Just days before leaving office, the Biden administration once again decided to impose further restrictions on the export of AI chips such as Nvidia AMD.

It is also his last-ditch effort to prevent American technology from falling into Chinese hands.

According to people familiar with the matter, the United States wants to limit the sale of AI chips at both the national and corporate levels. In other words, the ban on semiconductor trade will be extended to the entire world.

In this way, the United States can not only concentrate the development of AI in countries and regions with which it is friendly, but also promote global enterprises to align with American standards, and thus directly control the spread of AI technology.

It seems that this time, the United States is determined to kill.

It is reported that the new rules may be published as early as Friday, and will set up three tiers (Tier 1, Tier 2, Tier 3) chip restrictions.

• Tier 1, a small number of US Allies, with unrestricted access to US chips
• Tier 2 countries and territories will face country-by-country capacity limits
• Tier 3 countries and regions, data centers will be completely prohibited from importing chips

Limiting global AI chips

Companies based in the last group would be able to go beyond the limits of their home countries and receive higher limits if they agreed to comply with the U.S. government's security requirements and human rights standards.

If approved, these companies will be granted "Verified End User" (VEU) status.

In this way, the United States can establish a global system of "US-certified" entities and ensure that these entities develop and deploy AI in a controlled environment.

Hierarchical control measures

Details of the new rules are as follows, according to people familiar with the matter.

One-level control

Tier 1 includes 18 countries and regions, including the United States, Germany, the Netherlands, Japan and South Korea.

Companies are free to deploy computing power in these regions, and companies based there can apply for a blanket license from the U.S. government to supply chips to data centers in much of the rest of the world.

However, the following conditions must be met: the total computing power outside Tier 1 countries and regions must not exceed 25%, and the computing power in any Tier 2 countries and regions must not exceed 7%. At the same time, companies must strictly comply with U.S. government security requirements.

Companies based in the United States that want to apply for VEU certification must ensure that at least 50 percent of their computing power remains in the United States, according to people familiar with the matter.

The overall goal of these regulatory measures is to ensure that the United States and its Allies always have a strong computing advantage over the rest of the world.

Secondary control

For most countries and regions in the world, it is in the Tier 2 level.

In response, the US government has also very kindly set a computing power cap for them - between 2025 and 2027, each country and region will only get about 50,000 Gpus.

However, if a company applies for VEU status for each country in which it is building a data center, it can receive a higher cap, which will gradually increase over time.

To qualify for VEU, companies need to demonstrate that they have a track record of meeting U.S. government security and human rights standards, or at least provide a credible plan to meet those requirements.

When a company receives a VEU certificate from a country, its chip imports will not count towards the country's maximum total.

In other words, incentivize these AI companies to cooperate with the U.S. government and adopt U.S. AI standards.

Three-level control

Tier 3, which will face the strictest level of restrictions, applies to all countries subject to the US arms embargo, totaling about two dozen countries and territories.

The export of chips to data centers in these regions will be completely prohibited.

Nvidia and the semiconductor industry Association strongly oppose it

In a statement, Nvidia expressed its opposition to the proposal

A last-minute rule restricting exports to much of the world would be a major policy shift that, far from reducing the risk of abuse, would threaten economic growth and American leadership.

The widespread global interest in accelerated computing in everyday applications is a major opportunity for the United States to boost the economy and create more jobs.

Like Nvidia, the semiconductor industry Association also expressed opposition to the move

Policy changes of this scale and importance should not be rushed through in the midst of a presidential transition and without adequate consultation with industry.

The risk of avoiding a prudent decision-making process is simply too great. Only by ensuring the right policies in this area can our country maintain an edge in global competition.

Model weight control

In addition to semiconductor controls, the new rules also restrict the export of closed-source models.

Specifically, enterprises cannot deploy high-performance closed-source models in Tier 3 countries and regions. When deployed in Tier 2 countries and regions, security standards must be strictly adhered to. (For companies with VEU qualifications, they are not subject to the model weights.)

In contrast, the open source model (or open weight model) is not subject to these rules; Closed-source models with lower performance than existing open source models are also not regulated.

In this case, if an AI company needs to fine-tune the common open source model and the process requires a lot of computing power, it must apply for permission from the US government to operate in Tier 2 countries and regions.