Commerce closes the China cloud loophole, but the chokepoint is leaky The Commerce Department's Bureau of Industry and Security issued an interim final rule, Addressing Remote Access to Advanced Semiconductors, effective November 19, 2026, requiring a licence for any data center worldwide that supplies more than 12 million aggregate Total Processing Performance to a company headquartered in China, Macau or four other US-embargoed countries before it can take delivery of more advanced American chips. BIS separately named Brazil, Indonesia, Malaysia, Singapore and Thailand as presenting "a higher-than-normal risk of diversion," and the rule gives data centers already serving Chinese clients a 90-day grace period to roughly mid-February 2027 to obtain a licence, with public comments due January 19, 2027. The rule followed a letter from Sen. Maria Cantwell to Commerce Secretary Howard Lutnick sent one day before publication and formally rescinds the Biden-era AI Diffusion framework. A data center in Singapore can sell a Chinese company compute time without a single chip ever crossing into China. That is the mechanism the Commerce Department’s Bureau of Industry and Security says it shut down on November 19, when its interim final rule, Addressing Remote Access to Advanced Semiconductors, took effect immediately on publication in the Federal Register. The rule attaches the licence requirement not to where a chip is shipped but to who ends up using its output: any data center anywhere in the world that supplies more than 12 million aggregate TPP — Total Processing Performance, BIS’s own stand-in for raw compute speed — to a company headquartered in China, Macau or one of four other countries under a US arms embargo now needs a licence before it can take delivery of more advanced American chips. As Al Jazeera reported https://www.aljazeera.com/economy/2026/11/20/us-lawmakers-push-for-tougher-ai-chip-rules-amid-china-export-crackdown , US lawmakers who back the new rule still want more. That is the odd shape of this story: a regulation closing a specific hole, arriving under specific political pressure tied to a specific scandal, while the people who demanded it say the hole was never just the one BIS plugged. The chokepoint moves to the data center Until November 19, US export controls on advanced AI chips worked the way tariffs work on goods: control the border crossing, not the use. A chip sold to a data center in Thailand or Brazil was outside BIS’s reach once it cleared customs, even if a Chinese firm rented 100% of its output by the hour to train a model. That arrangement is what officials and lawmakers called the remote-access loophole, and it is the mechanism named in Sen. Maria Cantwell’s letter to Commerce Secretary Howard Lutnick, sent one day before the rule’s publication, accusing cloud operators of “exploiting a loophole in U.S. export controls by using advanced American chips in foreign data centers even when those chips are restricted from the People’s Republic of China .” Where the new licence sits The remote-access loophole, and the chokepoint BIS added on Nov. 19, 2026 | From | To | How | |---|---|---| | Nvidia/AMD chip, exported | Data center abroad e.g. Singapore | old rule: unrestricted | | Data center abroad e.g. Singapore | Compute rented by the hour | | | Compute rented by the hour | China-headquartered company | no chip crosses border | | BIS licence checkpoint new, 12M aggregate TPP line | Data center abroad e.g. Singapore | now requires licence if 12M TPP flows… | Based on Bureau of Industry and Security https://www.federalregister.gov/documents/2026/11/19/2026-23529/addressing-remote-access-to-advanced-semiconductors The 12-million-TPP line is the rule’s technical core; an earlier account on this site https://rews.cc/a/commerce-sets-a-758-chip-line-to-stop-china-s-cloud-workarou-db7f59 worked out that the threshold amounts to roughly 758 of Nvidia’s top-end chips pooled in a single cluster feeding a Chinese customer. Cross that, and the data center’s host needs a licence, regardless of which country the center sits in. BIS separately named five countries — Brazil, Indonesia, Malaysia, Singapore and Thailand — as presenting “a higher-than-normal risk of diversion,” meaning any advanced chip shipment there now needs a licence by default, with an exemption for consignees headquartered in the United States that meet a reporting requirement. Data centers already serving Chinese clients get a 90-day grace period, to roughly mid-February 2027, to get a licence in hand. Public comments on the rule are due January 19, 2027. The same rule formally rescinds the Biden-era AI Diffusion framework, which the Trump administration had already stopped enforcing. A scandal drives the timeline The rule’s speed is not explained by the merits alone. It follows the resignation of Nvidia’s Jensen Huang as chief executive on November 12, after Al Jazeera reported separately https://www.aljazeera.com/news/2026/11/12/ai-titan-jensen-huang-out-as-nvidia-ceo on his affair with Alice Huang, former chief executive of the Singapore-based Nvidia customer Megaspeed, whom the New York Times reported had been recruited by China’s Ministry of State Security. Huang’s fall, and Congress’s response to it https://rews.cc/a/mast-says-chip-export-curbs-will-likely-ride-the-defense-bil-38aa8c , is the political