- The Trump administration rescinded Biden’s Framework for AI Diffusion two days before its compliance deadline and replaced it not with a new rule but with a sequence of executive deals — including a revenue-share arrangement giving the US government 15–25% of Nvidia’s China chip revenues.
- The policy environment for global AI chip trade has become structurally unpredictable: the H20 was banned in April 2025, unbanned in August with a revenue share, and H200 access was granted in December — three policy reversals within eight months.
- China retains access to H20 and potentially H200 GPUs at volume limits, while domestic alternatives (Huawei Ascend 910C) approach approximately 60% of H100 performance — the performance gap is narrowing as the policy framework continues to shift.
Key Claim: The replacement of rules-based AI export controls with a deal-by-deal executive approach has transferred strategic uncertainty from chip performance metrics to US policy timing — a different and harder-to-model risk for global AI infrastructure investment.
The Biden administration spent three years building a rules-based framework for controlling advanced AI chips. The Trump administration dismantled the centrepiece of that framework two days before it took effect, then replaced it with a series of executive deals — including an arrangement that earns the US government a percentage of Nvidia’s China revenues. The second-order effect is not simply whether China gets more or fewer chips. It is that the policy environment governing global AI infrastructure investment has become materially unpredictable, with rules shifting within a single earnings quarter.
What the Biden Rule Actually Said
On 15 January 2025, the Bureau of Industry and Security (BIS) published the “Framework for Artificial Intelligence Diffusion” in the Federal Register (document 2025-00636) as an interim final rule, with a compliance start date of 15 May 2025. It was the most architecturally ambitious export control rule the US had ever applied to technology.
The rule did three things that previous controls had not. First, it sorted every country in the world into one of three tiers: Tier 1 (the US and 18 close partners, including the UK, Japan, South Korea, and most of Western Europe — unrestricted); Tier 2 (most of the rest of the world — licensed through a new data-centre Validated End-User programme with a presumption of approval); and Tier 3 (adversaries including China, Russia, Iran — near-total restriction). Second, it placed export controls on AI model weights for the first time, not just hardware. Third, it attempted to close the loophole of third-country data centres being used to provide Chinese entities with access to controlled compute.
The Tier 2 programme was designed to create a global ecosystem of accredited data centres that met US security standards — essentially an attempt to make the United States the licensing authority for large-scale AI infrastructure worldwide. For a broader view of how AI supply-chain constraints shape infrastructure investment, see our analysis of the AI semiconductor supply chain and TSMC capacity pressure.
Why Trump Rescinded It
On 13 May 2025 — two days before the compliance deadline — BIS announced it was initiating rescission of the rule. Three reasons were stated: the framework was “overly bureaucratic”; it “stifled American innovation”; and it “undermined US diplomatic relations” by placing close NATO allies such as Portugal and Austria in Tier 2 rather than Tier 1. The last point was substantive — the Tier 2 classification created compliance burdens for US chip companies selling to allied-country data centres and generated diplomatic friction with partners the US wanted to keep close.
No replacement rule was published. BIS said a new rule would follow but gave no timeline. In its place, BIS issued three enforcement guidance documents: warnings about US AI models being used for Chinese training and inference; supply-chain diversion tactics; and risks from Chinese advanced computing chips. BIS also issued a “high probability of enforcement” notice for AI and semiconductor violations — signalling that while the rule was gone, enforcement pressure on the underlying controls was not.
The Policy That Replaced It: A Deal-by-Deal Sequence
What followed was not a replacement framework but a sequence of discrete executive decisions, each reversing or modifying the one before.
April 2025 — H20 licence requirement. The Trump administration imposed a licence requirement on Nvidia’s H20 GPU, the chip Nvidia had specifically engineered to fall below the prior control thresholds. Nvidia disclosed it would record approximately $5.5 billion in charges for Q1 FY2026 (ending 27 April 2025) for inventory and commitments now subject to the new restriction. AMD disclosed a $1.5 billion revenue impact from concurrent restrictions on its MI308 chip.
August 2025 — Reversal with a revenue share. Three months after the ban, Trump announced that Nvidia and AMD could resume shipments of H20 and MI308 chips to China, on condition of paying 15% of those revenues to the US government. Trump noted he had originally sought 20%. The 15% figure is based on news reporting (NPR, Fortune); no formal regulation or executive order establishing the legal mechanism for the revenue-share arrangement had been published as of early April 2026 — itself a reflection of the ad hoc nature of the policy shift. According to Fortune, the arrangement was described by Georgetown CSET analysts as a “pay-to-export” model that risks “diluting US strategic leverage” by treating market access as a revenue instrument rather than a security tool.
December 2025 — H200 approved at 25%. On 8 December 2025, Trump announced Nvidia’s H200 — a substantially more powerful chip than the H20 — could be sold to approved Chinese customers, with the US government receiving 25% of those revenues. As with the August arrangement, the 25% figure is based on news reporting; no formal regulation or executive order establishing the legal mechanism had been published as of early April 2026. AMD and Intel were expected to be covered under a similar framework.
13 January 2026 — Final rule on case-by-case review. BIS formalised the policy shift in a final rule revising its licensing policy for AI semiconductors — including the H200 and AMD MI325X — to China and Macau. The rule shifted the review standard from “presumption of denial” to “case-by-case” subject to specified conditions. This is not the same as approval; it means each transaction is evaluated individually rather than defaulting to rejection. (The rule’s full conditions are summarised in Baker McKenzie’s analysis; the Federal Register notice had not been indexed as of publication.)
