Written by: Akasha2049

After three years of lockdown and seven rounds of regulation, the United States has exhausted its policy ammunition. In February 2026, the volume of AI model calls in China surpassed that of the United States for the first time. This is not just a reversal of numbers—it signifies that the "chokehold" strategy has entered a new phase of structural failure.

A number that caused Nvidia to evaporate 260 billion dollars

On February 26, 2026, Nvidia announced its record-breaking financial report: Q4 single-quarter revenue reached 68.1 billion dollars, a year-on-year surge of 73%. Normally, this should have been a celebratory night. However, the strong performance instead triggered pessimism in the market—Nvidia's stock price plummeted 5.5% overnight, resulting in a market capitalization loss of nearly 260 billion dollars (about 1.77 trillion yuan).

On the same day, across the ocean in the A-share market, the sectors of computing power leasing, cloud computing, and data centers saw a wave of trading limits hit. Behind the contrasting market trends is a signal hidden in the financial report: during the same week Nvidia released its earnings, data from OpenRouter, the world's largest AI model API aggregation platform, showed that in February 2026, the weekly call volume of AI models in China (measured in Tokens) for the first time surpassed that of the United States, with four out of the top five popular models globally coming from China.

The market understood the implications of this signal: AI prosperity no longer equates to linear growth in Nvidia's high-end chip sales. The underlying logic of the computing power landscape is being quietly rewritten.

The three-year history of the blockade: a continuously escalating policy list

To understand the current situation, one must first trace the unprecedented technological blockade imposed by the United States and how it has evolved step by step.

The United States first implemented export controls on advanced computing and semiconductor manufacturing equipment to China, placing high-performance chips like the A100 and H100 on the banned export list.

Regulations intensified, incorporating performance density into standards. Nvidia's "castrated" chips A800 and H800, specially supplied for China, were subsequently banned, and even the RTX 4090 was not spared.

The "Strengthening Framework for Overseas Key Export Restrictions" was passed, extending the scope of control from hardware to the export of top-tier AI large models.

Before leaving office, the Biden administration released the "AI Diffusion Export Control Framework," designating China as a "fully embargoed nation for GPU chips," with the scope of control extending to model weights.

The Trump administration rescinded the aforementioned framework but simultaneously released three more targeted guidelines: the global use of Huawei Ascend chips is viewed as a violation of U.S. export controls; and the "chip passport" plan was promoted to track technology flows.

According to Bloomberg, the Trump administration was brewing an expansion of chip export controls from 40 countries to a global scale, attempting to make the U.S. government the "total valve" of global AI computing power.

With seven rounds of upgrades and three administrations, the U.S. chip blockade has self-iterated at an incredible speed. However, while the policy list becomes longer, its actual effectiveness has begun to show significant diminishing returns.

The paradox of the blockade: pressure creates another ecosystem

The American think tank Rand Corporation, in its "2025/2026 AI Strategic Competition Assessment" report, acknowledged a troubling fact for Washington: relying solely on chip "chokeholds" can no longer prevent Chinese models from reaching top-tier levels.

The logic behind this is not complex. The blockade creates pressure, and pressure becomes fuel for innovation.

What is being restricted is not China's imagination, but rather China's right to use a specific vendor. When this path is blocked, another one is inevitably opened up.

—— A domestic AI infrastructure investor, in private communication

Breakthroughs on the algorithm side occurred first. DeepSeek has demonstrated in a shocking manner that through a mixture of expert frameworks (MoE) and extreme engineering optimization, similar inference performance can be achieved at a computing power cost far below that of the top American models. Data from Frost & Sullivan shows that the MoE architecture can reduce inference memory usage by 60% and increase throughput by 19 times—this means that the demand for computing power has shifted from "stacking cards" to "efficiency improvement."

The acceleration of hardware alternatives followed. Data from Bernstein shows that the penetration rate of domestic AI chip brands in China has rapidly climbed from about 29% in 2024 to about 60% in 2025. The performance of Huawei's Ascend series has approached 80% of Nvidia’s H200, and has been deployed on a large scale in multiple data centers; manufacturers like Alibaba, Cambricon, Birun, and Tensent ZhiXin are also accelerating penetration through heterogeneous computing.

