# Why US Chip Curbs Failed: 678K Chips, 34% Share > Source: > Published: 2026-11-14 11:16:35+00:00 Washington’s three-year effort to keep China off the most advanced AI hardware now has a failure verdict attached to it, and the evidence is piling up in public. Nvidia CEO Jensen Huang called the policy a failure, a think tank says overseas subsidiaries of Chinese firms got advanced chips without licenses, and The Register is relaying a claim that [Chinese AI labs are training models on Nvidia’s Blackwell hardware despite export controls](https://www.theregister.com/ai-and-ml/2026/11/13/chinese-ai-labs-are-training-models-on-nvidias-blackwell-despite-export-controls-says-ex-openai-engineer/5309891). The question this November 14 analysis answers is simple: why have US chip export controls failed to do what they were built to do? The short answer has five parts. The rules policed boxes shipped across borders while Chinese buyers rented computing power remotely. Third-country data centers in Malaysia and Singapore became the weak link. Rule changes kept the market guessing. The agency that enforces the rules lacks the staff to chase violators. And Chinese firms used the pause to build alternatives. Our earlier reporting on [six Chinese AI models tied to Blackwell with location unknown](https://shattered.io/chinese-ai-models-nvidia-blackwell-depue-six-models-location-unknown/) covered one slice of this story. This piece pulls the whole chain together. One caution before the details. The sources do not establish a single, officially documented cause that explains the entire failure of the controls. Claims that specific Chinese AI systems were trained on particular quantities of restricted Nvidia chips also remain unconfirmed. Everything below sticks to what named outlets and institutions have put on the record, and flags the rest as reports. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. [Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F) ## The Verdict: US Chip Export Controls Under Fire Jensen Huang, chief executive officer of Nvidia, said the US semiconductor export controls on China were a failure. He made the case at Computex in 2025, and the argument has aged into a talking point for critics on both sides of the Pacific. His logic runs like this: the restrictions cost American companies sales and pushed Chinese companies to build their own alternatives. Huang has an obvious commercial stake, so his view deserves scrutiny. But he is no longer alone. A Center for a New American Security (CNAS) analysis titled [The Nvidia Scandal Is a Wake-Up Call for Congress](https://www.cnas.org/publications/cnas-insights/cnas-insights-the-nvidia-scandal-is-a-wake-up-call-for-congress) argues from the opposite direction. CNAS wants tighter enforcement, not looser rules, yet it reaches a similar diagnosis: the controls leak. It identified loopholes involving remote access to advanced computing through data centers in third countries, including Malaysia and Singapore. So the critics span a wide range. A chipmaker says the rules went too far. A security think tank says they did not go far enough and were poorly enforced. A semiconductor professor says they backfired. Those camps disagree on the cure but converge on the symptom, which is why the story is breaking through now. ## The Numbers: 678,000 Accelerators and a 34% Share The cleanest hard data comes from IDC, as cited by Reuters. Nvidia shipped 678,000 AI accelerators to China during the first six months of 2026, a 44% decline from the same period a year earlier. Nvidia held 34% of China’s data-center AI-chip market in the first half of 2026. Those figures cut both ways. A 44% drop looks like a policy win at first glance. Washington wanted Nvidia’s position in China to shrink, and it did. But a 34% share is still more than a third of the market, and 678,000 units in six months is a large flow of silicon for a market the controls were supposed to choke. A policy aimed at denying capability does not look finished when the targeted vendor still sells that many accelerators there. Our own arithmetic on the IDC figures suggests the year-earlier half carried roughly 1.2 million units, since 678,000 is 56% of the prior level. Treat that as an implied estimate, not a published number. Readers who want the market-share angle in more detail can see our earlier look at [Nvidia’s China share slipping to 34%](https://shattered.io/nvidia-china-ai-chips-34-percent-share-600000-hong-kong-2026/). | Metric | Figure | Source | |---|---|---| | Nvidia AI accelerators shipped to China, H1 2026 | 678,000 | IDC, cited by Reuters | | Change vs. H1 2025 | Down 44% | IDC, cited by Reuters | | Nvidia share of China data-center AI-chip market, H1 2026 | 34% | IDC, cited by Reuters | | US