Every serious observer of Chinese semiconductors eventually runs into the same cluster of company names: SiCarrier, SwaySure, PXW, PST, SiEn, a Shenzhen investment vehicle called SZMII, and a lithography startup named Yuliangsheng. None of them admits a connection to Huawei. Huawei explicitly denies affiliation with all of them. And yet the US government sanctions them as Huawei's manufacturing network, industry insiders treat them as Huawei operations, and locals in Guanlan reportedly refer to the fabs as Huawei's. This report maps what is known, flags what is merely reported, and explains why this parallel fab system, not the commercial foundry industry covered in our companion primer, is the part of Chinese semiconductor manufacturing that Washington actually fears. A disclosure ground rule up front: almost nothing in this piece comes from audited filings. The sourcing is investigative journalism (Financial Times, Bloomberg, The Elec), US government entity listings, satellite analysis, and industry channel talk. We flag confidence levels throughout, and subscribers should treat every capacity and yield figure as an estimate with wide error bars.

Why the network exists

The logic is straightforward once stated. Huawei is the most heavily sanctioned company in the industry: cut off from TSMC since 2020, barred from American technology, and exposed in 2024 for having accessed TSMC capacity through an intermediary (the Sophgo affair, which cost TSMC a reported settlement and tightened everyone's KYC). Huawei's HiSilicon designs the only Chinese AI accelerator line with hyperscaler-grade traction, the Ascend series, plus Kirin smartphone SoCs, automotive silicon, and networking chips. The only legal manufacturing home for those designs is SMIC's sanctioned, DUV-limited advanced capacity, which is scarce, expensive, and shared. A company in that position, with roughly $25 billion in annual R&D, deep-pocketed municipal allies, and explicit state backing, does the obvious thing: it builds its own manufacturing system, and it does so through structures that preserve deniability, complicate sanctions enforcement, and let nominally independent entities buy what Huawei cannot.

The financing and control architecture centers on Shenzhen Major Industry Investment Group (SZMII), a state-owned vehicle under Shenzhen's SASAC that Bloomberg reported in 2023 was established substantially to support Huawei. SZMII wholly owns or controls the key manufacturing entities and has funded a dozen supply-chain firms, with personnel flows (former Huawei executives and engineers throughout) doing the coordination that ownership charts do not show. Huawei's formal position, stated to the FT, is that SiCarrier, SwaySure, UEA, PXW, and PST are not affiliated with it. Washington's position, expressed through the December 2024 Entity List action that added roughly 140 Chinese entities, is that SwaySure, SiEn, Pengxinxu, SiCarrier, and related firms are part of Huawei's chip production effort. Both positions are matters of record; readers can weight them as they see fit.

How we got here

The network's timeline tracks the sanctions ratchet almost step for step. Huawei's 2019 Entity Listing cut off American components and Google services; the May 2020 foreign direct product rule extension severed TSMC, which had manufactured every HiSilicon flagship, and by late 2020 HiSilicon's smartphone chip business was widely written off. The rebuilding phase ran quietly through 2021-2022: SZMII's formation and initial investments, ground broken on the Guanlan sites (construction is believed to have begun in 2022), SiCarrier's incubation, and the great tool-buying campaign that placed billions of dollars of foreign equipment into entities that did not carry Huawei's name. The August 2023 Mate 60 Pro moment, when TechInsights' teardown revealed a SMIC 7nm Kirin 9000s inside a phone launched during the US Commerce Secretary's Beijing visit, was the network's public debut by proxy: it proved the SMIC route worked and converted the whole effort from rumored to undeniable. The 2024 Sophgo affair, in which Ascend dies were found to have been fabbed at TSMC through an intermediary, simultaneously demonstrated how badly Huawei still wanted leading-edge wafers and closed the last gray channel for getting them, making the domestic buildout existential rather than supplementary. Washington's December 2024 Entity List action then formally named the network, and the 2025-2026 phase has been about execution: SEMICON China 2025 as SiCarrier's coming-out, the Ascend roadmap unveiled at Huawei Connect 2025 as a public multi-year commitment, and the first reports of all-domestic tool lines entering operation.

