Announcement No. 48 is signed by two ministries, MOFCOM and the Ministry of Science and Technology, both in its masthead number and again under the text. The body named on it as issuing it is MOFCOM's security and control bureau, and the date is 16 October 2026. Alongside the Foreign Trade Law, its opening sentence names a second instrument as the basis, the Regulations on the Administration of Technology Import and Export. Its scope section counts this consultation at 40 technologies across three fields, 23 in artificial intelligence, 5 in biomedicine and 12 in photovoltaics, and refers the technical scope and the policy explanation to its first attachment.
The consultation has a closing date and three routes. The public may put forward opinions between 16 October and 15 November 2026 by email at jinxianmulu@mofcom.gov.cn, by fax at 010-65197565, or by letter to that bureau at 2 East Chang'an Avenue, Beijing, postcode 100731. Opinions are invited on four subjects, the newly added scope of technologies, the indicators, the control policies and the facilitation measures, and the list is left open.
The announcement sets out the form as well. It asks that the subject line of the email, the first page of the fax and the envelope each carry the label it specifies, which it gives in Chinese, and that the submission state a contact person, their work unit and contact details. A submission made in the name of a unit shall bear that unit's seal.
Interesting article. It looks like there’s also a parallel here with foreign investment screening, particularly where a change in ownership or control triggers the requirement even though the technology itself remains in China. Different legal mechanism, of course, but potentially another consideration for cross-border acquisitions and restructurings.
Your reverse-question framing is the useful part. For years the fight was about who can export GPUs into China. Now China is asking which AI technologies it will allow to leave. Intelligence is being treated like energy or defense tech—not simply software.
I take that seriously. In my response to Hendrycks, Schmidt and Wang's Superintelligence Strategy, I argued that the hazard is not only foreign diversion of chips. It is the precedent: "silicon runs only by grace of third-party permission." Controls built for a temporary standoff "have a habit of hardening into permanent infrastructure." And compute access that hinges on continuous approval isn't leverage; "it is a revocable lease."
Registration-first, as you describe MOFCOM doing here, is more measured than an immediate ban list. The deeper civilizational question remains whether frontier containment protocols become the architectural precedent for everyone's devices.
A Human-Machine Intelligence strategy would keep heavy compute as a specialized resource without making personal machines run on someone else's permission slip.
The self-classification duty seems like the real pressure point. If a model can qualify through a top-ten specialist benchmark, whose ranking is authoritative when leaderboards move weekly or disagree? A binding reference list—or a safe harbor for good-faith classification—may matter more than the nominally simplified licence.
Announcement No. 48 is signed by two ministries, MOFCOM and the Ministry of Science and Technology, both in its masthead number and again under the text. The body named on it as issuing it is MOFCOM's security and control bureau, and the date is 16 October 2026. Alongside the Foreign Trade Law, its opening sentence names a second instrument as the basis, the Regulations on the Administration of Technology Import and Export. Its scope section counts this consultation at 40 technologies across three fields, 23 in artificial intelligence, 5 in biomedicine and 12 in photovoltaics, and refers the technical scope and the policy explanation to its first attachment.
The consultation has a closing date and three routes. The public may put forward opinions between 16 October and 15 November 2026 by email at jinxianmulu@mofcom.gov.cn, by fax at 010-65197565, or by letter to that bureau at 2 East Chang'an Avenue, Beijing, postcode 100731. Opinions are invited on four subjects, the newly added scope of technologies, the indicators, the control policies and the facilitation measures, and the list is left open.
The announcement sets out the form as well. It asks that the subject line of the email, the first page of the fax and the envelope each carry the label it specifies, which it gives in Chinese, and that the submission state a contact person, their work unit and contact details. A submission made in the name of a unit shall bear that unit's seal.
Interesting article. It looks like there’s also a parallel here with foreign investment screening, particularly where a change in ownership or control triggers the requirement even though the technology itself remains in China. Different legal mechanism, of course, but potentially another consideration for cross-border acquisitions and restructurings.
Your reverse-question framing is the useful part. For years the fight was about who can export GPUs into China. Now China is asking which AI technologies it will allow to leave. Intelligence is being treated like energy or defense tech—not simply software.
I take that seriously. In my response to Hendrycks, Schmidt and Wang's Superintelligence Strategy, I argued that the hazard is not only foreign diversion of chips. It is the precedent: "silicon runs only by grace of third-party permission." Controls built for a temporary standoff "have a habit of hardening into permanent infrastructure." And compute access that hinges on continuous approval isn't leverage; "it is a revocable lease."
Registration-first, as you describe MOFCOM doing here, is more measured than an immediate ban list. The deeper civilizational question remains whether frontier containment protocols become the architectural precedent for everyone's devices.
A Human-Machine Intelligence strategy would keep heavy compute as a specialized resource without making personal machines run on someone else's permission slip.
https://www.thesingularityproject.ai/p/superintelligence-strategy-the-perfect
The self-classification duty seems like the real pressure point. If a model can qualify through a top-ten specialist benchmark, whose ranking is authoritative when leaderboards move weekly or disagree? A binding reference list—or a safe harbor for good-faith classification—may matter more than the nominally simplified licence.