share under the H200 reimportation arrangement | 25% | CNAS | | Reported chip flow tied to Megaspeed network | About $7.5 billion (reported, unconfirmed) | Fortune | ## Remote Access: The Loophole in Malaysia and Singapore The biggest structural flaw is easy to describe. US rules mostly governed the shipment or transfer of hardware. They said little about who could log in to a rack of GPUs once it sat in a legal jurisdiction. A Chinese lab that cannot import a Blackwell-class card can still rent time on one in Kuala Lumpur or Singapore, and the controls did not clearly forbid that. Research gathered for this report indicates Nvidia chips could be sold legally to customers in Malaysia and Singapore and then allegedly resold, re-exported or made available to Chinese companies. The distinction between owning a chip and renting its output is the whole game. Cloud computing turned a border control into a paperwork exercise. Congress has noticed. Reports say the House passed the Remote Access Security Act, which targets the use of advanced AI processors hosted outside China, and that the bill was awaiting Senate action in November 2026. We could not independently confirm the bill’s text, so treat the details as reported. Our earlier coverage of [Chinese AI firms rushing to reserve Nvidia capacity abroad](https://shattered.io/chinese-ai-firms-rush-nvidia-overseas-capacity-spy-scandal-export-curbs-2026/) shows how quickly buyers move when they expect a door to close. ### How Megaspeed Fits the Pattern CNAS identified Megaspeed as a company that hosted advanced [AI chips](https://shattered.io/category/hardware/) for Chinese companies to access remotely and may have facilitated illicit transfers. Fortune reported that the company built a network in Singapore and Malaysia, with a potential flow of about $7.5 billion in Nvidia chips. That dollar figure comes from reporting and has not been confirmed by a court or regulator. The model is the point. Megaspeed allegedly let Chinese customers rent capacity without importing restricted chips into China. If true, no single shipment crossed a prohibited line. That is how [a Singapore chip hub at risk](https://shattered.io/nvidia-singapore-chip-hub-megaspeed-7-5-billion-breakingviews-2026/) becomes a policy problem rather than a one-company problem. ## Overseas Subsidiaries Got Chips Without Licenses CNAS also reported that overseas subsidiaries of Chinese companies obtained advanced Nvidia chips without licenses. This is a separate hole from remote rental. Here the buyer is a Chinese-owned entity registered abroad, and the rules did not stop it from taking delivery. The pattern shows how corporate structure defeats geography-based rules. A parent company in China can set up a subsidiary in a third country, buy hardware there, and run workloads for the parent. Unless regulators look through the corporate wrapper, the transaction looks like an ordinary sale to a local firm. Closing that gap requires ownership tracing and end-user checks, both of which cost time and people. Reports describe a proposed framework that would address part of this. Under it, seven additional countries would need export licenses for Nvidia Blackwell-class or more advanced processors. Those countries are Singapore, Malaysia, Thailand, Indonesia, Brazil, Norway and South Africa. The rules would also require disclosure of chip recipients and intended uses, plus continuing monitoring, and US-operated data centers could win an exemption if they block remote access from arms-embargoed countries such as China and Russia. The proposal has not taken final form, so its details may shift. We covered the fight over it in [our look at the seven-nation curbs and the H200 tariff](https://shattered.io/jensen-huang-exit-us-china-ai-chip-curbs-seven-countries-h200-tariff-2026/). ## Interconnect Bandwidth: A Spec Vendors Could Design Around The technical design of the rules created its own escape hatch. Export-control rules used interconnect bandwidth, the communication speed between processors or chips, as a control criterion. Vendors could reduce interconnect bandwidth while retaining much of the underlying processor capability. That matters because AI training is a team sport. Large models spread work across many chips, and the links between them set the pace. A rule that throttles the links slows big clusters. But it does less to stop smaller jobs, inference workloads and clever software that works around slow interconnects. A chip with trimmed links still computes. The Nvidia H20 is the textbook case. It was custom-built for the Chinese market and later banned by the US administration, which means regulators first accepted the design philosophy and then reversed it. Chinese buyers who stockpiled or rented around