More critically, 2026 will become a watershed for qualitative change: domestic AI chips are breaking through from merely handling "inference" tasks to tackling more challenging "training" scenarios. The transition from "capable of inference" to "capable of training" may seem like a slight improvement in performance, but it actually signifies a deep reconstruction across the entire technology stack.

Computing power diplomacy: another card for the United States

Beyond chip blockades, the United States is also playing another card—computing power diplomacy.

The logic is to offer allies and strategic partners a "full-stack AI export package" (hardware + models + software + standards) in exchange for GPU quotas, capital repatriation, and political endorsement, thereby establishing a global AI ecosystem dependency system centered around U.S. technology. A typical case is Saudi Arabia: it has promised to purchase tens of thousands of Nvidia chips and invest 600 billion dollars to support U.S. AI projects.

This strategy is effective in the short term. It has converted AI computing power into a "strategic resource"—much like oil, it can be priced, allocated, and politicized. Through tracking mechanisms like the "chip passport," the U.S. aims to bring the bandwidth of global AI infrastructure under its own monitoring and management.

But this card also has inherent contradictions: the more stringent the control, the more it forces other countries to accelerate the establishment of independent computing power systems. The results of global developers voting with their usage have already proven this logic—Chinese models, leveraging cost advantages (with equivalent Token costs about 1/17 of foreign counterpart products), are attracting migrations from global developers, and these call volumes do not rely on Nvidia's high-end GPUs.

The competitive logic of the second half: from "ceiling" to "floor"

At this moment, AI geopolitical competition is entering a new phase, with the dimensions of competition quietly shifting.

In the past three years, competition has mainly occurred at the "ceiling"—who possesses the strongest computing power, the largest models, and the most advanced chips. On this dimension, the U.S. still maintains a lead: the o-series models launched by OpenAI, etc., still enjoy generational advantages in complex logical reasoning, while Nvidia's Blackwell architecture continues to define the upper limits of computing power.

However, the decisive point in the second half is shifting to the "floor"—who can genuinely penetrate AI into the real economy and transform it into quantifiable productivity. In this dimension, China is forming a structural advantage due to its complete industrial chain, lower inference costs, and massive diverse application scenarios.

AI is no longer just a competition of technical parameters, but a question of who can truly enhance medical efficiency, lower manufacturing costs, and optimize government decision-making.

—— Microsoft President Satya Nadella, at the 2026 Davos Forum

Morgan Stanley's forecast confirms this trend: in 2026, the demand for inference computing power will exceed that for training for the first time. This indicates that the market's demands for chips have shifted from the peak computing power of a single card to the efficiency of Token output under unit costs—precisely the advantage range currently held by China.

Three variables worth continuous attention

Of course, it is still too early to conclude that "the blockade has completely failed." Three variables will profoundly influence the course of this competition:

First, EDA tools and advanced processes. The export controls on chip design software (EDA tools) have not yet been fully enforced. If the U.S. completely enacts these restrictions, it would substantially hinder the path for domestic chips to upgrade from "inference" to "training."

Second, the interconnectivity capabilities of clusters of ten thousand chips. Domestic chips have achieved breakthroughs in single card performance, but there remains a significant gap in interconnect stability and collaborative efficiency at the level of ten-thousand-card clusters compared to Nvidia's NVLink system. This is the core bottleneck for training large-scale foundational models and the most difficult hurdle that domestic manufacturers need to overcome next.

Third, the global landscape of computing power diplomacy is diversifying. As the U.S. extends export controls globally, an increasing number of "third-party countries" will face pressure to choose sides between China and the U.S. The competition for computing power infrastructure in Southeast Asia, the Middle East, and Africa will become the next battleground of this game.

We are witnessing a historic structural shift: computing power is becoming a sovereign asset of the 21st century. It is no longer just the product of a tech company, but a component of national strategic capabilities.

The blockade has created dependence; dependence has accelerated substitution; substitution has compelled innovation. The strategic misjudgment made by the U.S. in this chain may be repeatedly referenced in history—when you try to use walls to block a nation with a profound engineering culture, you only provide it with a reason to break through its own limitations.

The next question is not "Can China catch up?" but rather "In what way will the new global AI order split, and how will it restructure?"

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。