WireScreen Briefing: GPU & AI Compute Intelligence
Market Concentration, Export Controls, State-Linked Chipmakers, Entity List Risk  ·  June 2, 2026
National Security — The Compute Frontier
Strategic Review — GPUs & The Compute Chokepoint  ·  June 2, 2026
The graphics processing unit has quietly become the most strategically contested object on earth. Built to render video-game worlds, it turned out to be the perfect engine for AI — thousands of parallel cores doing exactly the math that trains every frontier model. Of AI's three inputs — data, algorithms, and compute — only compute must be physically manufactured, in volume, at the bleeding edge of the possible. You cannot download it, and almost no one can make it.

The market is a near-monopoly. Nvidia holds roughly 85% of AI data-center accelerators and about 92% of discrete GPUs; AMD is a distant second, Intel a vanishing third. But the deeper story is the chain behind the designer: ASML of the Netherlands alone builds the EUV machines that print leading-edge logic; TSMC of Taiwan alone fabricates and packages at scale; Korean memory makers SK Hynix and Samsung supply the rest. Control is split across five jurisdictions, complementary and not substitutable — stall any one stage and the whole AI build-out stalls with it.

This is the mirror image of the drone supply chain. There, China holds the chokepoint; here it holds none of the critical layers, and is locked out of the best chips. The US commands an estimated tenfold lead in total AI compute, and of 22 notable models developed exclusively in China by 2025, only two were trained on Chinese silicon — the rest ran on Nvidia chips, imported, stockpiled, or smuggled.

That tension has reshaped US policy. Since October 2022 Washington built the most aggressive export regime of the modern era around these chips — and in 2026 inverted it: a "revenue-for-access" deal letting leading chips reach China only if the US Treasury takes a 25% cut. Critics call it taxation by proclamation, challenging its legality under both the Export Control Reform Act and federal user-fee statutes such as the Independent Offices Appropriations Act. Beijing, locked out, responds on three fronts — building its own through Huawei and SMIC, smuggling foreign chips through shell networks, and pouring state capital into an industry whose ownership is deliberately hard to read. That last front is where WireScreen's lens matters most.
Nvidia AI-Accelerator Share
~85%
of AI data-center accelerators; ~92% of discrete GPUs (H1 2025)
US Compute Lead vs. China
~10×
estimated US advantage in total AI compute capacity (RAND)
AI Models Built on Chinese Chips
2 of 22
of 22 notable AI models built entirely in China by 2025, only 2 were trained on Chinese-made chips — the other 20 ran on Nvidia silicon
China's AI Chips: Home-Made Share
41%
of AI-server chips shipped in China in 2025 were domestically made (Huawei, Cambricon, etc.); the other ~59% were foreign — mostly Nvidia, imported or smuggled
Nvidia Data-Center Revenue, Q1 FY27
$75B
Data Center segment, quarter ended Apr 2026 — Nvidia reports ~$60B compute + ~$15B networking; total company revenue $81.6B
Market Structure & Supply-Chain Landscape
Principal Makers & Supply-Chain Controllers
Company
HQ / Role
Risk Flag
2025–26 Signal
— US & Allied Designers and Supply-Chain Controllers (No PRC State Link)
Nvidia
Santa Clara, USA
Designs GPUs/accelerators; CUDA software moat
US
Export-control subject; 15–25% revenue-for-access charge
~85% accelerators / ~92% discrete; Hopper→Blackwell→Rubin
AMD
Santa Clara, USA
Instinct MI300/MI325 accelerators; ROCm software
US
Export-control subject; MI308 China-spec
~6% accelerator share; distant but credible #2
TSMC
Hsinchu, Taiwan
