At a glance
Winners
Western WFE majors (Applied Materials, Lam Research, KLA, ASML, Tokyo Electron)Nvidia (H20/H20E/B30A China lane)Unitree (pre-IPO) and the Chinese robotics supply baseHuawei / SMIC / CXMT / Cambricon (China self-sufficiency stack)Apple / Qualcomm / MediaTek (asymmetric downside, not upside)Bottlenecks
HBM memory — the single binding constraint on Chinese AI accelerators; domestic CXMT can supply only ~2M stacks in 2026 (250–300k Ascends), so China runs on a finite foreign stockpile.TSV/advanced packaging & Korean memory enforcement — CXMT still needs Japanese TSV tools and OSATs (JCET, TongFu); Samsung/SK Hynix leaks are the main HBM entry point.<100 critical WFE components (Zeiss EUV projection optics, precision mask stages, RF plasma generators) — the highest-leverage chokepoint, currently un-restricted.Frontier-model compute — DeepSeek/Alibaba/Moonshot throttled by compute (multimodal models delayed, R1 served slow); the one thing the US can restrict.Actuators / strainwave-gearbox reliability — Harmonic Drive still the reliability leader; Unitree's QDD wins on cost/speed but Western humanoids gate on high-precision actuation.Risks
Policy reversal / license loosening — H20 resumed to unlock China's rare-earth exports; a B30A approval would hand China ~10× the FLOPs and accelerate DeepSeek/Alibaba.Rare-earth / magnet retaliation — China cut rare earths and magnets to the US; the Lutnick-linked H20↔rare-earth swap shows controls are now a two-way bargaining chip.Smuggling & re-export — H100s reach China via Malaysia/Oracle/Google rentals and shell OSATs; ByteDance trains Seed on US clouds; location tracking is 'technically intractable.'Chinese legacy-chip flood — massive subsidized mature-node capacity dumps on TI/STMicro/Infineon/NXP; the 'China Technology Cycle' erodes Western WFE share long-term.Unitree security / trust barriers — Unitree quadrupeds carry security concerns that bar select Western markets, a headwind to its otherwise dominant deployment path.Catalysts
Foreign HBM stockpile exhaustion (~end-2025) — the moment Huawei can't build even 1M Ascends next year; forces the real test of CXMT's ramp and of Nvidia's China share.CXMT entity-listing decision — SemiAnalysis calls the Biden-era omission a failure the Trump administration 'needs to solve immediately'; listing would slow HBM conversion.Nvidia B30A / next China SKU approval — sets how much frontier compute China gets; SemiAnalysis: only raise the bar when China can ship H20E-class in volume.Unitree Shanghai IPO (~$7B rumored) — the 'birth of the robotics moment'; discloses margins/verticalization and re-rates the Chinese robotics supply chain.0% de-minimis / upstream-component controls — if the US applies SemiAnalysis's blueprint (Huawei network at 0%, <100 components, monthly listing), Chinese advanced fabs could be handicapped within ~6 months. Overview
SemiAnalysis frames US–China technology policy as a full-scale economic cold war whose single most important lever is compute: the US controls ~70% of the world's deployed FLOPs and is trying to keep China off the frontier by choking AI accelerators, wafer-fab equipment (WFE), EUV/DUV lithography and — most decisively now — HBM memory. Their reporting is unusually granular and unusually opinionated: they argue controls are simultaneously (1) genuinely effective at the leading edge (China will make FEWER Huawei Ascends next year, not more, because it runs out of foreign HBM), and (2) riddled with self-inflicted loopholes — the 'fab whack-a-mole' of entity-listing one building while its non-listed twin across the street imports every advanced tool, allied-lag windows where Samsung dumped 7M HBM stacks in one month before enforcement, and Huawei's TSMC 'die bank' of 2.9M+ Ascend dies smuggled via shell companies like Sophgo/Bitmain. The corpus spans the 2022 opening salvo, the 2023 'controls have failed' broadside, the 2024 Fab Whack-A-Mole enforcement manifesto, and the 2025 Huawei Ascend production model. Running underneath is a second, non-semiconductor front: robotics. Here SemiAnalysis is bearish on the West — China is running the BYD/DJI playbook (own the hardest component, bootstrap on hobbyists, verticalize, undercut on price), Unitree is shipping its 10,000th humanoid at a ~$9K BoM and 67% gross margin while Tesla Optimus and Figure ship single digits, and general-purpose robots making robots is cast as an existential industrial threat the US is 'missing.' The throughline: China wants sovereignty over every layer from silicon to tokens to actuators, and predates — so will outlast — any US chip it is temporarily sold.
