Where Washington Draws the Line on Chinese Tech

On August 4, the same administration that was drafting a ban on Chinese optical transceivers for American data centers told leading AI companies it would not safety-test open-weight models. ^[1][2] The two moves are not a contradiction. They are the latest instances of a pattern that has been hardening for three months: Washington walls off Chinese hardware where American supply is adequate, retreats from oversight where it is not, and the private sector fills both voids faster than policy can close either one. The hardware pattern is the clearest. Since June, the FCC has banned Chinese equipment in three categories. On June 28, it expanded import restrictions on legacy telecom and surveillance gear from Huawei, ZTE, Hikvision, and Dahua. ^[3] On July 22, it banned Chinese military-grade drones from DJI and Autel and closed a loophole that had allowed Huawei and ZTE components into U.S. equipment. ^[4] On August 4, it began drafting a ban on Chinese optical transceivers for data centers, targeting Zhongji Innolight. ^[1] In every case, American or allied suppliers exist — Cisco and Juniper in networking, Skydio in drones, Coherent and Lumentum in optical components. Against that, two companies designated by the Pentagon as Chinese military entities have gone untouched: CXMT and YMTC, both memory-chip makers. YMTC is also on the Commerce Department's Entity List. Yet CXMT has secured a $7 billion deal with ByteDance and $3 billion with Tencent, is preparing an $8.6 billion IPO, and is building new fabrication plants. ^[5] The reason is not a stated policy — no official has articulated one — but the economics are unmistakable. The AI boom has diverted Samsung, Micron, and SK Hynix capacity toward AI customers, creating a global memory shortage that PC makers began responding to as early as February, when HP and Dell started qualifying CXMT chips for non-U.S. markets. ^[6] By August, HP, Asus, and Acer had integrated CXMT memory into notebooks sold abroad. ^[7] Apple made the pressure explicit. In June, it began testing CXMT chips and petitioned the Trump administration for a license to bypass national-security restrictions. ^[8] Tim Cook called the shortage

what I would characterize as a 100-year flood on memory pricing with exponential increases in memory prices. — Tim Cook

Apple's request has since expanded from China-only devices to worldwide products and now includes YMTC. ^[9] The political resistance is real. Representative John Moolenaar, chair of the House China committee, called the potential partnership

Apple choosing to partner with a Chinese military company would be a grave mistake. — John Moolenaar

Micron CEO Sanjay Mehrotra lobbied Commerce Secretary Lutnick and Treasury Secretary Bessent directly, warning that admitting state-subsidized Chinese memory firms would decimate the domestic semiconductor industry, comparing it to the decline of American steel and manufacturing. ^[9] But the petition remains undecided as of August 5, and the de facto outcome is already visible: U.S. PC makers are using the chips abroad, and the government has not moved to stop them. ^[10][7] The same pattern — action where alternatives exist, retreat where they do not — runs through the administration's approach to AI safety testing. The arc is compressed and sharp. In May, National Economic Council Director Kevin Hassett compared proposed AI reviews to an FDA drug-approval process. ^[11] By June 2, the executive order that emerged was voluntary only, with a 30-day review window — down from 90 days after tech industry lobbying — and explicitly prohibited mandatory government licensing. ^[12]

We’re leading China, we’re leading everybody, and I don’t want to do anything that’s going to get in the way of that lead. — Donald Trump

On June 24, Anthropic's Mythos model penetrated classified Defense Department systems "in hours" during Project Glasswing, yet the voluntary framework stayed in place. ^[13] By late July, White House AI advisor Sriram Krishnan had closed the door on anything stronger.

The rapid progress we’re seeing in AI requires a new approach to testing frontier AI model capabilities that is dynamic, adaptable, and rigorous. — Demis Hassabis

On August 4, the administration informed AI companies it would not safety-test open-weight models at all. ^[2] The retreat was complete: from mandatory to voluntary to abandoned in three months. The private sector is now filling both voids simultaneously. Microsoft is exploring a self-hosted, fine-tuned version of DeepSeek-V4, a Chinese open-source model, to power a lower-cost tier of Copilot — running it on Azure under its own security controls to manage the political risk. ^[14] Apple is lobbying to import chips from two Pentagon-designated Chinese military companies while simultaneously partnering with Intel on a $10 billion domestic chip-manufacturing deal in which the U.S. government holds a 10 percent equity stake. ^[15][9] And Chinese open-weight models from Moonshot and DeepSeek now handle nearly two-thirds of global AI use by token volume. ^[16] The market has already made the decisions the government has declined to make. The bans fall where substitution is possible — telecom gear, drones, optical transceivers. The inaction falls where it is not — memory chips, model testing. And the private sector is routing around both faster than policy can close either one.