Leading-edge foundry + CoWoS advanced packaging
Taiwan
Geographic concentration risk; $1B Sophgo fine 2025
~90% of leading-edge AI silicon; the fabrication chokepoint
ASML
Veldhoven, Netherlands
Sole maker of EUV lithography systems (~$400M each)
Netherlands
Allied export controls; EUV barred to China
Sole-source chokepoint; High-NA EUV for sub-2nm nodes
SK Hynix
Icheon, South Korea
High-bandwidth memory (HBM) — majority of Nvidia supply
S. Korea
HBM export-controlled to China since 2024
HBM supply chronically lags demand; Samsung, Micron trail
— Chinese / PRC-Linked Chipmakers (State Exposure)
Huawei / HiSilicon
Shenzhen, China
Ascend 910C accelerators; CANN software; fabbed by SMIC
China
100% Huawei (Trade-Union Cmte 99.5%) · Entity List · State Security Contractor · Section 889 · supplied Russia '22–23
Mandated domestic Nvidia alternative; ~$12B AI-chip revenue projected 2026 (from $7.5B)
Cambricon
Beijing, China
MLU training/inference accelerators; STAR Market listed
China
Entity List (2022) · Chinese Academy of Sciences owns 15.6% · Gov. Guidance Funds
CAS spinout (founders ex-CAS); supplies Huawei; Chen Tianshi 28.9%
Biren Technology
Shanghai, China
BR100-series GPGPU; HK IPO 2025
China
US Entity List (2023) · state-fund cap table
"Private" startup; ownership resolves to state vehicles
Moore Threads
Beijing, China
MTT GPUs (graphics + AI); STAR Market IPO 2025
China
Entity List (2023) · ByteDance & Tencent backed · China Mobile fund + SASAC of State Council on register
Founder a former Nvidia China exec; core GPU IP traces to Imagination (UK)
Hygon (Sugon)
Tianjin / Beijing
x86 server CPUs (AMD-licensed JV) + DCU AI accelerators
China
Entity List (2019, w/ Sugon) · ~23% government-owned · Military-Industrial Complex Co.
Parent Sugon/Dawning (28%); Chengdu SASAC, Sichuan & Tianjin finance on register
Entity Spotlight
From the Platform
Huawei Technologies Co., Ltd. (HiSilicon · private)
华为技术有限公司  ·  Shenzhen, Guangdong  ·  Private (employee-owned)  ·  Est. 1987  ·  US Entity List (2019)  ·  HiSilicon design arm
The WireScreen profile of Huawei surfaces what a product name hides. Barred from US technology since its 2019 Entity List designation, Huawei could not legally order leading-edge dies from TSMC. SemiAnalysis estimates it obtained them anyway through Sophgo, a Cayman-registered front — reportedly on the order of $500M of 7nm wafers, enough to yield well over a million Ascend processors (the 910B and 910C) across 2024–25. Reuters reports TSMC could face a fine of $1 billion or more over the diverted dies. Either way, that silicon feeds the Ascend program — now the backbone of China's AI clusters and the state-mandated alternative to Nvidia.

HBM memory is the other workaround. SemiAnalysis traces banned Samsung HBM into China via distributor CoAsia and packager Faraday/SPIL — bonded to a cheap logic die with weak solder so buyers can desolder and recover it — reportedly letting Huawei stockpile millions of HBM stacks. Its CloudMatrix 384, in SemiAnalysis's framing, is "a generation behind in chips but a generation ahead in scale-up." Even so, a CFR assessment puts Huawei's total AI compute near 2% of Nvidia's over 2025–2027.

The takeaway: the chip war is fought through corporate structure — offshore shells, relabelled resellers, and state funds layered three deep. "Private" champions like Biren and Moore Threads resolve upward into the Big Fund and provincial vehicles. Read name-by-name, the chain is invisible; resolved as an ownership graph, the shell, the sanctioned parent, and the state fund become plain.
Source: WireScreen platform profile, May 2026; SemiAnalysis & TechInsights  ·  WireScreen tracks every company in this briefing through a 20M+ Chinese entity dataset.
