In January 2025, NVIDIA served up a China-exclusive graphics card: the RTX 5090D. The “D” suffix was inherited from the previous-generation 4090D—a flagship architecture of the same generation, trimmed down just enough to sit right below the red line of U.S. export controls, designed exclusively for China. By April 2025, the U.S. Department of Commerce updated its control rules, pushing the 5090D across the line. NVIDIA promptly stopped shipments, and the first-version 5090D made its exit. In August of the same year, NVIDIA introduced the second version, the 5090D V2. This time, the spec cuts were even more aggressive: VRAM was slashed from 32GB to 24GB, memory bus width was narrowed, and AI compute performance was hard-locked at the firmware layer. On paper, the card was flawless and fully compliant. It survived for an extra nine months. Then in May 2026, China Customs notified major add-in card manufacturers and logistics customs brokers that this card would not be approved, no import permits would be issued, and retail channels would be entirely unable to clear customs. The authorities never released a single official notice, cited no legal basis, and provided no explanation. The second version of the 5090D was dead too.
One GPU model, two product revisions, killed by two different nations. This was no coincidence. It points to a structural reality in this chip rivalry that most coverage overlooks: the force blocking NVIDIA consists of two state machines, built on entirely opposing design philosophies.
Laying out the blockade landscape as of August 2026, the division of labor between the two sides becomes instantly clear. The United States blocks the apex of the compute pyramid: all flagship Blackwell-architecture chips (B200, GB200, GB300), the previous-generation flagship H100, the original consumer flagship RTX 5090, and the first-version 5090D halted in April 2025 due to new rules. China, on the other hand, blocks the China-tailored and enterprise lines: domestic tech giants effectively halted H20 purchases starting in September 2025, the Cyberspace Administration of China (CAC) issued a ban on the RTX Pro 6000D in September 2025, and China Customs refused to approve the second-version RTX 5090D in May 2026.
Stuck in a stalemate between the two state machines is the H200. The U.S. formally approved export licenses for China in January 2026; yet within hours of the U.S. rules taking effect, China Customs notified customs brokers that shipments were barred from entry. Procurements were subsequently routed into a case-by-case approval process managed by the National Development and Reform Commission (NDRC). According to the FT, over more than half a year, the actual volume delivered and cleared domestically reached only about 20,000 units, whereas the backlog of U.S.-approved export licenses exceeded 400,000 units. Currently in the Chinese market, the only high-end NVIDIA product that can still be legally purchased through official channels is a single card: the RTX 5080.
A comment by Jensen Huang during a Center for Strategic and International Studies (CSIS) symposium in December 2025 described this predicament with utmost clarity: “NVIDIA can no longer get into China, not to mention China has also banned NVIDIA from entering China. I think we are the first company in history to be banned from both sides simultaneously”.
The defining characteristic of the American machine is that all its actions must be codified in public text. Control provisions are published in the Federal Register, where anyone can download the full PDF; the Entity List is publicly searchable and regularly updated; and tariff rates are explicitly spelled out in presidential proclamations. It is a textbook game of cat-and-mouse played in plain view.
When the first round of controls was rolled out in October 2022, the Department of Commerce set two hard physical metrics: a ceiling on compute performance and interconnect bandwidth. The A100 and H100 exceeded the caps overnight. NVIDIA responded swiftly by cutting bidirectional interconnect bandwidth and introducing the China-tailored A800 and H800, successfully passing compliance checks.
In October 2023, the U.S. updated its regulations, eliminating the easily bypassed interconnect bandwidth metric altogether in favor of a composite coordinate system combining Total Processing Performance and performance density. The A800 and H800 immediately crossed the new threshold and were banned. NVIDIA downgraded specifications once again, engineering the H20—a chip with lower single-card compute performance but retaining large memory capacity and interconnect features.
In April 2025, rules tightened for the third time, pulling the H20 into the license-required regulatory scope, while cutting short the life of the newly released first-generation 5090D. In August of that year, NVIDIA unveiled its third-generation downgraded product, the 5090D V2, going so far as to completely lock down AI compute capabilities at the low-level firmware.
