The US-China AI Rivalry: Cost Efficiency Reshapes the Global Landscape
- Dr Frederick Wong
- Aug 4
- 5 min read
"Restraint is a strategy, not a compromise." — Liang Wenfeng, DeepSeek
I. Paradigm Shift: From Computing Power Arms Race to Efficiency Competition
The global AI race is undergoing a paradigm shift. The United States relies on a capital-intensive model to stack computing power, whereas China utilizes algorithmic innovation and engineering efficiency to achieve equivalent performance with significantly fewer resources. This represents a fundamental divergence in innovation philosophy and business logic.
Dimension | China | United States | Gap |
Total AI Investment | ~$20+ billion USD | >$2.8 trillion USD | US is approx. 100x China |
Daily Token Consumption | ~140 trillion times | <50 trillion times | China is approx. 3x US |
API Cost (per million tokens) | ~$0.8–1 USD | ~$15 USD | US is approx. 15x China |
Training Cost (R1 level) | ~40 million RMB | Hundreds of millions USD | US is approx. 20x China |
Despite investing a hundred times more capital, the market scale and cost-effectiveness generated by the US fall far short of China's. The foundation of "computing power determinism" is being shaken.
II. China's Dual Engines: DeepSeek's Restraint and Kimi's Breakthrough
DeepSeek: Defining boundaries by "what not to do"—abandoning trending tracks like video generation and consumer-end (C-end) traffic, with AGI as its sole main objective. The R1 model's training cost was only about 40 million RMB, developed by a team of fewer than 140 people, achieving equivalent performance with 1/20th of the computing power used in the US. By open-sourcing its most powerful model, it has established ecological standards and influence. It operates without a KPI-driven culture, uniting the team through a shared vision.
Kimi (K3): Featuring approximately 2.8 trillion parameters, a 1-million token context window, and an MoE (Mixture of Experts) architecture, it has improved long-context efficiency by 2.5 times. It topped the global Frontend Code Arena, marking the first time a Chinese large model has taken the crown on an authoritative coding leaderboard. Its API pricing has entered the top-tier range, with B2B (business-to-business) revenue accounting for over 70%, and its Annual Recurring Revenue (ARR) surging from $100 million to $300 million USD in just three months.
Implications of the Dual-Track Approach: DeepSeek disrupts competitors' profit margins with low costs, while Kimi proves the brand value of Chinese models through high premiums. Together, they dismantle the stereotype that "China can only compete on cost-effectiveness," forming a pincer movement against US AI giants.
III. Structural Risks of the US AI Bubble
The capital logic of US AI assumes that "increasing computing power investment will sustainably yield a leading edge." The practices of Chinese enterprises are currently falsifying this assumption:
Profit Margin Squeeze: China's API costs are only 1/15th of those in the US, forcing tech giants to lower prices and eroding their profits.
Declining Return on Capital: The $2.8 trillion USD investment has not yielded a corresponding scale of returns, exhausting investor patience.
Valuation Reassessment: The high valuations of tech giants are built on the premise of "infinite growth in AI demand." The shattering of these expectations will impact the US dollar capital market.
Historical Analogy: The case of China's rocket recovery technology surpassing SpaceX demonstrates that achieving technological breakthroughs and cost advantages in critical areas is enough to suppress competitors' valuations and financing capabilities. If China continues to achieve over 90% of the performance at 1/10th to 1/20th of the cost, US companies will be forced to continuously inject additional investments to maintain their lead, ultimately straining their cash flows. When investors realize that "stacking chips" is not the only answer, capital will reassess the true value of the AI sector.
IV. Domestic Computing Power: De-CUDA-ization and Ecosystem Autonomy
DeepSeek's self-developed compiler, TileLang, has enabled V3 training to become almost entirely independent of CUDA. This enhances resilience against blockades under chip export controls and opens up opportunities for domestic chips like Huawei's Ascend. Liang Wenfeng assesses that the performance of the Huawei 950 super node can substitute the GB200 at a ratio of "four top-tier cards to one," though the ecosystem and production capacity still require about 3 years to mature. The key lies in whether the improvement in algorithmic efficiency can continue to outpace the progress of hardware substitution.
V. AGI Roadmap: The Endgame Battle
Stage | Capability | Status |
Stage 1 | Chain of Thought (CoT) | Breakthrough achieved |
Stage 2 | Agent | In progress |
Stage 3 | Continuous Learning | Current biggest hurdle |
Stage 4 | Self-iteration | Approaching singularity |
Stage 5 | Embodied AI | Ultimate goal |
Continuous Learning — enabling models to self-evolve without needing to be trained from scratch—is the greatest technical bottleneck at present and the core focus of DeepSeek's research. Whoever breaks through this first will hold the key to AGI.
VI. Industry Consolidation and Investment Implications
Liang Wenfeng predicts that domestic foundational model companies will inevitably undergo consolidation, with the endgame being "two or three large enterprises + two or three small enterprises." The ultimate industry competition will be based on cost control, iteration speed, and technological moats. The investment dimension must be reassessed: algorithmic efficiency may trigger a valuation reassessment of the computing power supply chain; Chinese AI has already demonstrated brand premium capabilities; the maturity of the Ascend ecosystem is a critical indicator of resilience against blockades; and open-source strategies are dismantling closed-source monopolies.
VII. Conclusion: The Dual Code of Restraint and Breakthrough
The rise of Chinese AI represents the triumph of an innovation philosophy. DeepSeek defines boundaries through "restraint," while Kimi proves value through "breakthroughs." Computing power is not omnipotent, and capital is not omnipotent—what determines victory is the insight into "what is truly important." Will the US AI bubble burst? The answer perhaps isn't a simple "yes" or "no," but rather that China's low-cost, high-efficiency route is firmly altering how global investors calculate the value of AI. When these changes accumulate to a critical point, a bubble correction is merely a matter of time. AGI is the endgame, and everything happening now is just the prelude. Only the players who can master the dual code of "restraint" and "breakthrough" during this prelude will earn a seat at the main course table.
⚠️ Risk Warning: The data in this article is derived from media-compiled closed-door meeting minutes and public reports. It is not officially released, difficult to independently verify, and should only be used as a reference for analytical frameworks. The "bubble bursting" is a deductive viewpoint and should not be used as a sole basis for investment.
References & Highlights
[1] Liang Wenfeng Closed-Door Meeting Minutes (Media Compiled Version) — Strategic restraint, AGI roadmap, open-source strategy, organizational culture.
[2] Li Junhuai's WeChat Video Account — US-China AI investment comparison, bubble deduction.
[3] Moonshot AI Kimi K3 Release Information — 2.8 trillion parameters, MoE architecture, topping global programming leaderboards, interactions with Elon Musk.
[4] Third-Party Authoritative Evaluation Platforms (Artificial Analysis / Frontend Code Arena).
[5] Huawei Ascend vs. NVIDIA CUDA Technical Data.
[Limited-Time Expert Consultation Invitation]
FOFA sincerely invites visionary entrepreneurs and investors to engage in deep, practical exchanges regarding the aforementioned trends, we will provide you with a complimentary expert planning consultation to help you tailor a specific entrepreneurial path or investment blueprint, allowing technology leverage to serve your asset appreciation.
Book your strategic dialogue NOW:
Comments