Positioning: Who Wins and Why
Western WFE majors (Applied Materials, Lam Research, KLA, ASML, Tokyo Electron) — Counterintuitively the biggest near-term beneficiaries of controls: near-record revenue, expanding margins and best-ever 2025 guidance — driven BY China rush-orders (46–48% of revenue in peak quarters). Risk is long-term Chinese tool replacement, not the rules.
Nvidia (H20/H20E/B30A China lane) — Re-granted H20 licenses = hundreds of thousands of chips and billions in China revenue; a rumored Blackwell B30A would extend the franchise. SemiAnalysis thinks selling in is strategically wrong but commercially lucrative — and that China's demand is real, not fading.
Unitree (pre-IPO) and the Chinese robotics supply base — Cheapest humanoid at ~$9K BoM / 67% margin, ~70% of global quadruped volume, tripling revenue into a ~$7B-rumored Shanghai IPO; ~200 Chinese humanoid firms and a deep gearbox/motor/LiDAR ecosystem (Leaderdrive et al.) ride the same wave.
Huawei / SMIC / CXMT / Cambricon (China self-sufficiency stack) — The national-champion vertical stack — logic (SMIC), memory (CXMT, still un-listed), accelerators (Ascend, Cambricon popular at ByteDance), tools (SiCarrier, $2.8B raised). State-funded ($30B/yr to Huawei; Big Fund III) and impossible to short cleanly, but the structural long China thesis.
Apple / Qualcomm / MediaTek (asymmetric downside, not upside) — Inverse exposure: a Huawei phone resurgence via Kirin 9000S insourcing threatens ~$20B of Apple revenue and ~$7.6B of Qualcomm+MediaTek — the SemiAnalysis 'winner' here is really a risk flag for incumbents.
Key Data
| Metric | Value | Note |
|---|
| US share of world's deployed FLOPs | >70% | The US lead controls are meant to defend; compute is the one gating factor the US can restrict. |
| Huawei Ascend units shipped (2024 → 2025) | 507k → 805k | 653k of the 2025 units are the more advanced 910C; SemiAnalysis calls the reported '200k' figure significantly off. |
| TSMC Ascend dies smuggled into Huawei's 'die bank' | 2.9M+ | Via shell firms (Sophgo/Bitmain); usable for both 910B/910C; expected to run out ~9 months after Sept 2025. |
| Foreign HBM stacks procured by China | ~13M (Samsung 11.4M) | Sufficient for 1.6M Ascend 910C packages; 7M shipped in the 1-month pre-enforcement gap alone. |
| CXMT HBM capacity, 2026 | ~2M stacks | Only enough for 250–300k Ascend 910C; still NOT entity-listed despite $2B Big Fund III injection. |
| Huawei WFE spend, 2024 | $7.3B (+27% YoY) | 4th-largest WFE customer globally, from ~zero in 2022; 2nd-largest if SMIC+CXMT are included. |
| SMIC advanced-node capacity (7nm & below) | 45k → 60k → 80k wspm (2025→27) | Only single-digit % allocation needed to make >1M Ascend die/yr; logic is not the constraint. |
| SMIC N+2 (7nm) yield defect density D0 | ~0.14 | Roughly 2× TSMC N5/N6 (i.e. worse but 'decent'); debunks the '10% yield' narrative. |
| Advanced-logic gap SMIC vs TSMC | ~5 years | SMIC N+2 shipped 2023 vs TSMC N7 in 2018 / N5 in 2020; narrower than most Western estimates. |
| Nvidia B30A FLOPs vs H20 | ~10× | Rumored Blackwell-based China chip at ~half a B300's price/perf — 'buy 2× for parity.' |
| Revenue at risk if Huawei regains phone share | ~$20B Apple, ~$7.6B Qualcomm+MediaTek | Kirin 9000S insourcing threatens ~35–45M iPhone units gained after Huawei's 2019 ban. |
| Unitree G1 humanoid price cut / BoM / margin | $50K+ → $27.3K; ~$8,976 BoM; ~67% GM | Deals heard well under $20K; may ship its 10,000th humanoid vs Western single-digit units. |