Leading-Edge AI-Chip Fabrication — 2025 (Where Advanced Silicon Is Made)
Taiwan (TSMC)  ~90%
South Korea  ~6%
United States  ~3%
China (SMIC, trailing node)  ~1%
Concentration by Layer of the GPU Stack — 2025
Nvidia discrete GPU share
~92%
ASML EUV machines (sole)
100%
TSMC leading-edge fab
~90%
Nvidia AI accelerators
~85%
Top-4 CoWoS packaging
~85%
China domestic AI-server
~41%
China leading-edge fab
<2%
Export-Control & Regulatory Timeline
May 2019
Huawei Added to US Entity List
Commerce bars Huawei from buying US-origin technology — the foundational restriction that later forces its Ascend chips onto domestic foundry SMIC and into grey-market workarounds.
Aug 2022
US CHIPS & Science Act Signed
~$52B in subsidies to re-shore US semiconductor manufacturing — the carrot beside the export-control stick.
Oct 2022
First BIS Controls — A100/H100 Banned
Commerce bars export of top AI chips and key manufacturing tools to China — the opening move of the modern compute-control regime.
Sep 2023
Huawei Mate 60 Pro — SMIC 7nm Shock
Huawei ships a phone with a SMIC-made 7nm chip, signalling that China can produce advanced silicon without EUV — and triggering a sharp tightening of US policy.
Oct 2023
The A800/H800 Loophole Closes
Rules tighten to catch the cut-down parts Nvidia designed for the China market after the 2022 ban.
May 2024
China's "Big Fund" Phase III Launches
Beijing commits ~$47.5B in fresh state capital to semiconductors — the financing engine behind Huawei, SMIC, and the domestic GPU startups.
Dec 2024
HBM & Tooling Enter the Net
Controls extend to high-bandwidth memory and additional semiconductor manufacturing equipment.
Jan 2025
DeepSeek R1 — The Efficiency Shock
A Chinese lab trains a frontier-class reasoning model cheaply on constrained hardware. Nvidia falls ~17% in a day (~$589B wiped) — proof that compute limits can be partly engineered around.
Jan 2025
The "AI Diffusion" Three-Tier Framework
Licensing caps GPU exports across 120+ countries; ~20 states (China, Russia, Iran, NK) prohibited outright.
Apr 2025
Even the China Chip Is Pulled — H20 Licensed
The H20, Nvidia's China-only part, now requires a license. The last legal high-end option closes; Beijing later discourages purchases.
Aug 2025
Revenue-for-Access Begins (15%)
H20 licenses granted on condition Nvidia remits 15% of China sales to the US government — a novel and contested structure.
2025
Sophgo Exposed; TSMC Faces $1B+ Fine; "Operation Gatekeeper"
An estimated ~2.9M TSMC dies (SemiAnalysis) are found to have reached Huawei via a Cayman-registered front; Reuters reports TSMC could face a $1B-plus fine. BIS launches an enforcement push against grey-market routing through Singapore, Malaysia, and the UAE.
Sep 2025
China Tells Firms to Stop Buying Nvidia
Regulators discourage purchases of approved Nvidia parts and escalate an antitrust (SAMR) probe — turning the controls into Beijing's own bargaining chip and a captive market for Huawei.
Apr 2026
DeepSeek V4 Ships — Integrated with Huawei Chips
China's flagship open model arrives partly running on Huawei silicon, signalling a deliberate pivot away from reliance on Nvidia hardware.
May 2026
25% Revenue-for-Access on H200-Class Sales
Reports describe a 25% charge on H200-class chips sold to China, with licensing handled case-by-case. Critics dispute the government's authority to levy it under both export-control law and federal user-fee statutes.
Quick Glossary
GPU vs. AI accelerator. A graphics processing unit runs thousands of simple cores in parallel — the math AI training and inference depend on. "AI accelerator" refers to data-center GPUs (and related chips) sold for that workload; Nvidia dominates both the consumer and data-center categories.
EUV (Extreme-Ultraviolet Lithography). The technique required to print leading-edge logic. ASML of the Netherlands is the only company that builds EUV machines (~$400M each); China is barred from buying them, capping its fabs at 7nm-class nodes.
