Put the chip down. Look at the board. The really expensive engineering inside an AI cabinet sits, in large part, on a twenty-something-layer multilayer PCB (printed circuit board — the substrate that carries every trace between every chip in the box) packed with high-speed routing. By industry estimates, the PCB stack inside a single NVIDIA GB200 cabinet is worth on the order of $170,000. The real supply-chain divergence between Huawei and NVIDIA doesn't live on the GPU die. It lives on that board, on its copper-clad laminate, and on who gets to decide whose board to use.
Huawei and NVIDIA take opposite paths at this layer. One is closed and self-reliant. The other is open but tightly controlled. The difference is deeper than "domestic vs imported."
Protocol is supply chain
Start from the conclusion of the previous piece. Huawei replaced three protocol layers — PCIe + NVLink + InfiniBand — with a single in-house protocol, UnifiedBus (UB). NVIDIA keeps the layered stack of NVLink + PCIe + CXL. This isn't only a technical choice. It directly shapes the supply chain.
Full-stack in-house means Huawei refuses to outsource even the interconnect protocol. Protocol, memory addressing, system software, board-level hardware — all of it sits on one vertical line under one roof. The cost: a closed ecosystem that only serves Ascend clusters, with no third-party way in. NVIDIA goes the other direction. NVLink is proprietary, but the CUDA + NVLink ecosystem is open — every major framework and vendor can plug in. NVIDIA gives up "one protocol to rule them all" and gets back "supported by everyone."
The closed side has tight control and few seams. The open side has wide reach and lots of dependencies. Almost every other difference downstream — vendor lists, control models, which suppliers get rich on which order — grows out of this one root choice.
Two boards, two almost non-overlapping vendor lists
NVIDIA's side: clean list, global. The lead HDI-board supplier is China's Victory Giant (胜宏科技), which builds the GB200 Compute Tray and Switch Tray and has become NVIDIA's largest GPU-PCB supplier — with global share of AI-server PCBs above 30%. High-layer-count boards are led by WUS (沪电股份) in China and TTM in the US. The copper-clad laminate (CCL — the dielectric sheet that gives the PCB its insulating layers and its high-speed character) is migrating M8 → M9, supplied by Korea's Doosan, Taiwan's Taiwan Union Tech (台光电子), and Shengyi (生益科技), which broke into NVIDIA's chain. Concretely: the GB200 Compute Tray is 22 layers of HDI (18 M8 + 4 M4). GB300 switches back to a UBB + OAM design — a UBB (Universal Baseboard, the single full-cabinet baseboard) of 18 layers (14 M8 + 4 M4), ASP per UBB around $750–800, supplied by WUS and TTM. The Switch Tray's CCL is sole-sourced from Taiwan Union Tech — unit price jumped from ~$2,000+ in the M7 era to ~$4,000+ for M8U.
Huawei's side: short list, deep, local. The Ascend PCB anchor is Shennan Circuits (深南电路) — share of Huawei AI-server PCBs above 30%. Ascend-related boards are mostly built at its No. 2 plant and its W8 plant in Wuxi. On N+M-class high-layer-count processes its technical capability stands out, and it's one of very few mainland firms producing 14-layer-and-above FC-BGA (Flip-Chip Ball Grid Array — the chip-mounting package substrate) with ABF (Ajinomoto Build-up Film — the high-end dielectric for advanced packaging) carrier-board capability. That ABF capability is the link that makes Ascend's dense interconnect possible. Around Shennan: Huafeng Technology (华丰科技) for high-speed connectors. Xinsen (兴森科技) for Ascend ABF carrier boards. Shenling Environment (申菱环境) for liquid cooling — roughly 70% share. Brokerage estimates put Ascend's overall domesticization rate above 82%, and the per-card PCB value of the Ascend 910C is likely above the H100's roughly $407.
Worth naming the asymmetry: Shennan has not made it into NVIDIA's core chain to this point, and its Google share is low. That's the real difference between Shennan and Victory Giant / WUS — the latter two are players inside the global chain. Shennan is the anchor of the domestic chain. The same high-end PCB capability serves two non-overlapping systems.
Two control models
NVIDIA's "open + strongly controlled" model has structure. On the customer side, the hyperscalers (CSPs — cloud service providers) get the MGX modular spec, which lets them pick their own peripheral components — that's the open face. On the supply side, NVIDIA uses "buy and consign" (NVIDIA designates and contracts the parts; the CSP / ODM accepts and pays) to tightly control PCB, cold plate, and other high-value materials. That control is selective: NVIDIA picks the high-value items itself, and lets the contract manufacturers (Foxconn et al.) buy commodity capacitors and resistors themselves. Upstream of that, NVIDIA is locked all the way through to TSMC — fabrication, CoWoS advanced packaging, and the future CPO roadmap — which is the foundation of both its technology lead and its supply stability, and exactly the path Huawei would take if it could. At the ODM tier it's bound to Hon Hai (Foxconn), Quanta, and Inventec, and it's been shifting parts of PCB / PCBA capacity out of Taiwan to Mexico and Thailand (WUS's Thai plant has become one of its share-winning edges) — a global footprint that hedges geopolitical risk, though overseas build-outs aren't easy (operations and talent shortages are real). NVIDIA's control over the supply chain comes from being the chain anchor that every supplier wants to win.
