is always necessary. However, the article suggests that Wall Street might be wrong; reduced model costs could lead to an explosion in demand, as the need for computing power shifts from centralized training to ubiquitous inference. This decline is more of a trigger for market adjustment than a collapse in AI logic, indicating that AI competition has entered a new phase focused on efficiency.
Detailed Analysis
1. The Truth Behind Wall Street's Panic: It's Not That Kimi Is Too Strong, but the “Computing Power Myth” Is Crumbling
For three years, Wall Street has believed in a simple formula: the more powerful the model, the larger the number of parameters, the more expensive the training, and the more GPUs required—thus making NVIDIA more valuable. This logic drove NVIDIA’s market value from $1 trillion to $5 trillion, with giants advocating for increased investment in AI infrastructure, as if more spending would create a stronger competitive advantage.
However, Kimi K3 has broken this cycle by outperforming top models from the U.S. at just 60% of the cost, ranking first on multiple charts. Goldman Sachs even questioned, “How can Chinese laboratories, which can’t obtain large quantities of GPUs, narrow the gap?” This question hits the core issue: computing power is no longer the only factor; algorithms, synthetic data, and post-training techniques are also crucial. Wall Street suddenly realizes that if more companies can create good models with fewer GPUs, GPU sales might decline, and TSMC’s return on capital expenditure could be affected. As a result, stocks have plummeted as a precaution.
2. Chinese Companies’ “Low-Cost Counterattack”: A Strategy Forced by Reality
Chinese companies have found success on the low-cost path, not by luck but due to their circumstances. American giants, with ample funds and resources, rely on more computing power to overcome engineering flaws. In contrast, Chinese leaders faced shortages of GPUs and funding, so they had to optimize what they had:
- Synthetic Data: Using a small number of high-quality samples to train models with strong capabilities.
- Post-Training + Reinforcement Learning: Fine-tuning to compensate for shortcomings in pre-training.
These successes (with DeepSeek and Kimi) prove that the low-cost approach is viable. Wall Street’s fear is not just about one Chinese company; if this strategy is proven effective twice, it could become standard, turning the advantage of American giants into a liability.
3. Wall Street May Be Misled: Computing Power Demand Won’t Decline, but the Approach Is Changing
Just as Wall Street proclaimed the end of the “computing power expansion era,” Kimi’s official announcement revealed the opposite. The model’s high demand overwhelmed their systems, leading to the suspension of new user subscriptions and the sale of energy-intensive features separately.
This is a manifestation of the Jevons Paradox: improving efficiency actually increases demand. For example, cheaper computers led to higher overall sales, and cheaper cloud computing resulted in more servers worldwide. Lower model costs will spur more AI applications (such as 24/7 AI assistants and mobile local models), shifting the need for computing power from centralized training to continuous inference. The total amount of computing power may not decrease but will likely increase, with GPUs now being used for both training and inference. NVIDIA can no longer rely solely on its traditional revenue sources, but it won’t collapse either.
4. Kimi Is Just a Trigger: The Market Needed This Adjustment
The decline was not due to Kimi’s impact on AI itself but because the market was already overstretched:
- AI stocks have risen sharply this year (with the Philadelphia Semiconductor Index up 68%), so a correction was inevitable.
- Smart investors have been selling assets, with U.S. companies internally selling $77.6 billion in stocks in the first half of the year (the second-highest amount in 20 years).
- Other concerns include TSMC’s higher-than-expected capital expenditure and delays in Google’s Gemini model.
Kimi provided an excuse for these sales; in reality, it’s a shift in market behavior. Funds previously invested in NVIDIA are being reallocated to other sectors.
5. AI Competition Enters a New Phase: From “Resource Competition” to “Efficiency Competition”
The competitive landscape in the AI industry is changing:
- First Phase (2023–2025): A race for computing power, with NVIDIA and TSMC leading due to their GPU capabilities.
- Second Phase (starting 2025): A focus on efficiency—who can use fewer resources to create better models, launch products faster, and control inference costs will gain the upper hand.
Wall Street’s preference for simple narratives (e.g., “GPU dominance”) is outdated. The new era emphasizes efficiency: who can use resources more efficiently to create value.
Kimi’s significance lies in highlighting that AI competition now focuses on converting the same resources into greater capabilities. In the past, money determined success; in the future, efficiency will be the key.
Conclusion
This market decline marks not the end of AI but the beginning of a shift from wasteful spending to more efficient use of resources. The focus will shift from who has the most GPUs to who can use them most effectively.