There are no grey secrets in this logic: every time NVIDIA bypassed a restriction, it did so by pinpointing loopholes in public statutes; every time the Department of Commerce plugged a hole, it had to codify the new standard into public text. Codified rules are, in essence, a set of diode-like digital switches: every chip is either inside or outside the threshold, either on the control list or off it. The advantage of a digital switch is transparency and predictability—businesses have clear rules to follow. Its fatal flaw, however, is sluggishness and rigidity. The moment rules are set down in black and white, they are doomed to lag behind engineers’ spec-downgrade workarounds. Over the past four years, a single export control regulation has spent four full rounds playing catch-up with the exact same company.
Because rules are written in plain text, they are naturally open to lobbying—and the compromises born of lobbying are likewise laid out in the open. In July 2025, after Jensen Huang made frequent visits across Washington departments and White House corridors, H20 exports to China resumed. But the price of this clearance was an explicit deal: NVIDIA had to surrender 15% of its sales revenue in China to the U.S. government. This take was announced personally by Donald Trump during a press conference, where he revealed that his initial asking price had been 20%.
An interesting detail surfaced a little over a month later: NVIDIA’s Chief Financial Officer admitted on a subsequent earnings call that the government had indeed explicitly expressed its demand for a 15% revenue share, but the administration had to date published no formal regulation codifying the fee into law. A revenue split directly impacting tens of billions in revenue for a multinational giant ultimately rested solely on a president’s verbal statements at a press conference, unwritten in any formal rulebook. When the H200 was approved for export in early 2026, this regulatory mindset escalated further, evolving into a composite set of restrictions: enforcing case-by-case review, imposing a quota cap stipulating that total exports to China could not exceed 50% of historical U.S. domestic sales for the same model, mandating inspection by independent third-party testing organizations, and finally levying an exclusive 25% tariff via presidential proclamation.
Even when enforcing against underground illicit channels, the U.S. side follows public judicial procedures. In March 2026, federal prosecutors in the Southern District of New York formally indicted a co-founder of Super Micro along with two other individuals, charging them with illegally smuggling approximately $2.5 billion worth of NVIDIA servers to China since 2024. The indictment detailed absurd particulars of how those involved filled server rooms with fake hardware to fool on-site inspections by U.S. export control officials. Anyone can download this public document directly from the court’s website.
Turning the lens to the Chinese machine, the operational logic takes on an entirely different form. Searching through all publicly available policy databases and ministerial announcements in China from December 2024 to August 2026 yields not a single written document explicitly banning NVIDIA by name. Yet on the industrial front, NVIDIA’s product lines have ground to a practical halt one after another. Sustaining this silent shutdown is a toolkit operated through the coordinated action of five distinct authorities and mechanisms.
The Cyberspace Administration of China (CAC) moved first. In July 2025, CAC summoned NVIDIA for formal inquiries, demanding that the company clarify whether its H20 chips contained backdoors or posed threats to Chinese user data security. Immediately following in August, CAC convened Tencent, ByteDance, and Baidu for a symposium, cutting straight to the core: in compute procurements, why choose NVIDIA instead of buying domestic? As Reuters reported, officials issued explicit warnings during the meetings that materials submitted by NVIDIA to the U.S. government under revenue-sharing agreements might well contain sensitive data from domestic clients.
China Customs (GACC) managed physical checkpoint interceptions. On January 13, 2026, the very day U.S. rules took effect, Chinese customs notified customs brokers within hours that H200 chips were barred from entry—offering neither oral nor written reasons, nor clarifying whether it was a temporary regulatory measure or a formal ban. By May, facing the consumer-grade RTX 5090D V2, customs similarly refused point-blank to issue import permits.
The National Development and Reform Commission (NDRC) established a case-by-case approval channel. For the approved procurement of H200 chips, domestic authorities instituted a strict NDRC individual case review system, requiring every transaction to declare its intended use and undergo quota approval. According to the FT, by the time the first batch of H200s was delivered in July 2026, ByteDance and Tencent had each received only about 10,000 units.
The State Administration for Market Regulation (SAMR) launched an antitrust investigation. As early as December 2024, SAMR opened an antitrust investigation into NVIDIA; by September 2025, SAMR announced a preliminary determination that NVIDIA had violated anti-monopoly laws. Yet since then, no public administrative penalty has been issued, leaving the investigation hanging in the air.
The Ministry of Commerce held the rare-earth countermeasure leverage. In October 2025, the Ministry of Commerce announced extraterritorial export controls on components containing Chinese rare earths; following a November US-China de-escalation framework, it announced a partial suspension of these measures against the U.S. The rare-earth card served as reverse leverage on the negotiating table to offset semiconductor restrictions, dynamically traded against chip issues.