| Unitree quadruped entry price decline (2018→now) | 94–96% (Laikago $45K → Go2 ~$1.6–2.8K) | The DJI-style scaling curve that funded the humanoid program; ~70% of global quadruped volume. |
| Unitree quadruped gross margin (per IPO filing) | 42.36% → 55.49% | Scaling gave 'upstream bargaining power'; costs dropped nearly in half. Self-develops motors/gearboxes. |
| Unitree AI R&D spend planned | ~$300M | On revenue tripling YoY; ~250 Unitrees already in labor pilots below the $30/hr human line. |
| China local share of world's largest robot market | ~50% (from 30% in 2020) | ~200 Chinese humanoid firms; a US UR5e-clone arm costs ~2.2× the Chinese one. |
| SemiCap combined China sales, 3Q23 | ~$8.7B (+50% vs prior peak) | ASML 46% / Lam 48% of revenue from China on rush orders ahead of controls. |
| Critical WFE components that gate the whole chain | <100 | e.g. Zeiss EUV projection optics, precision mask stages, RF plasma generators — the tightest control point SemiAnalysis urges. |
Key Theses
1. Export controls are genuinely WORKING at the leading edge — because of HBM, China will make FEWER Ascends next year, not more.
SMIC logic is no longer the bottleneck (needs only single-digit % of capacity for millions of die); the binding constraint is HBM. China has ~13M foreign stacks (good for 1.6M 910C) that run out by end-2025; CXMT can supply only ~2M stacks in 2026 = 250–300k Ascends. Without smuggled HBM, Huawei can't build even 1M chips next year. Had controls not existed, the Ascend ramp would be fully realized and DeepSeek R2/V4 'would already be here.'
“China can easily make more than 805k Huawei Ascends this year from TSMC and SMIC capacity, but they will not because they do not have enough HBM.”— Huawei Ascend Production Ramp: Die Banks, TSMC Continued Production, HBM is The Bottleneck (2025-09-08), SemiAnalysis
2. The same controls are riddled with self-inflicted loopholes — 'fab whack-a-mole' makes the entity list structurally defeatable.
One SMIC building is entity-listed for advanced logic; its twin next door imports 'dual-use' tools for 'legacy' — connected by a wafer bridge / overhead track that makes them one cleanroom. Huawei spins up non-listed shells (Pengjin across the street from listed PXW; Pensun next to SMIC) faster than rules update. Offshore manufacturing (Lam's Malaysia etch chambers, AMAT/KLA Singapore) legally reaches entity-listed fabs if no US persons are involved; CXMT dodged controls by renaming an 18nm node to 19nm.
“In the race between shell entities and regulations, shells can be spun up faster than regulations are being updated.”— Fab Whack-A-Mole: Chinese Companies are Evading U.S. Sanctions (2024-10-28), SemiAnalysis
3. The single biggest enforcement failure was HBM: Samsung shipped 7M stacks to China in the one-month gap before enforcement.
BIS announced HBM2E-plus controls Dec 2, 2024 with a Dec 31 compliance date — telegraphed for months. Samsung alone directly provided 11.4M HBM stacks to China, including 7M in that single month; total foreign HBM into China reached ~13M stacks (majority Samsung), the majority of China's HBM. Because Korea has no re-export controls and Samsung failed to enter the Western accelerator supply chain, it sold to Chinese customers who funneled inventory to Huawei.
“Samsung alone has directly provided 11.4 million stacks of HBM to China, including a staggering 7M stacks in the 1-month gap between controls announcement and enforcement dates.”— Huawei Ascend Production Ramp (2025-09-08), SemiAnalysis
4. WFE suppliers' 'death spiral' lobbying is false — the two years under controls were among their best ever, driven BY China.
