Markets
China runs a more closed AI stack where advanced HBM access is often the stuck step. This door follows who gets paid inside that gate, what would loosen it, and the dates to watch — in plain English. Education only, not investment advice.
China
CXMT is paid on DRAM it already shipped: RMB 118.714 billion of third-quarter sales, at an 89.53% gross margin. Parent shareholders are paid on RMB 54.435 billion of profit. Minority owners of the operating fabs are paid on a larger share of the growth. The stuck step is the next bay.
China
Naura, China’s biggest maker of chipmaking machines, says orders keep growing, but its third-quarter sales rose just 1.3% because some parts are in short supply. Customers are paying ahead and the machines are waiting. The stuck step is one small part.
China
Cambricon on 23 Oct printed Q3 sales of RMB 3.53bn. Foundries are paid in cash and notes. Wafer and memory slots are still the stuck step.
China
China's official AI compute count rose 146% in a year, to 2,592 EFLOPS. The power used by its data services rose 35% in September. Chips and racks are being installed faster than paying work reaches them. That is the stuck step.
China
The first 10,000-card pool of Huawei's new Ascend 950DT chips is due by the end of October in a hall built for 30,000. The building, the power and the cooling are done. What nobody has shown yet is a weeks-long training run on it.
China
Hygon on 19 Oct printed nine-month sales of RMB 14.96bn. Foundries are paid in cash; advanced HBM is still the stuck step.
China
China's legacy 8-inch foundries have been running full on AI-server power semiconductors since the start of the year, while TSMC and Samsung cut capacity. Order books stretch about six months. Foundries are paid first; server builders wait on the wafers.
China
NBS printed September prices on 14 October. Who gets paid is the device step — storage, computers, phones. The stuck step is electricity: factory power prices are still falling.
China
China's biggest memory maker, CXMT, says it will launch server memory on a new, denser DRAM cell by the end of 2026. It already earned ¥77.6 billion in the first half. The big three are paid first while server DRAM stays short. The stuck step is server testing.
China
China’s markets reopened on 8 Oct and sold SMIC first. The fight is the foundry’s Q3 gross-margin guide of 26–28%: the first hard price tag on Huawei’s wafer-hungry LogicFolding design. SMIC gets paid on wafer invoices. DUV capacity and qualification still set the pace.
China
Alibaba says it will run more than 20 GW of data centers by 2032, and that global shortages limit how fast it can build. Its June-quarter capex hit RMB67.7 billion, up 75%, partly on higher chip-part prices. Chip and parts sellers get paid first. Chips set the pace.
China
October 2026 in one line: SMIC’s 2Q26 HKEX print shows $3.0B revenue, 25.3% gross margin, and 93.7% utilization — who gets paid is SMIC on wafer invoices at available nodes; the stuck step is 93.7% full ≠ leading-edge cash / new capacity still must qualify.
China runs a more closed AI stack: domestic accelerators, local cloud builders, and a policy push for compute capacity. The stuck step for advanced training clusters is often access to leading HBM and the most advanced foreign process nodes, not a lack of ambition on paper.
When advanced memory and packaging access is limited, domestic cards can get more expensive and less capable per watt. That changes who gets paid inside China — local chip designers, foundries, and cloud builders — and how fast the domestic stack can close the gap.
This door tracks the constraint, not a scoreboard of national champions. For the shared mechanisms, see HBM, explained and Semiconductors.
As at September 2026, advanced HBM access remains a binding gate for China's higher-end AI buildout even as policy targets for compute keep rising.
| Step | Who | How they get paid | Status (Sep 2026) |
|---|---|---|---|
| Domestic accelerators | Huawei Ascend and peers | System and card ASP inside China | Demand firm; capability gated by memory/tools |
| Advanced HBM access | Restricted foreign supply; domestic substitutes | Allocation and scarcity pricing | Still the stuck step |
| Local foundry / packaging | SMIC and domestic OSATs | Wafer and package pricing on available nodes | Behind leading foreign nodes |
| Policy compute targets | MIIT and provincial programmes | Subsidies and mandated buildouts | Targets rising (e.g. EFLOPS path to 2030) |
Status is Second Order's read from the latest China HBM Access Pulse and named policy releases.
| Name | What they do | Which step |
|---|---|---|
| Huawei (Ascend ecosystem) | Domestic AI accelerators and systems | Accelerators |
| SMIC | Domestic foundry on available nodes | Chipmaking |
| Domestic cloud builders | Consume local and permitted foreign compute | Demand |
| Domestic memory / packaging attempts | Try to substitute restricted HBM and advanced package access | Memory / packaging |
Examples of roles in the domestic stack — not recommendations. Listings and access rules differ sharply by jurisdiction.
| Date | Event | What to look for |
|---|---|---|
| Ongoing 2026 | Export-control and licensing updates | Whether advanced HBM / tool access widens or tightens |
| Ongoing 2026 | Domestic accelerator pricing prints | Whether scarcity pricing eases |
| Policy cycle | MIIT / provincial compute target updates | Whether EFLOPS targets outrun memory access |
| Quarterly | Cloud capex commentary inside China | Whether spend shifts to inference vs training |
Because export controls and domestic substitutes change who gets paid and which step stays stuck — even when the buyer list differs from the US.
Advanced HBM access (and the tools behind leading nodes) remains a binding gate for higher-end domestic AI clusters as at September 2026.
Not on the global leading-edge finish line today. This door maps the domestic constraint path, which is linked to — but not the same as — the Taiwan and Korea doors.
No. Education only. Access rules and listings vary by country; nothing here is a recommendation to buy or sell.
Updated 19 Sep 2026 · Next update after the next major HBM-access or MIIT compute-policy print