backdrop against which BIS moved. Notably, BIS’s own announcement states the rule was “written by and at the direction of” the Treasury Department and the National Security Council — not Commerce, which normally drafts export-control rules and which was, by this point, itself under scrutiny. A secretary with skin in the game That scrutiny centers on Lutnick. Treasury Secretary Scott Bessent has said four federal agencies backed blacklisting Megaspeed for months while Commerce, which controls the entity list, held out; the New York Times reported Lutnick separately pushed to narrow the scope of the restrictions now in force. Lutnick’s son Brandon chairs Cantor Fitzgerald, which held hundreds of millions of dollars in Nvidia stock and cut that position by about 49% between September 2025 and March 2026, as pressure to blacklist Megaspeed mounted. House Judiciary and Foreign Affairs ranking members Jamie Raskin, Gregory Meeks and Sydney Kamlager-Dove sent Cantor Fitzgerald a document request over it. Secretary Lutnick’s reported obstruction may have resulted in billions of dollars of restricted AI chips being diverted to China. A Commerce Department spokesperson disputed that account, saying the agency “conducts all enforcement strictly by the book” and “enforces aggressively across the board, including with respect to chip smuggling.” The pressure on Lutnick is not purely partisan: Republican Sen. Tom Cotton and House Foreign Affairs Chairman Brian Mast have sent their own letters demanding Commerce investigate Megaspeed and its affiliate Speedmatrix, with Mast writing that smuggling reports implicate the companies’ “broader corporate network, all under Jensen Huang’s nose.” The scale those letters are chasing is suggested by a separate figure Congress published this month: the Trump administration approved exports of 879,000 Nvidia and AMD chips https://rews.cc/a/commerce-approved-879-000-nvidia-and-amd-chips-for-china-con-a36245 to Chinese companies between December 2025 and May 2026, under rules that predate the November 19 rule. Lutnick and Huang are expected to testify before Congress next month; House Democrats gain subpoena power once the new Congress convenes in January. What lawmakers want next Rep. Pramila Jayapal, who introduced a National AI Charter Act framework on October 1, told Al Jazeera the export controls “should have been in place, certainly,” but framed chip leakage as one symptom of a larger problem: “We should not have this consolidation of all the AI knowledge in a couple of companies because that poses an enormous threat on multiple levels, including national security.” Her framework https://jayapal.house.gov/2026/10/01/jayapal-introduces-legislative-framework-establishing-national-charter-system-to-rein-in-ai/ — a policy document, not introduced legislative text — would require every AI company to hold a federal charter, modeled on bank charters and on pre-release federal testing regimes for nuclear materials: round-the-clock federal oversight, government-run testing before release, adversarial stress tests, a government-controlled kill switch, and a structural rule barring any chartered company from owning both AI infrastructure and AI systems at once. Companies valued above $1 billion would need congressional approval for a charter; smaller ones would apply to a new National AI Charter Authority. Rep. Everton Blair said Democrats intend to use incoming committee leadership to push the issue, arguing “the very industry is asking to be regulated by Congress.” Whether the November 19 rule can actually be enforced is a separate question from whether lawmakers want it. The rule’s language reaches “indirect access” — it covers not just a data center renting compute directly to a Chinese firm but any intermediary that trains models for one or takes part in that training, a standard that requires BIS to trace relationships rather than shipments. Stephen Ezell of the Information Technology and Innovation Foundation has argued enforcement will be difficult, and that China will keep building its own chip industry regardless of where Washington draws the line. The scale of what’s being guarded explains the stakes: Sherwood News, citing an Oxford University dataset published on SSRN, reported that only 32 countries https://sherwood.news/tech/ai-data-center-country-ranking-concentration/ host any specialized AI data center at all, and that 95.5% of the world’s installed AI accelerators run on chips designed by US companies. A separate Oxford figure cited by Al Jazeera put US firms’ global AI computing hubs at roughly 90, with 63 more under construction outside the US as of 2025 — the inventory of exactly the kind of overseas infrastructure the new rule now has to police. President Trump has pledged to penalize US companies for building data centers abroad at all — “We’re going to put tremendous penalties on companies that try and do it,” he said — a stance in some tension with an administration simultaneously pushing to add domestic data centers over local objections on energy and environmental grounds. The 90-day compliance clock and the January 19 comment deadline will test how much of this is enforceable rather than declaratory. Whether the Lutnick investigations produce documents before or after Democrats gain subpoena power in January is the more immediate number to watch.