The H200 Paradox: Washington Opened China's Biggest AI Chip Deal, Then Congress Moved to Kill It

Feb 22, 2026 Written By Blake Crosley

Four hundred thousand NVIDIA H200 GPUs sit approved but unshipped, frozen between a Commerce Department that greenlit the sale and a State Department that blocked the cargo.¹² On January 15, 2026, the Bureau of Industry and Security published Federal Register rule 2026-00789, shifting export review for H200 and AMD MI325X chips to China from "presumption of denial" to "case-by-case review," a 13x increase in permitted computing power over previously allowed exports.³⁴ Six days later, the House Foreign Affairs Committee voted 42-2 to advance the AI OVERWATCH Act, a bill that would ban Blackwell-class chip sales to China for two years and grant Congress veto power over every future AI chip export license.⁵⁶ The collision between executive action and legislative backlash has created the most consequential AI trade policy crisis since the original October 2022 semiconductor controls, and infrastructure teams worldwide now face a planning environment where the rules change faster than procurement cycles.

TL;DR

The Bureau of Industry and Security loosened H200 export restrictions to China on January 15, 2026, moving from presumption of denial to case-by-case review with a 25% Section 232 tariff. Chinese firms placed orders for over 2 million H200 chips, but the State Department stalled shipments pending security review. Congress responded with the AI OVERWATCH Act (H.R. 6875), which passed committee 42-2 and would impose a statutory two-year ban on Blackwell exports while granting congressional veto power over future licenses. Think tanks warn the export cap alone could increase China's domestic AI compute by 250%, and Huawei plans to ship 600,000 Ascend 910C chips in 2026 regardless of what Washington decides.

Section 1: The BIS Rule That Changed Everything

From Denial to Case-by-Case

On January 13, 2026, BIS published a final rule revising license review policy under 15 CFR Parts 742, 744, and 748.³ The rule took effect January 15 and fundamentally altered the U.S. approach to advanced computing exports to the People's Republic of China and Macau.⁴

The core shift: any commercially available chip meeting defined performance thresholds below 21,000 total processing performance and 6,500 GB/s DRAM bandwidth qualifies for case-by-case review rather than automatic denial.⁷ The NVIDIA H200 and AMD MI325X both fall under the new framework.⁸ Blackwell-class chips, including the NVIDIA B200, GB200, and GB300, remain under presumption of denial.⁴

The distinction matters enormously. Only direct exports from the United States qualify for the relaxed review; reexports, exports from abroad, and transfers of H200/MI325X to or within China still face presumption of denial.⁴⁹ Exporters must satisfy four certification requirements to obtain licenses:

Sources: Federal Register 2026-00789³; Morgan Lewis analysis¹⁰; Covington & Burling analysis¹¹

The Section 232 Tariff

A Presidential Proclamation issued January 14, 2026, imposed a 25% tariff on advanced semiconductor articles under Section 232 of the Trade Expansion Act of 1962.¹²¹³ The tariff covers H200, MI325X, semiconductor manufacturing equipment, and derivative products containing covered semiconductors.¹⁴

The mechanism works through a constitutional workaround: third-party U.S. testing facilities required by the BIS rule also serve as import-for-testing waypoints, enabling tariff collection as an "import" fee before re-export.¹⁴ Analysts project the tariff will push H200 unit pricing in China above $35,000.¹⁵

Exemptions cover U.S. data centers, research and development, startups, repairs, non-data-center consumer and industrial applications, and public sector uses.¹² The net effect targets a specific flow: American-made AI chips heading to Chinese commercial buyers.

At $14 billion in pending Chinese orders, the tariff could generate roughly $3.5 billion in U.S. Treasury revenue.¹⁶¹⁷

The Taiwan Deal

Simultaneous with the export rule and tariff, the U.S. and Taiwan reached a trade agreement granting Taiwanese chipmakers that expand U.S. production a reduced semiconductor tariff in exchange for $250 billion in American manufacturing investment, including new fabrication facilities.¹⁸ The deal locks Taiwan's semiconductor ecosystem more tightly to the U.S. market while TSMC's CEO publicly acknowledged that advanced-node capacity remains "about three times short" of global demand.¹⁹

Section 2: The AI OVERWATCH Act Strikes Back

Bill Structure and Vote

Chairman Brian Mast (R-FL) introduced H.R. 6875, the AI OVERWATCH Act, in December 2025.⁵ The House Foreign Affairs Committee advanced the bill on January 21, 2026, with a 42-2-1 vote signaling overwhelming bipartisan support.⁶²⁰