Despite lobbying that controls cause a 'death spiral,' AMAT/Lam/KLA saw near-record revenue, expanding gross margins and rising R&D during the control years, with 2025 guided as best-ever; the primary driver was China. AMAT grew faster than WFE for a 5th straight year 'despite headwinds.' SemiAnalysis's read: the cost of controls to Western suppliers is short-term and recoverable, so controls are worth it — the real long-term threat is Chinese domestic tool replacement (the 'China Technology Cycle'), not the rules.
“By nearly every metric the 24 months under export controls have been among the best in history for American WFE suppliers.”— Fab Whack-A-Mole (2024-10-28), SemiAnalysis
5. SMIC's 7nm is real and good — the gap to TSMC is ~5 years, not a decade, and a DUV-only '5nm' is feasible.
The Kirin 9000S on SMIC N+2 matches 1–2-year-old Qualcomm chips; on identical Arm A510 IP it's on par with Samsung 4LPX despite the node gap. Yield is decent (D0 ~0.14, ~2× TSMC N5). SMIC uses the same 1980i DUV tools as TSMC/Intel 7nm; ASML's own capacity plans BUILD IN Chinese semiconductor independence (1.5M excess wafers/month by 2030). A '5nm' at 130M+ transistors/mm² via ArFi multi-patterning would cost only ~20% more than EUV-based, since lithography is ~30% of process cost.
“Even with the lackluster export controls, this is a leading edge chip that would be near the front of the pack in 2021, yet was done with no access to EUV... We cannot overstate how scary this is.”— China AI & Semiconductors Rise: US Sanctions Have Failed (2023-09-12), SemiAnalysis
6. Selling Nvidia into China (H20, and especially a Blackwell B30A) is a strategic mistake dressed as diplomacy.
The H20 has more memory than the banned H100 (good for inference/RL, the current driver of progress); a rumored B30A could carry ~10× H20 FLOPs at half a B300's price — just buy 2× for parity. This gives DeepSeek/Alibaba more RL compute and lets China serve and PROLIFERATE its AI on its own apps, taking share from US products. 'China may be trying to psyop their way into getting the significantly more performant Blackwell chip.' The claim that Blackwell must be sold because Huawei is an alternative is false — Huawei is about to run out of HBM.
“The American AI stack is American AI on American chips, not just the chips.”— Huawei Ascend Production Ramp (2025-09-08), SemiAnalysis
7. Controls must move UPSTREAM: tighter on tools than chips, tightest on the <100 critical WFE components.
Current controls are strict downstream (accelerators) but loose upstream (equipment) and non-existent on components — which incentivizes China to build indigenous tools and eventually renders sanctions toothless. Example: EUV scanners are banned but Zeiss EUV projection optics (the #1 critical component) are not; fewer than 100 parts (EUV optics, precision mask stages, RF plasma generators) gate the whole chain. Margin-stacking math: $1B of restricted tools compounds to 2×+ at foundry and again at fabless, so keeping components onshore is high-leverage.
“The entity list must be expanded and updated often... This means monthly updating, as well as geographic checks. If we can find this with some satellite photos in trivial time, it is clear the rules are easy to skirt.”— Fab Whack-A-Mole (2024-10-28), SemiAnalysis
8. Robotics is the AI front where China leads outright — and the US is 'missing the new labor economy.'
Robots making robots is the first industrial input that is fully additive (24/7, higher throughput than a human), and only China is positioned to capture it. Chinese local firms are approaching ~50% of the world's largest robot market (up from 30% in 2020); Xiaomi runs lights-out factories making one phone per second with zero humans; a US-built UR5e-clone arm costs ~2.2× the Chinese one, and even 'Made in USA' parts depend on China-made components with no scalable alternative. This is the battery/solar/EV capture playbook applied to a general-purpose technology.