One more piece of context matters for interpreting everything that follows: the economics are not commercial. At a 20-40% yield, each good Ascend die carries a wafer cost multiple of what TSMC-fabbed competitors pay, before counting the capital burned on parallel fabs and duplicated toolchains. No profit-seeking firm would run this program. But the customer set (Chinese hyperscalers under government pressure to buy domestic, state data centers, and Huawei's own cloud) does not clear at market prices, and the financing (SZMII, Big Fund III, municipal land and utilities) does not seek market returns. The correct frame is industrial mobilization, and the correct performance metric is output growth and capability milestones, not margin.

The map

Shenzhen, Guanlan cluster. The FT's satellite-and-sources reporting describes three large sites under construction or ramping in Guanlan, northern Shenzhen. One is attributed to Huawei directly, intended for 7nm-class production of Kirin smartphone processors, Ascend AI chips, and autonomous-driving silicon. A second is operated by SwaySure (昇维旭), the DRAM venture whose founding technical leadership included the late Yukio Sakamoto of Elpida fame; The Elec's satellite analysis suggests SwaySure runs three fabs, one reportedly equipped entirely with domestic tools from SiCarrier, which if accurate would make it the industry's first all-domestic-toolchain 12-inch line. The third is attributed to SiCarrier (新凯来) itself, the equipment company spun out of a Huawei lab that debuted a full tool portfolio at SEMICON China 2025 and has raised $2.8 billion; its role spans building tools, replicating foreign ones (Huawei is reported to have purchased over $9 billion of foreign equipment partly for reverse-engineering reference), and operating fab space. Nearby sit PXW (鹏芯微), founded by a former Huawei executive and described in recent reporting as the network's largest node and a key R&D and pilot-production base, and PST (Pensun, formerly Pengxinshu), focused on mobile chip production. A sixth Shenzhen entity, Pengjin High-Tech, appears in the SZMII portfolio.

Beyond Shenzhen. SiEn in Qingdao (founded by SMIC founder Richard Chang, overseen by Qingdao SASAC, five fabs per industry reporting) was entity-listed in the same December 2024 action for Huawei-linked work; it also runs a legitimate commercial CIDM foundry business, which is why it appears in both this report and our foundry primer. DGGMT / Dongguan Guangmao is reported to target sub-10nm-class production in Dongguan, adjacent to Huawei's enormous Songshan Lake campus and a three-phase R&D complex whose final phase alone is estimated above $4 billion. Reporting also links the network to investments in Shanghai, Ningbo, and Qingdao, and, more loosely, to Fujian Jinhua (JHICC), the DRAM maker sanctioned since 2018, as an active fabrication participant. Counting generously, industry compilations credit the network with roughly eleven fabs across nine sites, mostly 12-inch, with perhaps five sites aspiring to sub-7nm-class capability. We would emphasize the word aspiring: announced cleanroom square footage is not qualified process capability, and the gap between the two is measured in years even for experienced operators.

The toolchain satellites. Around the fabs sits a supply ring: Yuliangsheng (宇量昇), led by former Huawei engineers and developing a domestic immersion DUV scanner that SMIC is now reported to be trialing; Cornerstone in photoresist; ZTOP in optical inspection; YourSend in process tools; plus SiCarrier's expanding subsidiary structure, including EDA unit Qiyunfang. This ring is the connective tissue between this report and our semicap and EDA primers: the same entities appear on both maps because the shadow network is, among other things, the launch customer and integration testbed for the domestic toolchain's most ambitious efforts, including the reported EUV research program (an LDP-source prototype reportedly under test, with commercial ambitions around 2028) that we treat with heavy skepticism until physical evidence emerges.

What it produces, and the SMIC connection

For all the construction, the Ascend chips actually shipping today are made principally by SMIC on its N+2 7nm-class process, using ASML immersion DUV with multipatterning. The publicly triangulated numbers: SMIC's advanced-node (7nm-class and below) capacity is estimated around 45,000 wafer starts per month at end-2025, heading toward 60,000 in 2026 and 80,000 in 2027, with Huawei reported to have procured the majority of it. Yields are the perpetual argument: TD Cowen estimated Kirin-class 7nm yields improved to 60-70% by 2025 from under 40% at launch, while estimates for the larger Ascend 910C die run at 20-40%, and SMIC's 5nm-class N+3 effort (achieved by even more aggressive multipatterning) is reported in pilot with yields cited around or below 20%, though TechPowerUp reported volume production status in 2026. Bloomberg reported Huawei's 2026 plan at roughly 600,000 Ascend 910C units, double 2025, within a total of up to 1.6 million Ascend-line dies; a February 2026 report described a national target of quintupling 7nm/5nm output within two years, and Hua Hong's Huali arm is now readying its own 7nm-class line at Fab 6 with Huawei's direct technical involvement, giving China a second advanced-logic operator with several thousand wafers per month targeted by end-2026. Every one of those figures is a report, not a filing; directionally they all point the same way, which is toward a serious scaling effort constrained primarily by lithography productivity and yield economics rather than by capital or will.