Mast framed the stakes bluntly: "Companies like Nvidia are requesting to sell millions of advanced AI chips, which are the cutting edge of warfare, to Chinese military companies like Alibaba and Tencent."²¹ His follow-up left even less room for interpretation: "This bill is very simple. It keeps America's advanced AI chips out of the hands of Chinese commie spies."²²

Key Provisions

The AI OVERWATCH Act contains six major provisions that would reshape the export control landscape:⁵⁶²³

1. Blackwell chip ban. An outright prohibition on NVIDIA Blackwell-class chip sales to China for a minimum of two years, codifying the existing BIS presumption of denial into statute.⁵

2. Congressional review authority. Congress gains power to block specific export licenses through a joint resolution, modeled on the 1976 foreign arms sale review protocols.⁶

3. License revocation. The bill revokes existing licenses for AI chip transfers to designated adversaries.²³

4. Temporary blanket denial. A blanket denial period takes effect until the administration submits a national security strategy on AI exports.⁵

5. 30-day review window. Any future approvals face a 30-day Congressional review period before taking effect.⁶

6. Adversary coverage. Restrictions apply to China, Russia, Iran, North Korea, and Cuba.²³

Bipartisan Coalition

Fifteen co-sponsors back the bill, including China Select Committee Chairman John Moolenaar, Intelligence Committee Chairman Rick Crawford, and NSA and Cyber Subcommittee Chairman Darin LaHood.²⁴ Supporting organizations range from the Foundation for Defense of Democracies Action to Americans for Responsible Innovation and the American Security Fund.²⁴

The breadth of support underscores a rare point of Congressional unity. Both hawks who want tighter China restrictions and moderates concerned about executive overreach found common ground in demanding legislative oversight of AI chip exports.²⁰

Executive vs. Legislative Collision

The tension between the BIS rule and the OVERWATCH Act creates a direct constitutional confrontation. The executive branch opened a pathway for H200 sales; the legislative branch moved to close it and lock the door. Mast drew the parallel explicitly: "If we are selling another country anything that could give them an advantage on the battlefield, it requires congressional notification. This bill creates the same congressional oversight for advanced AI chips that are being requested for sale to China, Russia, Iran, North Korea, Cuba... Because AI chips are used by our enemies' militaries just as much as they are used by our military."²²

Section 3: The Numbers That Alarmed Washington

Demand Explosion

Chinese firms ordered over 2 million H200 chips following the December 2025 announcement that H200/MI325X sales would proceed with a 25% fee.²⁵²⁶ NVIDIA currently holds approximately 700,000 units available and asked TSMC to ramp production for 2 million H200s across 2026.²⁶ China granted conditional approval for 400,000 GPUs to ByteDance, Alibaba, and Tencent, representing less than one-fifth of total orders.¹²⁷

The BIS rule caps exports at approximately 850,000 H200-equivalent chips, limited to 50% of cumulative U.S. sales of roughly 1.7 million units.²⁸ Combined with AMD MI325X allocations, the cap reaches nearly 900,000 H200-equivalents.²⁸

Compute Impact Assessment

Think tank analyses paint a stark picture of what those numbers mean for global AI competition:

Sources: CFR expert briefs²⁹³⁰; CNAS commentary²⁸

Chris McGuire, Senior Fellow for China and Emerging Technologies at the Council on Foreign Relations, called the overall policy "strategically incoherent and unenforceable."²⁹ His recommended alternative: "denying the export of all AI chips to China."²⁹

The Military-Civil Fusion Problem

The enforcement challenge compounds the volume concern. Nearly 75% of entities winning multiple PLA AI contracts operate as civilian firms with no self-reported state ownership.²⁸ Chinese law requires commercial firms to assist military organizations when requested, directly conflicting with the export certification requirements that underpin the BIS rule.²⁸³¹

Tencent, one of the three companies approved for the initial 400,000 GPU batch, carries an official Department of Defense designation as a "Chinese Military Company."²⁹ The Council on Foreign Relations noted that major Chinese purchasers including Alibaba, Tencent, and DeepSeek maintain "documented relationships with China's military and security services."²⁹

Precedent supports enforcement skepticism. Huawei used Sophgo as a front company to procure nearly 3 million TSMC chips, demonstrating how shell companies redirect semiconductor shipments after initial entry.²⁸ Cloud access creates another vulnerability: prohibited end users can access exported compute remotely despite physical chip certifications.²⁸

Michael C. Horowitz warned directly: "China's military is actively seeking Nvidia chips to fuel AI-enabled military capabilities."³¹ Both DeepSeek CEO Liang Wenfeng and China's Premier Li Qiang have cited U.S. export controls as "China's biggest constraint in AI development," underscoring precisely why those controls carry strategic weight.³⁰

