The $30B US–China Deal Doesn't Cover the Parts AI Buildout Needs — But Chinese Capital Is Coming Anyway

China and the US agreed to cut tariffs on $30B of reciprocal trade in 'noncritical' goods — agricultural and energy-related products — and to launch a bilateral AI dialogue, with the first substantive session scheduled for November. Treasury Secretary Scott Bessent called it the 'most favored nation' framework first floated at the May Beijing summit; the broader trade truce was extended by two months, from Nov 10 to Jan 10, to keep negotiations alive.

CNBC's summit takeaway was blunt: the visit produced 'bonhomie and pageantry' but 'few concrete agreements' on the technology that matters most to investors. Reuters noted the agreement was procedural — a 'mechanism' rather than a 'deliverable' — and the AI dialogue's value depends on whether the November talks produce anything enforceable. For markets, the asymmetric fact is this: the diplomatic opening is new, the technology decoupling is intact.

US hyperscaler AI capex is now projected at roughly $700B-$745B for 2026 — Microsoft, Amazon, Alphabet and Meta combined — up about 60% YoY. JPMorgan CEO Jamie Dimon has said the industry could clear $1T next year. Yet even with that spend, the buildout is bottlenecked at the physical layer, and a meaningful share of that physical layer sits in Chinese factories.

On Aug 4, the Trump administration began drafting a ban on imports of new models of Chinese data center components — exactly the optical, transformer and battery categories above. The Sep 26 deal did not touch that workstream. CSIS estimates that semiconductors are 54 cents of every dollar of data center buildout, and the US currently fabricates only ~12% of the chips it uses; the upstream components needed to host those chips are even more import-dependent than the chips themselves.

The CSIS scenario analysis makes the cost stakes concrete: a broad semiconductor tariff with no carve-out could lift the US data center buildout cost from $1.44T to $4.14T by 2030 — a 53% increase. That is the implicit price tag of any move to fully decouple from Chinese supply. The Sep 26 deal does not pay that price. It doesn't change it either.

Chinese capital is responding to the US buildout on three tracks: providers raising capital for global (not US) buildout, Chinese-listed component makers opening offshore listings to fund US-aligned customers, and prefab builders pitching directly to US hyperscalers.

The pattern is clear: most Chinese capital is going into the Chinese buildout (Alibaba, Tencent, ByteDance, Baidu, Kingsoft Cloud), while the Chinese suppliers most exposed to the US market — optical transceivers, prefab construction — are either raising offshore capital (Zhongji Innolight's Hong Kong listing) or courting US customers directly (Brightray, Eoptolink). Both expose them to the Aug 4 component-ban track.

The Sep 26 deal is, for now, a thinner headline than the capex numbers underneath it. The supply chain is bottlenecked at power and cooling, and the listed beneficiaries are almost entirely US-domiciled. Q2 2026 earnings released through July and August give the cleanest read.

Eaton is the cleanest proxy for the constraint Sep 26 didn't fix: data center orders at its Electrical segment are up 85% YoY, and the company's Electrical backlog grew by $5B in a single year. Vertiv's $15B backlog now provides 'revenue visibility well into 2027,' per CEO Giordano Albertazzi. Digital Realty's record $1.9B backlog is multi-year. Equinix raised its FY26 outlook twice in four months — to $10.21–$10.29B — as AI-related bookings accelerated.

On the power-utility side, the demand pull is structural: data centers are projected to need ~74 GW by 2028, with a ~38 GW shortfall in delivered capacity. That shortfall is the load growth that names like NextEra Energy and NRG Energy are racing to serve. The Sep 26 deal does not change the load; it changes only whether the equipment to serve it can come from China.

Short-term (days to quarters): the November AI dialogue is the next hard date. Any deliverable there — a verification mechanism for distillation complaints, a tariff exclusion for legacy optical or transformer SKUs, or simply a confirmed second session — moves both sides of the trade. The Aug 4 component-ban draft remains in motion; if finalized in current form, optical transceivers and select electrical gear would face immediate disruption, lifting US-made or Mexican-made capacity to multi-year backorder status.

Long-term (1–3 years): the structural driver is unchanged. The US is still short ~38 GW of data-center power by 2028, and replacing Chinese components in optical and transformer categories on a multi-year timeline requires capacity that does not yet exist. The Trump administration's tariff posture, the Federal Communications Commission's optical-transceiver rule updates (Sept 11 narrowed but did not kill the proposed ban), and Chinese retaliation risk all layer on top of the same underlying capex wave.