How Trump’s H200 Deal Reshaped the New AI-Based World Order

The US government walked away from the AI Diffusion Rule in May 2025, then spent the rest of the year shipping its most advanced chip to China. In December, the Trump administration reversed export controls on Nvidia’s H200, a chip nearly six times more powerful than the downgraded H20, opening a one-year window that hands the US Treasury a 15 to 25 percent cut of chipmaker revenues from Chinese sales. What is left is a market-share-first US policy that has rearranged the winners and losers of the AI race, and put every Tier-2 nation between allies and adversaries in a strategic bind.

The Trump administration’s bet, written into the National Security Strategy, is that the United States wins when the world builds AI on American technology. What is forming, the Atlantic Council’s Technology Programs argue, is an AI-based world order, with the rules set by where the next data center goes and whose chip powers it.

From Security to Sales: How US AI Policy Pivoted

The Biden administration published the Framework for Artificial Intelligence Diffusion on January 13, 2025, less than a week before leaving office. It split the world into three tiers and tried to write a tier system for advanced chip exports into US law. Seven months later, the Trump administration revoked the rule, a week before it was due to take effect on May 15. The replacement, framed as “much simpler” guidance, is being negotiated government by government.

The Atlantic Council and CSIS both describe the new direction as a transactional pivot. Commerce Secretary Howard Lutnick defended it in July 2025 by saying US firms should “sell the Chinese enough that their developers get addicted to the American technology stack.” The pivot culminated in two one-year concessions in late 2025: a waiver of the commerce department’s 50 percent rule effective November 10, and the December 8 reversal of H200 export controls.

On October 30, Trump and Xi Jinping met in Busan, South Korea. In return for the November waiver, China delayed its own rare-earth export controls for one year.

Under the December reversal, the US Treasury takes 15 percent of most chip revenues and 25 percent of advanced chip revenues from Chinese sales. Bremmer’s June 2026 column on the forces shaping markets and geopolitics reads this as part of a US strategy of “politically unconstrained AI development.” The reversal is set to expire in late 2026, by which point Washington will have to decide whether the concessions become precedent or remain exceptional.

The Tier System That Almost Was

The Diffusion Rule put 18 countries in Tier 1, the Five Eyes partners, twelve Western European and Nordic NATO members, and the semiconductor heavyweights Japan, South Korea, Taiwan, and the Netherlands. CSIS records the membership as a near-frictionless first-class cabin for advanced GPUs. Tier 2 was the rest of the US trading partners, lumped together from India and Switzerland to Saudi Arabia and Yemen. Each was capped at twelve months behind the frontier.

Defined as 10²⁶ FLOP, the frontier was an order of magnitude beyond the 10²⁵ FLOP where models like GPT-4, Gemini 1 Ultra, and Claude 3 Opus operate. The breakdown of the diffusion rule’s numeric caps put hard numbers on Tier 2: up to 100,000 H100-equivalent chips in 2025, 270,000 in 2026, and 320,000 in 2027. The 2027 cap used up only 0.2 GW of power, against Goldman Sachs’s projection of 68 GW in additional global AI data-center demand by 2027. That is a 340x gap. Tier 3 covered China, Russia, Iran, North Korea, Burma, Syria, and Venezuela, all with US arms embargoes, locked out of advanced American AI chips.

The rule had required US-headquartered companies to keep at least 50 percent of total AI compute in the United States, 75 percent in Tier 1, and no more than 7 percent in any single Tier 2 country. SemiAnalysis projected more than ten gigawatt-scale campuses coming online in the United States over the next two years, with roughly four times the planned data-center capacity of either Europe and the Middle East or Asia Pacific through 2027. Even so, the chokepoint of advanced US chips remains intact, and the compute gap between the US and its rivals has not narrowed.