The single greatest commonality among these five mechanisms when blocking specific NVIDIA products is that they yield not a single referenceable, formal administrative decision. While writing this article, I stumbled upon a self-evidencing phenomenon: for every U.S. action, I could provide links to official documents or court filings above; yet for every Chinese blocking action against specific NVIDIA products, the only citations I could provide were media reports and anonymous source accounts. This structural divergence in citations is itself proof of how the Chinese machine operates.
Unpacking this subtlety reveals four functional advantages. First is deniability. Since there is no written decision, there is no target to challenge on legal grounds, allowing the Ministry of Foreign Affairs to calmly assert that purchases are simply autonomous enterprise choices.
Second is scalability. Per a single-source FT report, the U.S. granted export allocations exceeding 400,000 units for the H200, yet domestic authorities released only 20,000 units over more than six months—an actual fulfillment rate under 5%. Case-by-case approval provides regulators with a valve to adjust the flow at will, controlling the degree of constriction entirely through the pace of approvals.
Third is reversibility. Because no formal ban was ever enacted from start to finish, should industrial demand shift in the future, the valve can be reopened at any moment without anyone having to publicly admit fault or retract a decision.
Fourth is un-lobbyability. Lobbying targets specific rules; without written rules, there is no target for modification. Jensen Huang could knock on every door in the White House and Department of Commerce in Washington, but in Beijing, he cannot even find a door with a matching nameplate to lobby.
The silent operation of the Chinese machine is by no means arbitrary; it possesses clear triggering conditions. A September 2025 Financial Times (FT) report disclosed the regulatory execution sequence: regulators first meet with domestic chipmakers to verify that the comprehensive performance of domestic chips in real-world business scenarios has caught up with NVIDIA’s corresponding China-tailored offerings; only after confirming that domestic substitution capabilities are ready do they formally block that specific imported product line. Performance parity first, ban follows. This strict temporal causality comes directly from statements made by sources in the report, with very little derived purely from reporting logic.
The entities summoned are nominally “domestic vendors,” but under the current industrial reality, the only player capable of volume delivery and absorbing the compute deficit in the short term is Huawei—a conclusion that represents a highly definitive inference within the industry. Companies like Cambricon and Moore Threads certainly exist, but in terms of scale, ecosystem, and delivery capacity, a significant gap remains between them and Huawei. Consequently, the underlying process can be roughly decoded as: ask Huawei if it’s ready first; once Huawei is ready, close the corresponding door.
The cadence between Huawei and regulators demonstrated remarkable synchronization in September 2025: on September 15, SAMR published its preliminary anti-monopoly determination; on September 17, the FT reported that the purchase ban had rapidly expanded to the workstation-grade RTX Pro 6000D; on September 18, Eric Xu unveiled a three-year roadmap for Ascend compute power at Huawei Connect, announcing the 950 series for 2026 and the 960 series for 2027 with full integration of proprietary HBM, while showcasing the Atlas 950/960 supernode systems. Three actions in three days, seamlessly linked in rhythm.
Whether Huawei can truly plug the compute gap across the entire market depends on its actual coordinates. At the single-chip physical performance level, Eric Xu candidly acknowledged on stage that, restricted by external sanctions, Huawei’s chips cannot be manufactured at TSMC, leaving single-die performance noticeably trailing NVIDIA.
Huawei’s engineering breakthrough lies in leveraging massive system-level engineering to offset generational single-chip gaps: its CloudMatrix 384 supernode clusters 384 Ascend 910C chips, achieving system-level compute that surpasses NVIDIA’s flagship GB200 NVL72 by roughly 40%. However, this system-level advantage comes at a massive cost in energy consumption and physical footprint, with overall power consumption reaching nearly four times that of the GB200. Commercially, this is essentially a resource trade-off: using five times the number of chips to offset a one-third single-chip gap, and relying on domestic Chinese industrial electricity prices—roughly one-quarter of U.S. rates—to keep this high-power, brute-force stacking solution economically viable.
