Nine hundred thousand chips. That's how many H200-equivalent processors China could legally acquire under the Bureau of Industry and Security's revised export policy—a volume representing twice the country's entire domestic AI chip production capacity for 2026.
The policy shift, effective January 15, 2026, replaces a presumption of denial with case-by-case licensing for NVIDIA's H200 and AMD's MI325X accelerators destined for China and Macau. The change follows President Trump's December 8, 2025 directive permitting exports to "approved customers" in exchange for a 25% federal surcharge on every chip sold.
For data center operators and enterprise AI buyers, the implications extend beyond geopolitics. The policy introduces new supply chain dynamics, testing requirements, and volume constraints that will shape GPU availability through 2027. Understanding these mechanics matters whether you're planning infrastructure expansion or evaluating competitive positioning in the global AI race.
The Technical Framework: What's Actually Allowed
The revised policy targets a specific class of chips defined by two technical thresholds established in the Export Administration Regulations. Chips qualify for case-by-case review if they fall below both limits:
- Total Processing Performance (TPP): Less than 21,000
- Total DRAM Bandwidth: Less than 6,500 GB/s
NVIDIA's H200 specifications place it well within these bounds. The chip delivers 989.5 TFLOPS of FP16 throughput, translating to a TPP score of 15,832. Its 141 GB of HBM3e memory provides 4.8 TB/s bandwidth—both metrics safely below the thresholds.
AMD's MI325X cuts closer to the limits. The chip offers 1,300 TFLOPS of FP16 performance (TPP score of 20,800) with 256 GB of HBM3E and 6 TB/s bandwidth. Both figures remain compliant, but the tighter margins explain why BIS specifically named this chip in its guidance.
The thresholds matter for a practical reason: they exclude next-generation hardware. NVIDIA's Blackwell architecture and AMD's MI400 series exceed these limits, meaning the policy permits exports of current-generation training accelerators while maintaining restrictions on frontier hardware.
Four Requirements That Define "Case-by-Case"
BIS didn't simply remove restrictions—it replaced blanket denial with conditional approval. License applicants must satisfy four requirements before Commerce will approve exports:
1. U.S. Supply Priority
Exporters must demonstrate that sales to China "will not reduce global semiconductor production capacity currently available to U.S. customers." In practice, this means proving:
- No U.S. orders are delayed
- Domestic demand is fully met
- Advanced-node foundry capacity serving U.S. customers is not diverted
- Aggregate China-bound shipments don't exceed 50% of the same product shipped to the United States
This last point—the 50% volume cap—establishes the hard ceiling. NVIDIA has sold approximately 2 million H200 chips to U.S. customers cumulatively. Assuming 85% went to domestic buyers, China could purchase up to 850,000 units before hitting the threshold.
2. Purchaser Compliance Programs
Chinese buyers must adopt export compliance procedures including customer screening. The requirement sounds straightforward, but enforcement poses challenges that critics have highlighted. Companies like Alibaba and Tencent—likely first in line for H200 purchases—maintain documented connections to Chinese military projects through China's Military-Civil Fusion strategy.
3. Independent U.S. Testing
Every shipment must undergo verification in an independent, U.S.-headquartered testing laboratory with no financial ties to the exporter or importer. Testing covers four parameters:
- Total Processing Performance
- Memory bandwidth
- Interconnect bandwidth
- Co-packaged DRAM capacity
The requirement adds cost and time to each transaction while creating a bottleneck: limited testing capacity could constrain export volumes even below the 50% cap.
4. Certification Requirements
Exporters must certify compliance with standards that critics describe as inherently unfalsifiable. Proving that a shipment causes "no delay" to U.S. orders requires demonstrating a negative across complex supply chains. The Council on Foreign Relations analysis argues that semiconductor supply chains are sufficiently inelastic that any export necessarily affects domestic availability.
The 25% Surcharge: Revenue Mechanics
President Trump's December announcement included a provision that generated as much attention as the policy reversal itself: a 25% federal surcharge on all H200 and similar chip exports to China.
The surcharge creates interesting economics. An H200 currently sells for approximately $25,000-$30,000 depending on configuration and volume. A 25% surcharge adds $6,250-$7,500 per unit, potentially generating billions in federal revenue.
Chinese buyers have already placed orders for over 2 million H200 units for 2026—far exceeding both NVIDIA's current inventory of 700,000 units and the policy's volume caps. The demand signal suggests Chinese customers view the surcharge as acceptable given the alternative: continued reliance on domestic chips that lag substantially in performance.
The surcharge also closes a routing loophole. The policy explicitly limits licensing to exports from the United States. Chips fabricated outside the U.S. must first be imported (and pay the 25% duty) before they can be exported to China under a license. This prevents companies from shipping directly from TSMC's Taiwan fabs or Samsung's Korean facilities.
Compute Implications: The Numbers That Matter
The Center for a New American Security published analysis quantifying what permitted exports mean in practical compute terms.
850,000 H200 chips represent:
-
Twice China's 2026 domestic production capacity. Analysts estimate Chinese fabs will produce roughly 390,000 H200-equivalent chips domestically this year. Export permissions effectively triple available supply.
-
Sufficient compute to train frontier models. CNAS analysts conclude this volume provides "sufficient compute to train models matching or exceeding current American frontier models." The aggregate capacity approaches OpenAI's entire deployed compute as of late 2025.
-
More than twice the world's largest data center. Meta's Prometheus facility in Ohio—the first 1+ GW data center—doesn't approach this GPU concentration.
The comparison to domestic production capacity carries particular weight. Huawei plans to produce approximately 600,000 Ascend 910C chips in 2026, roughly double its 2025 output. But Huawei faces severe constraints that limit how many chips actually reach customers:
-
HBM bottleneck: China's domestic HBM production can support only 275,000-300,000 Ascend chips in 2026. SK Hynix controls 60% of global supply and remains sold out through 2026.