March 2026 — Worldwide licensing draft reported. Bloomberg reported on 5 March 2026 that Commerce had drafted regulations requiring US government approval for AI chip exports to any country worldwide. The draft uses a volume-based three-tier structure: shipments of fewer than 1,000 GPUs receive standard review; medium-volume deployments require preclearance before a licence application can be submitted; orders of 200,000 GPUs or more require formal government-to-government certification from the destination country. As of early April 2026, the draft had not been published in the Federal Register for public comment.
What China Can and Cannot Access
The practical question for China’s AI infrastructure is which chips are available and what performance gap remains.
The H100, A100, and their direct successors remain fully restricted to China under controls that predate both administrations. The H20 — a stripped-down Hopper variant — is now available under the August 2025 licensing arrangement. H200 exports to approved Chinese customers are possible under the January 2026 case-by-case regime; no completed H200 shipments to China had been publicly disclosed as of early April 2026.
On the domestic side, Huawei’s Ascend 910C is the benchmark for Chinese alternatives. CSIS analysis from early 2025 found the 910C achieves approximately 60% of H100 inference performance and that internal Huawei employees described the CANN software ecosystem as “difficult and unstable.” CFR analysis found that even under aggressive production assumptions — 800,000 Ascend chips in 2025 — Huawei’s aggregate computing output would represent roughly 5% of Nvidia’s total capacity. Huawei’s projected 2026 chip is expected to underperform the 910C, and analysts assess no competitive successor is imminent.
DeepSeek’s January 2025 open-source release complicated this picture. The model’s apparent efficiency on restricted hardware — Nvidia H800s — prompted debate about whether chip controls were effective. CSIS found that DeepSeek chief executive Liang Wenfeng explicitly cited chip bans as “the greatest challenge” facing his company, and that Chinese firms required two to four times more computing power when restricted to H800s versus H100s. MIT Technology Review reported that DeepSeek had in fact used a substantial stockpile of H100s accumulated before tighter controls via shell companies and third-country intermediaries — raising questions about enforcement gaps rather than control effectiveness.
The Strategic Debate
The policy shift has divided analysts along a specific axis: whether the primary goal of export controls is to maintain a performance gap in AI hardware, or to generate leverage that can be converted into diplomatic and economic outcomes.
CFR argues the US currently holds a commanding and growing hardware lead — best US chips are approximately five times more powerful than Huawei’s — and that loosening controls sacrifices that advantage for short-term revenue. Approving H200 exports would provide China with computing power equivalent to years of domestic Huawei production, potentially enabling Chinese companies to build globally competitive AI infrastructure services.
RAND and Brookings have each argued the controls need to be smarter rather than simply tighter — that blanket hardware restrictions create pressure for algorithmic efficiency breakthroughs (as seen with DeepSeek) while enforcement gaps allow hardware stockpiling. ITIF contends that overly broad restrictions harm US AI leadership by cutting domestic companies off from significant revenue while pushing allied-country customers toward non-US suppliers.
What neither side addresses directly is a structural question that, in our editorial assessment, deserves scrutiny: the revenue-share model creates a direct financial incentive for the executive branch to approve chip sales, which sits in tension with the independent security-assessment function that export controls are designed to perform. No policy institution has publicly raised this conflict-of-interest concern as of early April 2026, but the incentive structure is inherent in the model.
This governance gap mirrors a pattern we have tracked in domestic AI policy. The US federal framework — built on executive orders and agency guidance rather than statute — has consistently struggled to keep pace with industry developments, as we detailed in our analysis of the US AI regulation governance gap. The state-level dimension of that gap is examined separately in our piece on US state AI law fragmentation.
Implications: What to Watch
Four signals will indicate where this policy lands.
The replacement rule. BIS committed in May 2025 to publish a replacement for the Diffusion Rule. Eleven months later, no replacement has appeared in the Federal Register. If the worldwide licensing draft reported in March 2026 becomes the replacement, it would represent a significant expansion of executive discretion over AI supply chains — extending beyond China to cover allied-country data-centre buildouts in the Middle East, Southeast Asia, and Europe.
Allied coordination. The Biden rule’s Tier 1 / Tier 2 structure was designed to create a multilateral control architecture. The Trump administration’s bilateral deal model requires negotiating separately with each country and company. Japan, the Netherlands, and the EU have their own semiconductor export control regimes. As of early April 2026, no allied government had publicly announced a formal policy response to the US shift from the Diffusion Rule to the deal-by-deal approach. Whether those regimes align with, diverge from, or are instrumentalised by the US approach will determine the practical effectiveness of any controls. CSIS analysis notes that allies’ legal authority to implement parallel controls varies significantly.
Enforcement against diversion. CSIS identified large-scale chip smuggling networks as the principal gap in the existing control architecture. DeepSeek’s H100 stockpile — assembled before tighter controls via shell companies and third-country intermediaries — illustrates the problem. Enforcement resources at BIS have historically been limited relative to the scope of monitoring required. The May 2025 “high probability of enforcement” notice suggests awareness of the problem; whether resourcing follows is a separate question.
Chip company earnings disclosures. Nvidia and AMD will quantify the revenue impact of the 15% and 25% government revenue-share arrangements in forthcoming quarterly filings. The scale of those figures will indicate how much of China’s AI infrastructure buildout is flowing through the licensed channel versus alternatives.
This article was produced with AI assistance and reviewed by the editorial team.Related posts:
Section 232 Semiconductor Tariff Report: What AI Infrastructure Buyers Must Model Now
The EU’s High-Risk AI Rules Take Effect in August. The Compliance Infrastructure Barely Exists.
RAG vs Fine-Tuning vs Prompting: Enterprise LLM Architecture Decision
The Packaging Bottleneck: Why TSMC’s CoWoS Lines — Not Its Fabs — Now Control AI Chip Supply