Section 4: NVIDIA's Financial Tightrope and China's Alternatives

The Revenue Paradox

NVIDIA faces a financial equation with no clean answer. The company logged a 69% year-over-year revenue jump in Q1 FY2026 despite export controls.³² Simultaneously, NVIDIA absorbed a $4.5 billion Q1 charge from H20 excess inventory and purchase obligations triggered by the April 2025 H20 ban.³³³⁴ The Q2 FY2026 outlook carries an $8.0 billion revenue loss from H20 export limitations.³⁴

NVIDIA estimates the China AI market will grow to nearly $50 billion annually.³⁴ ByteDance-led orders for 2026 total upward of $14 billion.¹⁷ CEO Jensen Huang personally secured the December 2025 agreement with President Trump to re-enter the China market, staking significant corporate and political capital on H200 sales proceeding.³⁴

The State Department review, however, has stalled all shipments. Commerce completed its analysis, but State pushed for tougher restrictions including mandatory third-party testing, detailed end-use reporting, requirements that half of shipments serve U.S. customers, and measures preventing Chinese military and intelligence access to advanced AI computing power.²³⁵ AMD MI325X approvals face identical delays, indicating an industry-wide standstill.³⁵

Huawei's Parallel Track

China has not waited for Washington to resolve its internal disputes. Huawei targets 600,000 Ascend 910C chips in 2026, doubling its 2025 output, with total Ascend production reaching up to 1.6 million dies.³⁶³⁷³⁸ Each 910C delivers roughly one-third the BF16 throughput of NVIDIA's B200, and Huawei's strategy relies on horizontal scaling rather than per-chip performance parity.³⁶

Sources: Tom's Hardware³⁶; Bloomberg³⁷; Huawei Central³⁸; SemiAnalysis³⁹

Manufacturing constraints remain real. SMIC produces on 7nm-class DUV lithography without access to EUV equipment, and yields trail TSMC and Samsung.³⁶³⁹ Reports of smuggled dies and Samsung HBM memory in Ascend 910C units raise questions about Huawei's true supply chain independence.³⁶

The CFR assessed the situation clearly: "AI chips are uniquely hard for China to produce" given domestic manufacturing constraints.²⁹ Yet the same analysis acknowledged that large numbers of lower-performing chips can aggregate into world-class AI data centers, and China matches or surpasses the U.S. in data, talent, algorithms, applications, and electricity.³¹

China's Retaliatory Leverage

Beijing holds a significant countermeasure. In October 2025, China announced licensing requirements on rare-earth oxides, metals, and magnet products, applying its own version of the foreign direct product rule to regulate foreign-made products incorporating Chinese rare earth materials.⁴⁰⁴¹ Rare earths underpin critical defense systems including F-35 fighters, Virginia-class submarines, Tomahawk missiles, radar systems, and Predator UAVs.⁴¹

China suspended these controls for one year until November 2026 following a Trump-Xi meeting, maintaining leverage for the April 2026 summit while keeping the threat active.⁴² Companies not diversified from Chinese rare earth sources face production slowdowns, higher costs, and product launch delays if restrictions resume.⁴⁰

Section 5: The Historical Whiplash

The current policy collision represents a 180-degree reversal from both Trump's first term and the Biden-era containment framework.⁴³

Policy Evolution Timeline

Sources: Congress.gov CRS Report⁴³; BIS press releases⁴⁴; Gibson Dunn analysis⁴⁵

The policy simultaneously acknowledges that H200 exports pose national security risks through certification requirements, creates a pathway to permit the sales through case-by-case review, and attempts to extract revenue from the transaction through the 25% tariff.⁴⁵ Congress, having watched the whiplash from containment to commerce and back, decided to assert statutory authority over a domain the executive branch has controlled since 2022.

CFR's assessment resonates with added force against the historical backdrop: the policy tries to treat AI chips as both a strategic weapon and a commercial product, satisfying neither the national security community nor the commercial interests it purports to serve.²⁹

The Xi 2027 Timeline

One data point looms over every policy discussion. Chinese President Xi Jinping reportedly directed military readiness for a potential Taiwan invasion by 2027.³⁰ Every H200 shipped to China between now and that date adds to the computing power available for military AI applications. The CFR explicitly connected export policy to the Taiwan timeline, noting that "export controls on AI chips are the only U.S. policy capable of slowing China's AI progress" during the critical window.²⁹

Section 6: Infrastructure Implications for AI Deployment Teams

Supply Chain Uncertainty

The 850,000 H200-equivalent export cap, combined with SK Hynix controlling roughly 60% of the global HBM market (sold out through 2026) and Micron fulfilling only 50-66% of core customer demand, means every GPU allocation carries geopolitical weight.¹⁹²⁸ Infrastructure teams face a procurement environment where delivery timelines depend on diplomatic outcomes rather than manufacturing schedules.

