On 17 September 2026 the Center for a New American Security published the insight "America Must Expand AI Export Controls Amid Talks with China," by Michelle Nie and Daniel Remler. The piece anticipated the Trump–Xi meeting of 24–25 September and argued that the "controls failed" narrative is incomplete: the U.S. still leads in compute and models, and much of China's recent progress still depends on American and allied technology.
The first gap—and this piece's focus—is remote access. Per CNAS, Chinese firms are accessing substantial compute, including export-controlled Nvidia Blackwell chips, through cloud providers in Malaysia, Thailand, Indonesia and elsewhere. One cited estimate puts remote access at around one million H100-equivalents—about one-twentieth of world compute—used in part to train increasingly capable models.
The legal gray zone: the U.S. Commerce Department does not currently interpret its authority to cover remote access to export-controlled chips. The Remote Access Security Act (H.R. 2683) passed the House on 12 January 2026 but remains stalled in the Senate (Banking). Even if enacted, it would require specific Commerce regulations that do not yet exist. CNAS also points to the MATCH Act as a bipartisan vehicle to close gaps.
The second gap is adversarial distillation: Chinese labs use more capable U.S. models to train their own systems. CNAS cites Anthropic's report on the largest distillation attack it has observed, attributed to Alibaba (3,500 accounts, 151 million Claude exchanges). The FBI, NSA and CISA have called for sharing information on distillation campaigns, but antitrust uncertainty slows lab-to-lab coordination.
The third gap is allied servicing of chipmaking equipment inside China. Maintenance by ASML and Tokyo Electron remains legal and can extend DUVi machine life to ~30 years, with tacit know-how transfer. CNAS argues allies are unlikely to stop without sustained U.S. pressure.
The insight answers Beijing's signaling—Moonshot's Kimi K3, reported DUVi production by Aishengna and Yuliangsheng—that seeks a narrative of inevitability. CNAS counters with data: Goldman estimates China can domestically produce only ~8% of the advanced AI chips it needs; SMIC's 7nm yield (~26%) versus TSMC (~90%); Huawei's Ascend 950 at roughly half the power of a four-year-old H100; and combined 2026 Chinese DUVi output (~17 systems) versus ~130 DUVi ASML expects to ship.
Assessments by the UK AI Security Institute and the U.S. Center for AI Standards and Innovation placed Kimi K3's cyber capabilities about five to six months behind leading U.S. models; SaferAI found Z.ai's GLM-5.2 two to four months behind the U.S.-led frontier. Chinese executives (Tencent's Martin Lau; Z.ai's Zhang Peng) have publicly acknowledged GPU shortages.
Neural Pulse reads the insight as a reminder that controls were never meant to "prevent indigenization forever"—China has pursued it since at least the 2014 fund—but to lengthen the U.S. lead. On that metric, Nie and Remler argue, controls are working: China remains a fast follower in AI and lags in leading-edge semiconductors.
The risk of leaving gaps open is eroding that lead: distillation extracts capabilities that cost billions to train; remote access harnesses H100/Blackwell without building them; servicing transfers knowledge of machines that took decades. Chinese models, CNAS warns, are already reportedly used by the PLA for military and intelligence operations.
For post-summit policymakers the agenda is concrete: pass MATCH and the Remote Access Security Act; clarify Commerce authority over remote access; reduce antitrust uncertainty for collective defense against distillation; and press the Netherlands and Japan on China servicing. The window, CNAS says, is narrow: leverage lasts only while China depends on allied systems.
Neural Pulse adds a method note: the "1 million H100-equivalents" figure is a cited estimate, not an audited census. Treat it as an order of magnitude, not millimeter precision. The same applies to yields and DUVi timelines: they are analyst reconstructions.
Close: while the U.S.–China incident channel owns diplomatic headlines, CNAS reminds readers that the compute contest also runs through racks in Johor and Jakarta. Without closing remote access, export control on the physical chip is a door with the window open.
