According to the external media review, much of the criticism surrounding the AI Chip Regulatory Programme “Plan A” was based on an erroneous reading of the content of the programme. According to the article, this set of governance proposals, which is aimed at US-China coordination, focuses on registration, review and calculator tracking rather than comprehensive monitoring for ordinary users.
Programs for high-end algorithms
By article combing, Plan A requires high-performance AI chip factories, buyers and data centres where the chips are deployed to be registered with the Government and inspected. A permit is also required if the enterprise resold the chip.
The programme also requires relevant data centres to disclose basic information on training assignments, including the scale of training, and to prove that they operate with declared codes. According to the article, the chip would be equipped with encryption control software and could be stopped remotely by the United States or China if necessary.
The unit cost of the H100 chip mentioned in the paper is approximately US$ 40,000. According to the author, this type of regulation is mainly directed to the high-end computing power required by Frontline AI, excluding consumer electronic devices such as mobile phones, laptops, etc.
Cost impacts are limited
The article compares this set of regulations with those of controlled medicines in the United States, arguing that high-intensity registration, inspection and flow-tracking do not necessarily evolve into a system of day-to-day public awareness surveillance.
The authors estimate that Plan A may have pushed the regulatory intensity of the AI chip industry to higher levels in the United States, but the direct economic impact is mainly reflected in a small increase in the price of the chip, which is about a few percentage points. At the same time, the article notes that AI ' s reasoning costs themselves continue to decline rapidly.
Restrictions on open weights in 2030
According to the article, the most critical part of the package is not to inspect itself, but to allow both sides of the United States and China to see the flow of chips and to have a technical constraint on the relevant calculus, thereby reducing the situation where one side is secretly breaking through and the other is forced to catch up.
The article acknowledges that the most controversial part of Plan A is the prohibition on training new open weight models after 2030. This is considered to be a real restriction on open ecology, but in the author's view it is not the same as “supervising common user equipment”.
Instead, Plan A supports the “open algorithm” approach, whereby an enterprise does not have to disclose the weight of the final model, but does need to publicly train the research behind it. The article argues that, as long as other companies have sufficient capacity, there is an opportunity to replicate similar modelling capabilities.
The Trump government has taken similar measures.
The article also mentions that in January this year, the Trump government introduced a number of controls on the AI chip, which is close to Plan A, but did not trigger public debate on the same scale.
As listed, these measures include requiring chip manufacturers to perform bank-like KYC verification of customers, requiring customers to confirm their downstream customer compliance, requiring physical security certification, and placing chips at third-party laboratories for performance verification.