If Congress passes the AI OVERWATCH Act and the 30-day review window takes effect, every large export license will face potential legislative challenge.⁶ Procurement teams need contingency plans for scenarios where approved orders face mid-cycle review or revocation.

Domestic Capacity Acceleration

The Taiwan deal's $250 billion investment commitment signals a structural shift toward U.S.-based semiconductor manufacturing.¹⁸ Combined with existing CHIPS Act investments, domestic fab capacity will expand significantly over the next three to five years. Infrastructure planners should map procurement strategies to the expanding domestic supply base rather than assuming current allocation patterns persist.

The organizations deploying AI infrastructure at scale need partners who understand both the technical demands and the shifting policy landscape. Introl's 550 HPC-specialized field engineers operate across 257 locations globally, delivering GPU deployments of up to 100,000 units. Ranked #14 on the Inc. 5000 with 9,594% three-year revenue growth, Introl provides the deployment expertise that translates available GPU allocations into operational compute capacity regardless of which policy scenario materializes.

Planning for Multiple Scenarios

Infrastructure leaders should model at least three scenarios for 2026-2027 GPU procurement:

Scenario A: BIS rule holds, OVERWATCH stalls. H200 exports to China resume at scale after State Department review concludes. Global supply tightens as 850,000+ units flow east. Domestic procurement faces allocation competition. Plan for longer lead times and premium pricing.

Scenario B: OVERWATCH becomes law. Blackwell exports face a statutory two-year ban. H200 exports require 30-day Congressional review. Chinese demand redirects to Huawei Ascend and potentially gray-market channels. Domestic supply improves as China allocation frees up. Plan for greater availability but continued HBM constraints.

Scenario C: Policy oscillation continues. Export rules shift quarter to quarter based on diplomatic momentum. The April Trump-Xi summit, China's suspended rare earth controls (expiring November 2026), and midterm election dynamics create multiple inflection points. Plan for maximum flexibility with modular procurement contracts.

Key Takeaways

For Infrastructure Planners

• Model GPU procurement around at least three policy scenarios through 2027
• Map supply chain exposure to HBM shortages (SK Hynix 60% market share, sold out through 2026) and TSMC capacity constraints (3x short of demand)
• Evaluate domestic manufacturing timelines from CHIPS Act and Taiwan deal investments for medium-term planning
• Build relationships with deployment partners who maintain allocation access across policy shifts

For Operations Teams

• Track State Department review outcomes as leading indicators for GPU availability shifts
• Prepare for H200 pricing volatility as the 25% tariff, demand surges, and policy uncertainty interact
• Assess existing infrastructure for Blackwell-class readiness given the separate restriction tier
• Plan liquid cooling and power upgrades around confirmed rather than projected GPU deliveries

For Strategic Decision-Makers

• The AI OVERWATCH Act's 42-2 vote signals durable bipartisan support for export restrictions regardless of executive branch direction
• China's 600,000 Ascend 910C production target in 2026 means the competitive compute gap narrows whether or not H200 exports proceed
• Rare earth export controls, suspended until November 2026, represent Beijing's most direct countermeasure against U.S. semiconductor policy
• The Xi 2027 military readiness timeline drives urgency in both U.S. restriction efforts and Chinese domestic chip development

Looking Ahead

The H200 export saga will not resolve cleanly. The AI OVERWATCH Act must clear the full House and Senate before reaching the President's desk. The State Department review carries no published timeline. China's conditional approval for 400,000 GPUs has not converted to actual shipments, and Chinese regulators reportedly find the BIS license conditions "too restrictive" for purchases to proceed.⁴⁶

Meanwhile, the chips keep getting more powerful. NVIDIA's Blackwell generation already sits under presumption of denial. The CFR warned that if the case-by-case framework ever expanded to Blackwell, permitting 50% of all GB300 chips sold to the U.S. could greenlight 2.5 million GB300s to China.²⁹ The OVERWATCH Act exists precisely to prevent that escalation.

For infrastructure teams, the signal through the noise remains clear: build for flexibility, secure supply through diversified partnerships, and assume that AI chip trade policy will generate surprises faster than anyone in Washington or Beijing can predict.

References