“The only country that is positioned to capture this level of automation is currently China, and should China achieve it without the US following suit... posing an existential threat to the US.”— America Is Missing The New Labor Economy - Robotics Part 1 (2025-03-11), SemiAnalysis
9. Unitree is being born as the next BYD/DJI — cheapest humanoid on earth at a ~$9K BoM and 67% gross margin.
Unitree cut G1 pre-tax pricing from $50K+ to $27.3K in 12–18 months yet still hits ~67% gross margin on a ~$8,976 BoM, with deals heard well under $20K; it may ship its 10,000th humanoid while Tesla/Figure ship single digits. It owns the actuator (50–70% of humanoid BoM), self-develops motors (30–40% of Western cost), gearboxes, LiDAR and cameras. Quadruped margins already went from 42.36% to 55.49% per its IPO filing. ~250 Unitrees are in labor pilots today, already passing below the $30/hr human-labor line even under conservative full-teleoperation assumptions.
“Unitree has slashed pre-tax pricing from $50K+ to $27.3K over the past 12-18 months. Even at that price, we estimate they still hit 67% gross margins on their flagship G1.”— China's Unitree Will Dominate Global Robotics (2026-06-08), SemiAnalysis
10. Unitree's QDD bet — mocked for overheating — is now defying expectations and building an ecosystem.
Critics said quasi-direct-drive (big motor, tiny gearbox) would draw too much current and overheat: the original G1 held 2kg outstretched for seconds before a 30-min cooldown. Two years of iteration (smoother magnets to cut cogging/torque ripple, 'low copper consumption' coils, targeted active cooling) pushed it to 5kg for 10–15 min arms-bent (~2× payload, ~5× duration). QDDs are 95–98% efficient (vs 85–90% strainwave), up to 80% cheaper, use standard hobbing so a new sample actuator comes in weeks vs 3+ months for Western firms — and other Chinese humanoids are now switching to QDD, thickening the supply base.
“While most stuck with the strainwave (HarmonicDrive) route, Unitree took the chance and iterated.”— China's Unitree Will Dominate Global Robotics (2026-06-08), SemiAnalysis
11. China's datacenter/power advantage means chips — not grids — are its only real AI ceiling.
Where US frontier clusters are 'primarily capped by regulatory environment and industrial capacity' (forcing unproven cross-datacenter training), China has 'no issues securing power infrastructure' — multi-gigawatt substations convert from aluminum mills to datacenters in <6 months. Chinese AI chips are today decentralized (largest clusters ~1/3 US size), but a concentrated effort 'could lead to cluster sizes that dwarf those in the US in less than a year.' So compute — the one thing the US can restrict — is the decisive lever.
“There are enough chips being shipped to China + manufactured domestically to create the world's largest AI training cluster.”— Fab Whack-A-Mole (2024-10-28), SemiAnalysis
12. China exploits IP/JV loopholes the same way it exploits tool loopholes — Arm China went fully rogue.
SoftBank sold 51% of Arm's China ops to a Chinese consortium for a paltry $775M; the JV then went rogue under ousted-but-entrenched CEO Allen Wu (who held the company chop, hired security paid by Arm China that locked Arm out, and literally sued himself). Arm China ('安谋科技') rebranded its own Power Core / XPU IP (NPU/SPU/ISP/VPU) with a 400+ person China R&D team — 'the tech heist of the century.' Same pattern in RF/power: STMicro's SiC JV with Sanan risks transferring tech repurposable for military RF.
“Arm has been shaken to its core with the 2nd largest market snatched from underneath it. According to Arm, no IP has been stolen, but Arm cannot license directly license to any firm in China.”— The Semiconductor Heist Of The Century | Arm China Has Gone Completely Rogue (2021-08-27), SemiAnalysis
Article Deep-Dives
The definitive quantified model of China's domestic AI-chip capacity. Huawei shipped 507k Ascend in 2024 and 805k in 2025 (653k 910C), carried by a 2.9M-die TSMC 'die bank' running out in ~9 months. SMIC advanced capacity (45k→80k wspm) means logic is NOT the constraint — HBM is. China has ~13M foreign stacks (11.4M from Samsung, 7M in a one-month enforcement gap) good for 1.6M 910C; domestic CXMT can supply only ~2M stacks in 2026 (250–300k Ascends). Verdict: controls WORK — assuming no smuggling, China makes fewer Ascends next year, not more — but Nvidia's H20/B30A and un-listed CXMT are the leaks. Also details the KunPeng CPU and CM384 switches made at TSMC to free SMIC for compute die.