HBM (High-Bandwidth Memory). Stacked memory that feeds a GPU its data, made chiefly by SK Hynix and Samsung. A critical input China cannot yet produce at scale — and restricted to China under US controls since 2024.
CoWoS. TSMC's chip-on-wafer-on-substrate packaging, which fuses logic die and HBM onto one interposer. The current binding chokepoint on global AI-chip output; the top four designers take ~85% of capacity.
Entity List. US Commerce Department list restricting export of US-origin technology to named firms. Huawei and Hygon/Sugon (2019), Cambricon (2022), and Biren and Moore Threads (2023) are all listed — barring them from US tools and chips.
Revenue-for-access (H20 / H200). The 2025–26 policy under which Nvidia may sell certain chips to China only by remitting 15% (H20) and then 25% (H200) of sales to the US government. Its legal basis is contested under both the Export Control Reform Act and federal user-fee statutes (the Independent Offices Appropriations Act, 31 U.S.C. § 9701).
Ascend / SMIC. Huawei's domestic accelerator line (910B/910C), fabricated by SMIC on a 7nm-class DUV process without EUV. The state-mandated domestic alternative to Nvidia inside China.
Big Fund. China's National Integrated Circuit Industry Investment Fund, now in its third phase with $47B+ committed — the state-capital engine behind domestic chipmakers, often co-investing with provincial SASAC vehicles.
Strategic Implications & Outlook
1
The chokepoint is allied, not American alone
US control rests on partners it does not own — ASML in the Netherlands, TSMC in Taiwan, SK Hynix in Korea. The regime is only as durable as that coalition, and Taiwan's concentration of leading-edge fabrication and CoWoS packaging is its single largest point of fragility.
2
Export control is mutating into trade policy
The 2026 "revenue-for-access" charge blurs the line between security and revenue. Critics argue a percentage-of-sales fee exceeds the government's statutory authority — under export-control law and federal user-fee statutes alike; the dispute is unresolved, and the policy whipsaws with each administration shift.
3
China's domestic surge is real but capped
Domestic chips meet ~41% of China's AI-server units — but that is a large slice of a small, lower-end pie. The volume is real; the high-end capability is not. SemiAnalysis and a CFR assessment estimate Huawei's leading-edge output at only ~200–300K chips a year and China's total AI compute near 2% of Nvidia's over 2025–2027. Without EUV the fabs stay at 7nm-class and HBM is the missing input. Capability is rising fast; leading-edge parity is not imminent.
4
The real risk hides in ownership
Smuggling fronts (Sophgo) and state capital obscure who controls the chips, and "private" labels conceal it. WireScreen resolves the rest: the Chinese Academy of Sciences sits behind both Cambricon (15.6% direct) and Hygon; Hygon is ~23% government-owned; Moore Threads' "venture" register includes a China Mobile fund and the State Council's SASAC; and HiSilicon's records flag supply into Russia's war. Entity by entity, flag by flag, that resolution is precisely WireScreen's edge for compliance, procurement, and investment diligence.
Report generated June 2, 2026  ·  Sources: SemiAnalysis ("Huawei AI CloudMatrix 384"; Accelerator / Datacenter / Wafer-Fab models), TechInsights (Ascend teardowns), Council on Foreign Relations (Huawei capacity assessment, Dec 2025), Congressional Research Service "U.S. Export Controls and China: Advanced Semiconductors" (2025–26), CSIS, RAND, CNAS, U.S. Bureau of Industry and Security rules & 2025–26 revenue-for-access measures, Nvidia & SMIC filings, Morgan Stanley, J.P. Morgan, Reuters, Financial Times, Bloomberg, Baker McKenzie, Morgan Lewis & Lawfare (legal), WireScreen entity profiles & corporate-graph data (20M+ Chinese entities; SAMR, MOFCOM, exchange filings, SEC)
Intelligence Alerts
Get briefings when they publish
New reports delivered to your inbox.
We use cookies to ensure that we give you the best experience on our website. If you continue to use this site we will assume that you are happy with it. View our Privacy Policy for more information.