Where does the shadow network fit? Three roles are visible. First, incremental capacity: some reporting suggests the network's combined production could eventually exceed SMIC's, and SMIC itself has seconded engineering teams to help stand the facilities up (as has SMEE, even after losing staff to Huawei poaching), which implies the fabs are real and progressing, not Potemkin cleanrooms. Second, vertical integration beyond logic: SwaySure covers DRAM (with Huawei's self-developed HiZQ HBM featuring on the Ascend 950 roadmap, and CXMT covering the volume HBM supply covered in our memory work), while advanced packaging capacity in Dongguan and partner OSATs closes the loop from wafer to accelerator module. Third, sanctions insulation: an all-domestic-tool line at SwaySure, if it works, is the existence proof that matters more than any single node claim, because it is the configuration that no future BIS rule can touch.

Assessment

What is well established: the entities exist, are massively funded through SZMII and adjacent state channels, are staffed heavily with Huawei alumni, are sanctioned by the US as Huawei's network, and are physically building at scale visible from orbit. What is reported but unverified: essentially every specific capability claim, including node targets, tool self-sufficiency percentages, yield figures, and timelines. What is genuinely unknown: whether an organization can compress twenty years of fab operational learning into five, even with unlimited money and conscripted expertise from SMIC and SMEE. The skeptics' case, voiced even by Chinese chip investors, is that rivals with decades of experience have failed to match ASML and TSMC; the believers' case is that Huawei has repeatedly done what the skeptics said it could not, most recently shipping a 7nm smartphone chip under full sanctions. Our own view sits in the middle: the network is best understood not as a bid to catch TSMC but as a bid to make China's AI compute supply un-embargoable at a two-generation lag, and on that definition of success it is visibly on track, with the Ascend roadmap (950PR in Q1 2026, 950DT in Q4 2026, 960 in 2027, 970 in 2028) as the public commitment against which to measure it.

The implications radiate outward from the network into every layer of the ecosystem, which is why this report sits alongside the sector primers rather than apart from them. For the commercial foundries, the network is simultaneously a relief valve (Ascend demand that might otherwise consume SMIC capacity) and a talent drain (SMEE and SMIC losing engineers to better-paying Huawei-adjacent employers). For the equipment names, it is the most demanding customer in the country and the only one funding full-portfolio domestic toolchains at scale; NAURA and AMEC sell into it even as SiCarrier grows into their competitor. For the memory complex, SwaySure represents a second DRAM aspirant behind CXMT, with the open question being whether China's memory strategy consolidates or fragments. For the AI chip designers outside Huawei (Cambricon, Biren, MetaX and peers), the network is double-edged: it expands total domestic advanced capacity, but Huawei's priority claim on SMIC wafers is precisely why rumors of tripled Cambricon output run into SMIC capacity math, and Hua Hong's 7nm entry (with Biren already taping out there) matters most to this constituency. And for Nvidia and Washington, the network defines the policy dilemma of 2026: the January loosening of H200 sales to China was justified partly as slowing the domestic ecosystem's momentum, while every subsequent Ascend milestone is cited as evidence the controls failed. Both things can be true, and the network's growth rate is the variable that decides which reading ages better.

What to watch, in rough order of signal value: teardown evidence of any chip fabbed outside SMIC (the first PXW- or Guanlan-origin die identified by TechInsights or peers would be the single most important datapoint); SwaySure DRAM appearing in commercial devices; Yuliangsheng scanner trials at SMIC progressing to production insertion; further Entity List actions and, in the other direction, any BIS move against the secondary-market DUV fleet and its servicing, which the April 2026 AEI report identified as the binding loophole; Ascend roadmap execution against the published dates; and the HBM supply question, since memory, not logic, has been the bottleneck on Ascend volumes. For subscribers doing company-level work, the practical note is that none of these entities is investable directly, but the network's progress reprices half the database: every SiCarrier tool win is a NAURA and AMEC comp, every SwaySure ramp is a CXMT signal, and every Ascend volume increment flows through the optical, power, packaging, and analog names covered across this primer series.