Winners: American Chipmakers and Their Treasury

Nvidia’s valuation crossed five trillion dollars in 2025, the first company to reach that milestone according to the Atlantic Council’s tally of last year’s records. The December reversal opens a Chinese market that had been operating under the Biden framework’s de facto ban. The Treasury takes a 15 percent cut on most chips and 25 percent on advanced ones, turning what had been a security perimeter into a revenue stream. Investors are pricing US firms as the global default, even as the rule of law around their access is going transactional. The same chips that doubled in price on China’s black market are now flowing through Treasury-cut official channels.

Hyperscaler AI spending underwrites the bet. Trump announced Stargate at the start of 2025 with the stated aim of investing $500 billion in AI infrastructure over five years. Inside Washington, the policy is held together by a network of White House AI adviser David Sacks, the commerce department’s licensing machinery, and a draft Chip Security Act introduced in May 2025 in both chambers of Congress. The bill mandates new on-chip verification, including so-called quantum-dot identifiers and telemetry that allows remote verification and, in some cases, remote disabling of deployed chips.

Losers: The Tier-2 Middle Powers

The Biden framework had been designed to keep Tier-2 nations at least a generation behind the AI frontier. Its revocation does not necessarily improve their position; it removes the standing rules that governed their access and replaces them with a case-by-case bargaining process. The CSIS analysis warns that a bloc of Tier-2 nations may collectively chart a “Third Way” between Washington and Beijing, carving out a mid-tier AI stack that offers strategic autonomy without open confrontation. The default CSIS projects is that Tier 2 nations either climb into Tier 1 or drift into China’s orbit, with the disruptive third option of a coordinated non-aligned AI stack. Poland, the Baltic states, Israel, India, and the Gulf monarchies expressed unhappiness or felt slighted at being placed in Tier 2.

The Trump administration’s preference for government-to-government deals means that each of them now negotiates separately, with compute access contingent on bilateral terms.

Some will lose more than others. Malaysia was on track to become the world’s third-largest data-center country by 2026, CSIS records, and India, Indonesia, Saudi Arabia, Singapore, and the UAE all staked their AI plans on expanded compute access that the Diffusion Rule would have rationed. China’s own response has been to position itself, through DeepSeek and a fast-growing set of open-source models, as the alternative patron. For proof of how the race is already tilting on cost, Chinese open-weight models now cost a fraction of US frontier tokens per call. The wider pattern is a move from the logic of globalization to zero-sum thinking, in which middle powers must choose a side or build their own.

• Poland, Latvia, Estonia, Lithuania: excluded from Tier 1 despite NATO membership
• Israel: advanced AI sector but outside Tier 1
• India: legacy of nonalignment, large compute needs
• Saudi Arabia, UAE: large AI investors but outside intelligence-sharing orbit
• Malaysia: third-largest planned data-center capacity by 2026
• Singapore, Indonesia, Mexico, Philippines, Switzerland, Turkiye: all placed in Tier 2

China’s Open-Source Play and the Diffusion Counter-Bet

The Atlantic Council’s Ryan Pan expects China to double down on open-source AI strategy in 2026, the same lever that has already drawn major US tech companies to Chinese large language models. DeepSeek published a new AI training paper in early 2026, almost exactly a year after its headline-making January 2025 release that was followed by the DeepSeek-R1 launch. China’s bet is that free, deployment-ready models will capture global market share faster than controlled access can lock it down.

The Diffusion Rule inadvertently accelerated that bet. The CSIS analysis of the three-tier framework notes that the framework’s controls on model weights trained on US-controlled hardware create an incentive for Tier 2 nations to migrate to open source. Beijing has obliged with DeepSeek and Huawei’s Ascend series chips, the latter positioned as an Nvidia substitute inside China. In May, the commerce department notified private firms that using Huawei’s Ascend 900-series chips and other Chinese-made chips would likely constitute a violation of US export controls.

Howard Lutnick’s addiction framing was meant to lock Chinese developers into US toolchains. Beijing has read it the same way. After the Trump pivot, China moved to discourage large-scale H20 purchases, arguing that reliance on US-origin chips would create long-term strategic vulnerabilities.