SemiAnalysis's enforcement manifesto and policy blueprint. Four claims: (1) controls work but need constant updates; (2) Huawei's state-backed fab network — $7.3B WFE in 2024, 4th globally — is the key threat, spinning up non-listed shells (Pengjin, Pensun) next to listed fabs faster than rules update; (3) the WFE 'death spiral' lobbying is false (record margins, China-driven); (4) enforcement must move upstream. Concrete asks: a 0% de-minimis threshold on Huawei's network (could cripple Chinese fabs in 6 months), monthly entity-list updates with satellite checks, and restricting <100 critical components (Zeiss EUV optics). Documents wafer-bridge tricks, offshore manufacturing (Lam Malaysia, AMAT Singapore), and CXMT's 18nm→19nm rename dodge.
The bull case on China's robotics champion via the BYD/DJI verticalization lens. Unitree owns the actuator (50–70% of humanoid BoM), self-develops motors (30–40% of Western cost), gearboxes, LiDAR and cameras; a full teardown puts the G1 BoM at ~$8,976 with ~67% gross margin even after cutting price from $50K+ to $27.3K. Its contrarian QDD (quasi-direct-drive) bet — mocked for overheating — now sustains 5kg for 10–15 min (2×/5× improvement) after two years of magnet/coil/cooling iteration. ~250 G1s are in labor pilots, already below the $30/hr human line even under full-teleoperation assumptions. With a $7B-rumored Shanghai IPO, tripling revenue and ~$300M AI R&D, SemiAnalysis argues Unitree keeps a structural first-mover cost advantage even as it scales.
The 9,600-word 'call to action' framing robotics as the next strategic-industry capture. Robots making robots is the first fully-additive industrial input (24/7, above-human throughput); only China is positioned to capture it, and if it does without the US following, that's 'an existential threat.' Evidence: Chinese local firms near ~50% of the world's largest robot market (up from 30% in 2020); Xiaomi lights-out factories make one phone per second with zero humans; KUKA's Guangdong plant has robots building robots; a US UR5e-clone arm costs ~2.2× the Chinese one, and even 'Made in USA' parts depend on China. Uses DJI/GoPro as the template: Shenzhen iteration speed + scale/oversupply crushed every Western drone maker. Grounds the whole robotics thesis the later Unitree pieces build on.
The one-year-post-2022-controls verdict: sanctions had failed at that point. The Kirin 9000S on SMIC N+2 (true 7nm, no EUV) matches 1–2-year-old Qualcomm silicon; on identical Arm A510 IP it's on par with Samsung 4LPX. Yield is decent (D0 ~0.14). SMIC can ramp 7nm to 30k+ wspm on allowed 1980i DUV tools; a DUV-only '5nm' (130M+ transistors/mm²) is feasible at ~20% cost premium. Equipment makers sell 'every tool they offer' to China under the 28nm pretext while it's used for 7nm. China will have >1M A100-class Nvidia chips by end-2024 and multiple firms training beyond GPT-4. Ends with a 20-point 'what can be done' arsenal (lithography, resist, masks, metrology, CMP, EDA, JVs) to shut China out.
The line-by-line breakdown of the Oct-2022 opening salvo that started the chip cold war. Two actions: (a) sweeping export controls on advanced computing + semiconductor manufacturing, and (b) 31 firms to the Unverified List — which, crucially, lets BIS cut ANY Chinese firm off global supply chains in 2 months simply because Beijing blocks end-use audits (a 'nuclear salvo'). The AI-chip test (600 GB/s I/O AND ~4800 TOPS×bit) bans A100/H100/MI250X/Biren BR100; a supercomputer test; a US-persons rule that locks Americans out of Chinese advanced logic/NAND/DRAM; equipment bans at 16nm logic / 128-layer NAND / 18nm DRAM. Flags the enforcement ambiguity ('no such thing as equipment for 16nm but not more') that would define the next three years of loophole reporting.