the rules came out ahead of buyers who waited for clarity. Different chip architectures trade raw compute against interconnect in different ways, which is why one threshold rarely fits every design. ## Policy Whiplash: H20, H200 and the AI Diffusion Rule A deterrent only works if buyers believe it will hold. This one flipped repeatedly. Reports indicate the H20 was blocked in April 2025, then sales resumed in July 2025 under revised US policy. The administration later permitted sales of the [Nvidia H200](https://shattered.io/pla-linked-university-nvidia-h200-two-chips-123000-blacklist-2026/) to China under a reimportation arrangement that gave the United States a 25% share. Reports say H200 licenses moved to case-by-case review in January 2026. On top of that, the US government suspended the AI Diffusion Rule without clarifying which provisions would continue to be enforced. Huang had attacked that framework in 2025, saying its founding assumptions were flawed. Whatever one thinks of the rule, a suspension with no guidance left companies to guess, and guessing favors the side that moves fast. The incentive effects are predictable. When a rule might tighten next quarter, buyers reserve capacity now. When a ban might lift, suppliers lobby for the lift. Neither response advances the original goal. Our piece on [Lutnick’s softer line on Nvidia export controls](https://shattered.io/lutnick-softer-line-nvidia-export-controls-13-months-no-entity-list-2026/) tracks the same drift from the enforcement side. ## Enforcement: An Under-Resourced Bureau of Industry and Security Rules are only as strong as the agency that polices them. CNAS described the US Bureau of Industry and Security (BIS) as under-resourced and understaffed for the investigations and enforcement required. That is a blunt judgment about the body that writes licenses, vets end users and chases diversion cases. Consider the workload. Tracing a chip from a distributor to a data center to a remote customer means reading corporate filings in several jurisdictions, cooperating with foreign regulators and sorting through shell structures. Each case consumes analysts for months. Meanwhile, the volume of advanced accelerators moving through global supply chains keeps growing, and demand for the memory side of the stack is climbing too. CNAS framed the Nvidia scandal as a wake-up call for Congress, which controls BIS funding and authorities. That framing points to a fix that has nothing to do with chip specs. More agents, more license reviewers and clearer legal authority over cloud access would raise the cost of evasion without a single new rule. ## Where Export Controls Did Work: Memory and Lithography The record is not uniformly bleak. Export controls worked more effectively for specialized technologies that are difficult to reproduce, including high-bandwidth memory and advanced lithography equipment. A GPU is a product that gets sold in thousands of boxes. A leading-edge lithography tool is a rare machine built by a tiny set of firms, and no one can smuggle one in a suitcase. That contrast carries a lesson for policy designers. Controls bite where supply is concentrated, hard to copy and slow to ship. They leak where the product is a commodity with a global channel and where the regulated activity (using compute) can happen anywhere there is a network cable. The strongest case for the current regime rests on the chokepoints that held, not the ones that leaked. | Control target | Outcome | Why | |---|---|---| | High-bandwidth memory | More effective | Specialized and hard to reproduce | | Advanced lithography equipment | More effective | Specialized and hard to reproduce | | Chip shipments to China | Leaky | Overseas subsidiaries obtained chips without licenses (CNAS) | | Interconnect bandwidth limits | Partly circumvented | Vendors cut link speed but kept much of the processor capability | | Remote access to hosted chips | Gap | Rules focused on hardware transfer, not rented compute | | Enforcement capacity | Weak | BIS described as under-resourced and understaffed (CNAS) | ## Expert Voices: Huang, CNAS and Rakesh Kumar Huang’s public comments supply the sharpest sound bites. In remarks covered by [Reuters](https://www.reuters.com/world/china/nvidia-says-us-export-controls-ai-china-were-a-failure-2025-05-21/), the [Nvidia CEO](https://shattered.io/johny-srouji-nvidia-ceo-report-one-source-zero-confirmed/) said: “I think, all in all, the export control was a failure.” In the same Reuters coverage he added: “The fundamental assumptions that led to the AI diffusion rule in the beginning, in the first place, have been proven to be fundamentally flawed.” He also made an argument about Chinese industry. As reported by [DW](https://www.dw.com/en/nvidia-chief-calls-us-export-curbs-on-ai-chips-to-china-a-failure/a-72617920), Huang said: “The local companies are very, very talented and very determined, and the export control gave them the spirit, the energy and the government support to accelerate their development.” That is a claim about incentives, and it is the part of his case that outside experts find hardest to dismiss. Rakesh Kumar, a professor, reached a similar conclusion when he spoke to [Fortune](https://fortune.com/2026/10/16/china-chip-export-controls-failed-rakesh-kumar/). Kumar argued that the controls pushed China to spend heavily on building an ecosystem that can now serve these applications on its own, and he said the policy had the opposite of its intended effect. CNBC’s 2025 report, Nvidia’s Jensen Huang, semiconductor experts think US chip curbs failed, shows the view was already mainstream among chip specialists last year. The critique has limits. Huang sells the hardware in question, and an academic’s reading of Chinese industrial strategy is a judgment, not a measurement. CNAS offers the counterweight: the evidence of leakage points toward better enforcement, which is a different lesson than abandoning controls. ## Historical Context: From 2022 Rules to the 2026 Scandal The US has restricted Nvidia’s most advanced AI chips from China since 2022, aiming to slow Chinese military and artificial-intelligence capabilities. That original logic assumed China could not match the best American silicon for years and that denying access would open a lasting lead. Four years later the lead in raw chips remains, but the denial does not. The sequence since then reads like a cycle of tighten, loosen and patch. Each tightening produced a compliant product, such as the H20. Each compliant product then drew political fire or got banned. Loosening arrived through licensing deals, including the 25% arrangement on the H200. And each new loophole, from remote rental to overseas subsidiaries, called for yet another patch. | Date | Event | Basis | |---|---|---| | 2022 | US begins restricting Nvidia’s most advanced AI chips to China | Reports | | April 2025 | H20 blocked (reported) | Reports | | May 2025 | Huang calls the controls a failure at Computex | Reuters, CNBC | | July 2025 | H20 sales resume under revised policy (reported) | Reports | | January 2026 | H200 licenses move to case-by-case review (reported) | Reports | | H1 2026 | 678,000 accelerators shipped to China, down 44% | IDC via Reuters | | October 16, 2026 | Kumar tells Fortune the controls backfired | Fortune | | November 11, 2026 | Fortune reports on Megaspeed and Singapore, Malaysia chip flows | Fortune | | November 13, 2026 | CNAS calls the Nvidia scandal a wake-up call for Congress | CNAS | ## Market Impact: Nvidia, Chinese Rivals and Cloud Hosts For Nvidia, the impact is mixed but real. China still accounts for a meaningful slice of its accelerator business, yet shipments fell 44% year over year in the first half of 2026. The company faces a policy environment that can change by memo, plus a wider scandal that has put its leadership in the headlines. ### Chinese Chip Designers and Cloud Buyers Chinese chip designers and data-center operators gain from every quarter of uncertainty. Kumar’s argument, relayed by Fortune, is that the controls handed them funding and a captive customer base. Chinese cloud buyers, meanwhile, hedge by booking foreign capacity and maintaining domestic options at the same time. That dual-sourcing behavior is exactly what a strict regime would try to prevent. ### Data-Center Hosts in Third Countries Hosts in Singapore, Malaysia and other named jurisdictions face the heaviest compliance burden if licensing arrives. The reported framework would force recipient disclosure and ongoing monitoring, and would exempt US-operated sites that block remote access from embargoed countries. Operators that cannot prove who their customers are risk losing access to top-tier Nvidia parts. Operators that can prove it may gain a premium for being clean. China’s government has also reacted. Al Arabiya reported on November 13 that Beijing warned against new export controls on US chips after the Nvidia scandal. Expect diplomatic friction to grow alongside any new rule. ## Competitive Comparison: Four Ways to Control AI Compute Policymakers have four main levers, and the record so far shows different strengths for each. The comparison below draws on the confirmed findings above plus analysis. It rates each lever’s reach, not its political appeal. | Approach | Strength | Weakness | |---|---|---| | Hardware shipment bans | Simple to state and to audit at the border | Defeated by subsidiaries and third-country resale | | Performance and interconnect thresholds | Technical, targets training clusters | Vendors redesign chips to sit just below the line | | Licensing and end-user disclosure | Creates a paper trail and consequences | Needs enforcement staff that BIS reportedly lacks | | Remote-access restrictions | Closes the rented-compute gap | Hard to monitor, depends on host cooperation | The analysis suggests no single lever suffices. The reported seven-country plan is a hybrid, pairing licensing with remote-access conditions. That is a sensible direction on paper, but it inherits the staffing problem that CNAS flagged. ## Predictions: What Comes Next for US Chip Export Controls These forecasts are our analysis, not statements from any named source. They rest on the patterns documented above. - **Remote access becomes the main battleground.** The House-passed Remote Access Security Act, if the Senate acts, would turn the rented-compute gap into a legislative target. Expect hosts to start screening customers before any law forces them to. - **Licensing expands to more countries.** The reported seven-nation framework, with its Blackwell-class threshold, is likely to land in some form. The final list may differ from the draft. - **Congress funds enforcement.** The CNAS argument that BIS is understaffed gives lawmakers a low-controversy move. Added headcount may arrive before any new technical rule. - **Nvidia’s China share keeps sliding.** With shipments already down 44% in the first half of 2026 and domestic alternatives maturing, the 34% share looks more likely to fall than to recover. - **Policy swings continue.** The H20, H200 and AI Diffusion Rule episodes show the administration changes course often. Companies should plan for another reversal. Near-term catalysts include congressional hearings and Nvidia’s next earnings report, both of which will force executives to explain their China exposure on the record. ## What This Means for Hardware Buyers and Builders For readers outside the policy world, the practical effects are indirect but real. Export fights ripple through GPU supply, pricing and availability. When Chinese buyers scramble for capacity, they compete for the same advanced parts that US and European labs want. Hardware coverage on our [hardware desk](https://shattered.io/hardware/) tracks those knock-on effects for chips, memory and cooling. Cloud customers should also watch their providers’ compliance posture. If licensing and recipient disclosure arrive, hosts will ask more questions of tenants, and some workloads could face new checks. Teams that rent GPU time abroad should expect paperwork, not just invoices. ## Frequently Asked Questions ### Have US chip export controls really failed? It depends on the metric. Nvidia’s China shipments fell 44% in the first half of 2026 and controls on memory and lithography held up better. But Nvidia still held 34% of China’s data-center AI-chip market, and CNAS documented leaks through overseas subsidiaries and remote access. Huang and Kumar both call the policy a failure, while CNAS calls for tighter enforcement instead. ### What is the remote access loophole? Rules mostly governed hardware shipments. Chinese companies could still rent advanced computing hosted in third countries such as Malaysia and Singapore, because the use of hosted compute fell into a regulatory gap. ### Who is Megaspeed? CNAS identified Megaspeed as a company that hosted advanced AI chips for Chinese companies to access remotely and that may have facilitated illicit transfers. Fortune reported a potential flow of about $7.5 billion in Nvidia chips. These are reports and allegations, not court findings. ### What is the Nvidia H200 arrangement? The administration permitted H200 sales to China under a reimportation arrangement that gave the United States a 25% share. It followed the H20, a chip custom-built for China that the administration later banned. ### Which countries could face new licensing rules? Reports say BIS prepared controls covering Singapore, Malaysia, Thailand, Indonesia, Brazil, Norway and South Africa for Blackwell-class or more advanced processors. The plan was not final at the time of writing. ### What happened to the AI Diffusion Rule? The US government suspended it without clarifying which provisions would continue to be enforced. Huang had criticized its underlying assumptions in 2025. ### Did Chinese labs train models on restricted Blackwell chips? The Register relayed a claim from an ex-OpenAI engineer that they did. Claims about specific models and chip quantities remain unconfirmed, and no regulator has publicly verified them. ### Which export controls worked best? Controls on specialized technologies that are hard to reproduce, including high-bandwidth memory and advanced lithography equipment, worked more effectively than controls on finished GPUs.