Both sides are now running a bet on speed: the US on diffusion outrunning indigenisation, China on indigenisation outrunning diffusion. On the test bench for that wager sits Huawei’s chip superclusters, with the December 2025 pivot on AI diffusion controls noting claims that the systems outperform Nvidia’s AI chips on some measures, a claim treated as a signal rather than a confirmed benchmark. Beijing has chosen to treat compute self-reliance as a national-security priority. The company is pushing turnkey AI systems to emerging markets.

The Labor Side of the Same Coin

The compute arms race has a parallel race on the labor side, with both running on the same AI stack. The November 2025 analysis of work hours that could be automated estimates that today’s technology could, in theory, automate activities accounting for about 57 percent of US work hours. Demand for AI fluency, the ability to use and manage AI tools, has jumped nearly sevenfold in two years, the fastest of any skill in US job postings. More than 70 percent of the skills employers seek today are used in both automatable and non-automatable work.

McKinsey sizes the upside at about $2.9 trillion in US economic value by 2030, in its midpoint scenario, if organizations prepare their people and redesign workflows around people, agents, and robots working together. The AI fluency gap, eight million US workers in occupations already calling for at least one AI skill, is a fraction of what may be needed. McKinsey is explicit that the 57 percent figure is a technical potential, not a forecast of job losses.

Compute and labor are both governed by who controls the underlying platform. On a country scale, the same trade plays out: nations that wait for Tier 1 access or settle for Tier 2 rationing risk the same stranded-asset problem as firms that delay adoption. The planned expiry of the 50 percent rule waiver and the advanced chip export allowance, both in late 2026, will force Washington to decide whether these concessions become precedent or remain exceptional measures. For Europe, the test runs through its own rule book: EU AI Act fines are reshaping Europe’s AI compliance posture on the same horizon. The wider pattern is a move from the logic of globalization to zero-sum thinking.

• 57 percent of US work hours automatable in theory (McKinsey, November 25, 2025)
• 7x growth in demand for AI fluency in two years
• $2.9 trillion potential US economic value by 2030 (McKinsey midpoint scenario)
• $500 billion Stargate AI infrastructure plan over five years
• 70%+ of employer-sought skills apply to both automatable and non-automatable work

Frequently Asked Questions

What was the AI Diffusion Rule?

A January 13, 2025 Commerce Department rule that split the world into three tiers for advanced AI chip access. Tier 1 included 18 allies with near-frictionless access, Tier 2 was capped at a generation behind, and Tier 3 (China, Russia, Iran, North Korea, Burma, Syria, Venezuela) was locked out. The Trump administration revoked it a week before its May 15, 2025 enforcement date.

What changed in December 2025?

On December 8, 2025, the Trump administration reversed export controls on Nvidia’s H200 chip, a full-performance Hopper chip nearly six times more powerful than the downgraded H20. The window is open for one year. In exchange, the US Treasury takes 15 percent of most chip revenues and 25 percent of advanced chip revenues from Chinese sales.

Why did the US reverse course on AI chip exports?

Commerce Secretary Howard Lutnick framed the approach in July 2025: US firms should “sell the Chinese enough that their developers get addicted to the American technology stack.” The strategy reflects the administration’s view that architectures and training methods diffuse too quickly to fence off, and that locking in global users, developers, and data flows to a US-centered stack is the goal.

Which countries are most affected by the new bargaining system?

Tier 2 nations, including India, Israel, Singapore, Switzerland, Saudi Arabia, the UAE, Mexico, and the Philippines, lost the standing rules that governed their chip access and must now negotiate bilaterally. Poland, Latvia, Estonia, and Lithuania were excluded from Tier 1 despite their NATO membership. CSIS records the omission as reflecting enforcement capacity and diversion risk alongside alliance status.

How big is the AI compute gap between the US and China?

China’s most advanced chip is four years behind the US and its total compute capacity is about ten times smaller, according to the Steptoe analysis. The Center for a New American Security estimated up to 12,500 advanced chips could be smuggled into China annually. China’s DeepSeek V3 and R1 models have demonstrated near-parity in AI engineering despite the gap.