Yet Huawei’s most lethal bottleneck today lies not in chip fabrication capacity, but in high-bandwidth memory (HBM). Stockpiles of original Samsung HBM memory accumulated earlier are nearing exhaustion, while domestic HBM supply this year is estimated to support packaging for only a few hundred thousand chips. Whether the 950DT, equipped with in-house HBM, can achieve smooth, stable mass production by late 2026 stands as the crucial litmus test for this substitution system.
Tracing the thread this far yields an answer to the core question running through this article: why must China’s blocking method be so subtle? Industrial substitution in semiconductors becomes ready section by section. The inference market catches up first, followed by medium-and-small model training of specific configurations, and only later by top-tier frontier foundation models. The readiness timeline for each segment is entirely distinct, as are the product lines that must be closed. This demands a control valve capable of segment-by-segment, timed, and variable-aperture adjustment. The NDRC’s case-by-case approval alongside Customs’ tacit cooperation serves precisely as such a valve: regulators can dynamically decide how much to approve today, to whom, and for what workloads, based on the actual delivery status of the domestic supply chain.
American written regulations cannot achieve this. The Federal Register cannot publish a clause reading: “Release quantities adjusted daily based on Huawei’s mass production status.” Digital switches are strictly all-on or all-off; and per the single-source FT report cited above, only an analog valve like China’s can dial in a faint, lifeline-sustaining aperture of under 5%.
One more structural layer completes the picture: official channels shut down only incremental procurements, while grey-market stockpiles have continuously sustained domestic industry. Chinese tech giants stocked up in advance during the first half of 2025, hoarding over 1.3 million H20 chips; underground smuggling networks never ceased, with the $2.5 billion Super Micro case marking merely the tip of the iceberg; and for a full year between May 2025 and May 2026, scores of Chinese firms legally purchased Blackwell chips directly through overseas subsidiaries established in Malaysia, Singapore, and the UAE. According to an industry source cited by Reuters, as many as hundreds of thousands of chips flowed out through this compliant channel before U.S. authorities issued emergency weekend guidance to plug the loophole. Freezing official channels while tolerating grey channels can be understood as an intentional design: it grants domestic chips room to grow within the official market, while allowing pressing frontier foundation model training needs to draw support from grey avenues.
Jensen Huang once valued China’s compute market at $50 billion annually. Today, the two machines have jointly torn that market in two.
Inference workloads and routine training inside China are transitioning wholesale toward fully domestic hardware. Huawei’s AI chip revenue climbed from $7.5 billion in 2025 and is projected to reach approximately $12 billion in 2026. Meanwhile, training demands for frontier foundation models scrape by on tightly controlled approval channels, overseas compute, and grey imports. Bernstein predicts NVIDIA’s share of China’s AI chip market will plummet from around 40% in 2025 to 8% in 2026.
Even if the U.S. were to lift all export controls tomorrow, research firm Omdia assesses that buyer confidence has suffered an irreversible shift: once buyers determine in practice that domestic chips are fundamentally sufficient for business operations, that conviction will not be easily reversed simply because external controls ease. Jensen Huang himself has completely stripped revenue expectations for the Chinese market from earnings guidance, telling investors bluntly that he holds zero expectations for sales in China.
The endgame of this contest displays a fascinating symmetry: two machines built on opposing premises are moving toward one another. The American machine, originally grounded in explicit written rules, is sliding rapidly toward executive discretion under a transactional administration: taking a 15% revenue cut from NVIDIA remains without codified regulations to this day, core policy moves are routinely broadcast via social media platforms, and levying 25% tariffs on high-tech products relies on presidential tariff proclamations to bypass regular procedures. In response, the U.S. Congress introduced emergency legislation demanding that all export control measures be re-codified, essentially forcing the legislature to retroactively lecture the executive branch on written law. Conversely, the Chinese machine—historically defined by subtlety and discretion—is accelerating toward institutionalization: mandates dictating that state-funded data centers use exclusively domestic chips have become explicit directives, nine domestic AI chips have been formally listed in government procurement catalogs, and according to the FT, top domestic cloud providers were instructed to reserve roughly 30% of their procurement quotas for domestic silicon.
One side slides from codified rules into discretion, while the other hardens discretion into written mandates. The eventual outcome will almost certainly be a thoroughly fractured dual-stack world: in China’s vast incremental compute market, NVIDIA’s market share will languish in low single digits priced by political maneuvering, while two state machines on opposite sides of the ocean work their respective switches and valves, each remaining convinced that its own way of blocking is the smarter one.