China’s “domestic capital” AI doctrine makes US AI capex less dollar-dominant—look for memory suppliers to win while GPU capex gets quietly rerouted

Reuters reports CXMT raised 57.92B yuan (~$8.6B) in its IPO and that its market capitalization jumped to 3.3T yuan (~$487.73B) after its Shanghai debut.

That matters because CXMT is not only “a chip company”; it’s a market-made pipeline for memory capacity—exactly the kind of bottleneck investors map onto AI buildouts.

In a grant/subsidy model, the state can underwrite projects even if the marginal return is uncertain. In a capital-markets model, projects must clear pricing and liquidity conditions: what gets funded, when, and at what scale depends on market appetite.

That changes timing and mix. Component makers that are part of onshore “capacity narratives” can get funded earlier (or at least financed through equity), while the US stack can see a more muted incremental demand effect—because the market “imports less capacity narrative” per unit of AI compute demand.

This is why CXMT’s debut matters even to non-China portfolios: it’s a proof that AI capacity can be funded through tradable market vehicles, not only through off-cycle state support.

Reuters reports DeepSeek told prospective investors in its second fundraising round it is suspending the deal for now (citing people familiar with the matter). Reuters also provides context that the planned next round was at a valuation of about 500 billion yuan (~$74 billion), following an earlier round of about $7.4B.

GPU demand is the visible story, but memory is the constraint that turns models into real data-center builds. That’s why a capital-markets prototype like CXMT shows up in the same portfolio logic as DRAM/NAND and why it can re-price the marginal US capex chain.

On the US side, Micron’s own filings show how directly investment and shipments depend on capex timing. In Micron’s SEC filing (for the quarter ended May 28, 2026), the company states it completed a wafer fabrication facility acquisition in Taiwan for cash consideration of $1.8B and expects product shipments from that site to begin mid-calendar 2027.

CXMT trades at a very high price-to-sales (~54.5x) in the latest data snapshot, which is consistent with a “capacity story” being priced ahead of cash-flow realization.

That’s exactly what you would expect if Beijing’s doctrine increasingly relies on capital markets: the market can price forward the capacity build, and the company can then convert that priced equity into capex plans.

If a larger share of AI capacity funding in China happens through equity and domestic capital plumbing, the marginal dollar capex intensity for US hyperscaler supply chains can look weaker even while global AI demand grows.

This isn’t a claim that AI demand falls. It’s a claim about who captures the financing impulse first. When memory capacity can be funded earlier via onshore equity narratives, part of the cycle’s capex gravity can shift away from the US dollar-denominated trading corridor.