Reference: Value Chain
WFE / tools & subsystems (the chokepoint)ASML, Applied Materials, Lam, KLA, Tokyo Electron, ASMI, Zeiss, VAT Group; China: Naura, AMEC, SiCarrier — The true bottleneck. SemiAnalysis argues controls should be TIGHTER upstream (tools > chips, components > tools): <100 critical parts (EUV optics from Zeiss, precision mask stages, RF plasma generators) gate the whole chain. Huawei spent $7.3B on WFE in 2024 (4th-largest globally); its SiCarrier copies foreign tools and raised $2.8B.
▼
Advanced-logic foundrySMIC (N+2 7nm), Huawei captive fabs (Pengjin, Pensun, PXW); TSMC (the leak); Cambricon (fabless) — SMIC makes the Ascend/Kirin dies on DUV 7nm; Huawei is building its own fab network of non-entity-listed shells next to listed fabs. SemiAnalysis forecasts SMIC advanced capacity of 45k→60k→80k wspm (2025→27) — enough that logic is NOT the binding constraint; HBM is.
▼
Memory / HBMCXMT (DRAM/HBM), Swaysure (3D DRAM), JCET & TongFu (OSAT/TSV); foreign leaks: Samsung, SK Hynix, Micron — The binding constraint on Chinese AI accelerators. CXMT ships DDR5 ~2 yrs behind, targeting 257k wpm (~15% of world DRAM) in 2026; still NOT entity-listed despite the Big Fund III $2B injection. Foreign HBM (13M stacks, mostly Samsung) is what China actually runs on until it dries up.
▼
AI accelerators & systemsHuawei Ascend 910B/910C + CloudMatrix CM384 + KunPeng CPU, Cambricon, Biren, MetaX; Nvidia H20/H20E/B30A (the sanctioned imports) — Huawei is 'Nvidia's greatest adversary.' 507k Ascend shipped 2024, 805k 2025 (653k are 910C). System > microarchitecture: CM384 scale-up switches & KunPeng 930 CPU are made at TSMC to free SMIC for compute die. Nvidia backfills via H20 (more memory, fewer FLOPs) and a rumored B30A with ~10× H20 FLOPs.
▼
Frontier AI models (compute-gated)DeepSeek, Alibaba/Qwen, Moonshot/Kimi, ByteDance/Seed — China's labs have talent but are throttled by compute. DeepSeek's next-gen model was delayed by Huawei-chip use; most DeepSeek tokens still inference on Western hardware; R1 served deliberately slow to save compute. ByteDance trains Seed in the US on US clouds and rents Blackwell via Malaysia/Oracle/Google.
▼
Robotics / embodied AI (China ahead)Unitree, UBTech, AGIBot, DEEP Robotics, DJI legacy; West: Boston Dynamics, Agility, Figure, Tesla Optimus, ANYbotics — The one AI front where SemiAnalysis says China leads outright. ~200 Chinese humanoid firms; Unitree ~70% of global quadruped volume, shipping its 10,000th humanoid at ~$9K BoM / 67% margin vs Western single-digit units. A US-built UR5e-clone arm costs ~2.2× the Chinese one; even 'Made in USA' parts depend on China.
Reference: Core Concepts
Foreign Direct Product Rule (FDPR) & de minimis threshold. Two distinct legal mechanisms. The Foreign Direct Product Rule (FDPR) is a binary 'direct product' test: it lets the US control any foreign-made item that is the direct product of specified US technology/software (or made by a plant that is itself a direct product of such technology) — with NO percentage threshold. The SEPARATE de minimis rule instead turns on US-origin CONTENT by value: a foreign item is normally exempt below 25% US content, but that cutoff is lowered to 0% for the most sensitive cases (e.g. tools like ASML's 1980i lithography). SemiAnalysis's central policy recommendation is applying a 0% de minimis threshold to Huawei's fab network, which they argue could handicap or halt Chinese advanced fabs within six months.
TPP / Performance-density threshold (the AI-chip test). The 2022 rule banned chips above ~4800 (TOPS × bit-length) with 600 GB/s I/O — a two-part test Nvidia gamed with the A800/H800 by cutting only interconnect. The 2023 rule dropped the bandwidth prong and added a performance-DENSITY threshold (TPP÷die area, 5.92 absolute / 3.2 license), which SemiAnalysis says forces any compliant chip back to a 2017-era V100 and can't be re-gamed with blank silicon.