Huawei's "closed and self-reliant" model isn't designed; it's forced. Under export controls, "global diversification" is not on the menu. The only available move is to pull the chain inside the border as far as possible — deep partnerships with Shennan, Shengyi, Shenling — trading the certainty of domestic orders for supply-chain autonomy. Control comes from vertical integration and state industrial intent, not from chain-anchor status. That autonomy carries a cost: closed is a forced choice. While building the domestic chain, Huawei loses the global frontier — TSMC's manufacturing, EUV lithography, the most advanced packaging. The bill won't show up this quarter. It will show up in the iteration cadence over time.
The asymmetric landing: the one layer that can't be broken through is HBM
Sort the chain layer by layer by "can this be domestically substituted." You get a clean answer.
Protocol layer — Huawei builds it. UnifiedBus is full-stack in-house. No external dependency.
Board level — Huawei is essentially self-sufficient. Shennan + Huafeng + Shenling cover the PCB, connectors, and liquid cooling of an Ascend server without going outside the border. Export controls don't bite here.
Copper-clad laminate (the layer above the board) — China is structurally strong here. Shengyi's high-speed CCL share went from 18% in 2023 to 27% in 2025, and it supplies both chains — main supplier of the Switch-Board CCL at NVIDIA (above 50% share), plus Huawei.
Logic die — SMIC is climbing. Yield trails TSMC, but with enough capacity, million-die Ascend volumes aren't the blocker.
Going up the stack box by box, Huawei finds a domestic answer at almost every cell, or will eventually find one. Until the last cell. HBM. That cell can't be broken through.
Not "temporarily can't" — structurally can't. The reason isn't a single point, it's a stack of four:
- HBM is not ordinary DRAM. It needs TSV (Through-Silicon Via — vertical wires drilled through the silicon to stack dies) and advanced stacked packaging. CXMT trails the leaders by only one to two years on commodity DRAM, but on HBM the gap is another order of magnitude. Best case next year: ~2 million HBM stacks, enough for 250–300K Ascend NPUs. A drop in the bucket.
- The foreign stockpile is one-off. The roughly 13 million stacks already smuggled in are enough for about 1.6 million 910Cs, and they burn down by year-end.
- There's no substitute. GDDR and LPDDR bandwidth can't sustain frontier-model training or large-scale inference. Downgrading is exiting the game.
- This layer is the precise target of export controls — it's exactly what's being blocked. And it's the only thing Beijing specifically asked for in negotiations, not TSMC capacity, not lithography equipment.
Stack the four together and what it means is: without foreign HBM, Huawei won't make even 1 million Ascend NPUs next year. SemiAnalysis runs the math to its sharp end — assume no smuggling, and China's Ascend production next year is lower than this year, not higher. Logic, process node, board, protocol — all of these are catching up. Only HBM is moving the other direction.
That's the real landing of the whole piece, and it lines up exactly with the system fight's conclusion. The protocol unification, Shennan's high-layer board, Shengyi's CCL — everything in the earlier sections exists to set up one contrast: those are layers Huawei can build. No matter how loud an article makes them sound, none of them gets around the one layer Huawei can't build. UnifiedBus can let 8,192 cards behave like one machine. Shennan can stack a board to forty layers. As long as those handful of HBM stacks on every card still get airlifted from Korea, the whole "ten-thousand-cards-as-one-machine" building has a borrowed foundation.
Huawei wins every layer it can build itself. What pins it is the one layer it can't build, can't buy, and can't substitute. How far closed self-reliance carries depends not on how tightly it's closed, but on how long those HBM stacks outside the wall hold up. That countdown isn't a switch Huawei has a hand on.
Source note
PCB / CCL / packaging vendor mapping and the GB200 / GB300 layer-stack numbers come from industry teardown reports and brokerage research. The HBM math (~13M stacks smuggled, ~2M from CXMT next year, "no foreign HBM → <1M Ascend next year") is drawn from Huawei Ascend Production Ramp: Die Banks, TSMC Continued Production, HBM Is The Bottleneck (SemiAnalysis, 2025-09). UnifiedBus / system-architecture facts trace to Xu Zhijun's Huawei Connect 2025 keynote and the MWC Barcelona 2026 debut. The previous piece — Why Huawei's Bet Isn't on the Chip — covers the system-side argument this piece extends.