China's Stack
The biggest risk on the map, in both directions: China builds most of the optical modules and is closing on commodity DRAM — but has no EUV, no HBM3E, no laser, no DSP. And it holds the minerals the West's lasers need.
China builds most of the optical modules and a growing share of ordinary memory, but it cannot buy the EUV machines, cannot yet make the fast HBM3E memory, and cannot make the laser or the DSP chip inside its own optical modules. It does control the raw minerals the West's lasers need. So every sanction has a counter-sanction, and the companies on this map are exposed to both.
What it is
What China has: CXMT at ~10% of global DRAM revenue and ~300k wafers/month after a 466% IPO pop; InnoLight and Eoptolink at ~60% of NVIDIA's 800G modules; Huawei's Ascend line with in-house HBM (HiBL/HiZQ) from the 950 series; Montage's PCIe 6 / CXL 3 retimer on the PCI-SIG integrators list; JCET's XDFOI 2.5D packaging line in production. What it lacks: EUV, HBM3E at scale, EML lasers, optical DSPs — and a leading-edge logic die that isn't a pre-ban TSMC stockpile.
The business
For the tickers on this map China is a risk in both directions. Downside: Beijing's ban on buying NVIDIA's China-specific chips (H20, RTX Pro 6000D; a B30A approval now unlikely) has all but closed the China TAM; the gallium/germanium reprieve to the US expires 27 Nov 2026; and a proposed FCC rule to bar Chinese transceivers from US networks — not enacted, status unverified — would hit InnoLight and Eoptolink's core customers if it ever landed. Upside for the incumbents: every year China stays without EUV and HBM3E is another year of pricing power for ASML, SK Hynix, Lumentum and Marvell. Watch the BIS review of the HBM rule (3A090.c) due December 2026 — it is the next scheduled shock.
Bottleneck severity
The chokepoint runs both ways. China dominates transceiver assembly (7 of the top 10 makers) and is closing on commodity DRAM scale — but cannot buy an EUV machine, cannot yet make HBM3E, cannot make a high-speed DSP, and localises under 15% of its 25G+ lasers. Meanwhile it holds the gallium/germanium/rare-earth inputs that the West's lasers and magnets need. Every export-control move now has a counter-move, and both sides' supply chains are hostage to it.
Who makes it
CXMT688825.SH
China's only scaled DRAM maker: ~10% of global DRAM revenue in Q2 2026, ~300k wafers/month heading for 350k, and a 466% first-day pop on its July 2026 STAR listing. But its IPO capex names no HBM line; HBM3 is a 2026–27 target and HBM3E not before 2027 — years behind on the part AI actually needs.
SMIC688981.SH
China's leading logic foundry; fabs Huawei's Ascend on 7nm-class DUV multi-patterning with no EUV. Held near 7nm through 2026, which caps every domestic accelerator's performance per die.
Huawei / HiSiliconPRIVATE
Ascend AI accelerators. Roadmap: 910C (2025) → 950PR/950DT (2026) → 960 (2027) → 970 (2028). From the 950 series it uses in-house HBM: HiBL 1.0 (128 GB, 1.6 TB/s) and HiZQ 2.0 (144 GB, 4 TB/s). Today's 910C still runs on stockpiled Samsung/SK Hynix HBM2E. HBM supply, not compute, gates its volume.
InnoLight300308.SZ
World's largest transceiver maker: ~35% of global 800G share, Q1 2026 revenue +207% YoY on the 1.6T ramp; ~60% of NVIDIA's 800G volume with Eoptolink. Depends on US DSPs + JP/US lasers; Thailand plants dodge tariffs.
Eoptolink300502.SZ
#2 transceiver maker globally (overtook Coherent in 2025; revenue +189%). Cost-performance leader; Thailand Phase II live for US-bound 1.6T. Same DSP-and-laser dependency as InnoLight.
Montage Tech688008.SH
CXL memory controllers + DDR5 RCD/MRDIMM buffers. China's CXL silicon play.
ASMLASML
Monopoly on EUV lithography — every leading-edge logic + DRAM die needs it. Ultimate chokepoint.
LumentumLITE
The only vendor shipping 200G/lane EML lasers at volume — the exact part 1.6T optics needs — with all EML capacity under long-term agreements through 2027. NVIDIA invested $2B (Mar 2026) to lock it. Purest laser-chokepoint play.
MarvellMRVL
~70% optical-DSP share + custom AI silicon + Celestial AI photonic memory fabric ($3.25B). Most complete optical stack in public equities.
Metrics that move this layer
- CXMT HBM capexany amended filing naming an HBM3/HBM3E line
- BIS 3A090.c reviewDecember 2026 — tighten or loosen the HBM rule
- NVIDIA China banenforced, relaxed, or replaced by B30A?
- Domestic 25G+ laser share<15% today; 25–40% would de-risk the module makers
- JCET/Tongfu hybrid-bond yieldChina's realistic route to competitive accelerators