Die bank. Huawei's stockpile of foreign-fabbed logic dies. Before SMIC could ramp, Huawei procured 2.9M+ Ascend dies from TSMC (via shell firms) usable for both 910B and 910C. This inventory — not domestic production — is what carries Ascend volumes through 2024–25; SemiAnalysis expects it to run out within ~9 months of Sept 2025.
HBM as the binding constraint. High-Bandwidth Memory (stacked DRAM with TSVs) is the real chokepoint for Chinese AI ASICs — modern LLM training/inference needs it, GDDR/LPDDR won't substitute. China procured ~13M foreign HBM stacks (enough for 1.6M Ascend 910C packages) then hit a wall; domestic champion CXMT can make only ~2M stacks in 2026 (~250–300k Ascends). Beijing's trade-deal ask targets HBM, not lithography — telling you where it hurts.
Allied-lag / re-export window. The US exempts Japanese and Dutch tool makers, who then don't immediately match new controls (often delayed 6 months or never), and neither restricts re-export. Chinese firms rush-order years of equipment into this gap; Samsung shipped ~7M HBM stacks to China in the one-month window between the Dec-2024 control announcement and its Dec-31 enforcement.
SMIC N+2 (7nm) via DUV multi-patterning. SMIC produces true 7nm-class logic (Kirin 9000S, Ascend) with no EUV, using ASML's older 1980i DUV immersion scanners plus multi-patterning. SemiAnalysis argues yield is actually decent (D0 ~0.14, roughly 2× TSMC N5), the gap to TSMC is ~5 years not a decade, and even a DUV-based '5nm' at ~130M transistors/mm² is feasible given subsidies (lithography is only ~30% of process cost).
QDD (Quasi-Direct-Drive) actuator. Unitree's core bet: a big motor + tiny (<20:1) planetary gearbox, vs the industry-standard small motor + big strainwave (Harmonic Drive) gearbox. QDDs are 95–98% efficient (vs 85–90%), up to 80% cheaper, built with standard gear-hobbing so many suppliers exist, and iterate in weeks not months. Early ones overheated; after two years Unitree pushed payload/duration ~2×/5× — making the cheap architecture viable.
The DJI/BYD verticalization playbook. SemiAnalysis's frame for how Chinese hardware giants are born: own the single hardest/most-expensive BoM component (BYD=battery cell, DJI=flight controller, Unitree=actuator), bootstrap on a willing hobbyist/researcher market with a half-finished cheap product, ride the Shenzhen ecosystem, and let each generation unlock a new market while in-housing more of the stack until the cost structure is untouchable.
Military-civil fusion & the entity/unverified list. Because China's national-security law forbids US end-use checks and blurs civilian/military R&D, BIS uses two lists: the Unverified List (a 2-month on-ramp that can cut any firm off simply because Beijing blocks audits) and the Entity List (Huawei, YMTC). SemiAnalysis argues the whack-a-mole entity approach is structurally defeated by shell-company spin-ups faster than rules update.
Open Questions
- Will CXMT close the HBM gap fast enough to matter? SemiAnalysis pegs 2026 at only ~2M stacks (250–300k Ascends), but flags that Chinese manufacturing 'knows no bounds' and HBM3e in 2026 is plausible — the whole China-compute forecast hinges on this ramp.
- Does the US have the political will to move upstream (0% threshold, <100 components, monthly entity updates), given ally friction (ASML/Netherlands), WFE lobbying, and rare-earth leverage China now holds?
- Is China's H20 'disinterest' real demand-signal or brinkmanship? SemiAnalysis believes it's top-down government pressure to extract a more powerful B30A — but if wrong, the compute-diplomacy calculus changes.
- Can Unitree hold its structural cost lead through mass production, or will UBTech/AGIBot's verticalization drive (and Western supply-chain reliance on China) compress the moat once volumes scale?
- Does forking into a different silicon idea-space (compute-in-memory, neuromorphic) let China escape the transformer/GPU control regime entirely, or is it a dead end that just buys time?
Sources (SemiAnalysis)