-
Yield challenges: Estimates put Chinese yields for Ascend chips between 5% and 20%, compared to 60-80% for NVIDIA Blackwell at TSMC.
-
Performance gaps: Huawei's own roadmap shows the Ascend 950PR and 950DT planned for 2026 have lower TPP scores than the current 910C—a regression rather than advancement.
Permitted H200 exports would overwhelm these domestic limitations, fundamentally changing China's AI compute trajectory.
Congressional Response: The AI Overwatch Act
The House Foreign Affairs Committee responded to the policy shift by advancing H.R. 6875, the AI Overwatch Act, on January 21, 2026. The committee voted 42-2-1 to advance the measure.
The bill would require Commerce Department approval for AI chip exports exceeding certain performance criteria to China, Cuba, Iran, North Korea, Russia, and Venezuela. More significantly, it gives the House Foreign Affairs Committee and Senate Banking Committee at least 30 days to review and potentially block approved exports—a process modeled after congressional oversight of foreign arms sales established in 1976.
Chairman Brian Mast (R-FL) framed the stakes directly: "Companies like Nvidia are requesting to sell millions of advanced AI chips...to Chinese military companies like Alibaba and Tencent...This bill keeps America's advanced AI chips out of the hands of Chinese commie spies."
The bill has 15 cosponsors including House China Select Committee Chair John Moolenaar and House Intelligence Committee Chair Rick Crawford. Supporting organizations include the Foundation for Defense of Democracies Action, American Compass, and Americans for Responsible Innovation.
Bipartisan opposition to the export policy extends beyond Congress. Anthropic CEO Dario Amodei stated at Davos: "It would be a big mistake to ship these chips. I think this is crazy. It's a bit like selling nuclear weapons to North Korea."
White House AI and Crypto Czar David Sacks has criticized the AI Overwatch Act, arguing it undermines executive authority over export policy. The bill faces a long path: passage in the full House, Senate approval, and either presidential signature or veto override.
Enforcement Questions
Critics from across the political spectrum have identified enforcement gaps that could render the policy's restrictions meaningless.
End-user verification challenges: China's Military-Civil Fusion strategy eliminates meaningful distinctions between private sector and military entities. A chip certified for commercial use at Alibaba Cloud can be redirected to defense applications without violating any agreement the original purchaser signed.
Cloud access loopholes: Remote access to H200 compute via cloud services remains virtually impossible to prevent. Chinese researchers can access U.S.-hosted H200 instances, train models, and export weights without any chip physically entering China.
Smuggling precedent: Rep. Gregory Meeks (D-NY), top Democrat on the House Foreign Affairs Committee, cited a recent Justice Department action breaking up a $160 million chip smuggling network as evidence that diversion risks are not theoretical.
The Council on Foreign Relations concluded that the regulation is "strategically incoherent" because it simultaneously acknowledges national security risks while creating pathways that those risks materialize. Faithful implementation would block most exports; loose implementation ignores the original concerns.
NVIDIA's Position: Revenue and Reality
NVIDIA's financial exposure to China explains the company's support for the policy shift. The company lost $4.5 billion in Q1 FY2026 due to excess inventory and purchase obligations from the H20 export ban. Q2 guidance reflects an additional $8 billion loss from continued H20 restrictions.
Analysts expect the H200 policy to partially offset these losses. Orders for over 2 million H200 units have already been placed by Chinese customers for 2026, though volume caps limit actual shipments. Analysts project FY26 revenue of $206.6 billion, with a $275 billion backlog extending into 2027.
But the H200 policy doesn't reverse NVIDIA's broader China decline. Analysts from Bernstein expect NVIDIA's share of China's AI processor market to drop from 66% to approximately 8% as Huawei, Cambricon, and other domestic vendors approach 80% combined share.
What This Means for Data Center Planning
For enterprise AI buyers and data center operators, the policy creates several planning considerations:
Supply chain dynamics: The 50% volume cap means U.S. customers have priority, but increased total demand from China could extend lead times. NVIDIA's current inventory of 700,000 H200s falls far short of the 2+ million units ordered by Chinese customers alone.
Testing bottlenecks: Independent testing requirements for China-bound chips may create capacity constraints that affect overall production scheduling.
Pricing pressure: The 25% surcharge doesn't apply to U.S. customers, but increased China demand could support higher pricing across all markets. Cloud H200 rates currently range from $1.50-$4.31/hour depending on provider.
Competitive positioning: Chinese hyperscalers gaining access to H200 chips accelerates their AI capabilities. For global enterprises competing with Chinese companies, the compute gap narrows.
The Larger Context
The H200 policy represents one move in a complex semiconductor chess game. The same week BIS revised export rules, it also:
- Approved annual equipment authorizations for TSMC, Samsung, and SK Hynix to continue importing U.S. chipmaking equipment to their China fabs
- Maintained Blackwell restrictions, keeping next-generation chips beyond China's reach
- Preserved Entity List restrictions on specific Chinese companies
The policy attempts to thread a needle: permitting current-generation exports that generate revenue and satisfy industry pressure while maintaining restrictions that preserve U.S. technological leadership. Whether this balance holds depends on enforcement effectiveness and congressional action.
The AI Overwatch Act's 42-2 committee vote suggests broad bipartisan concern that the balance tips too far toward commercial interests. The bill's path through the full House and Senate will test whether those concerns translate to legislative action.
For now, the policy stands. Nine hundred thousand chips—twice China's domestic capacity—can legally flow across the Pacific. The question is whether that flow strengthens American companies or accelerates a competitor that views AI supremacy as a national security imperative.
The global competition for AI infrastructure continues intensifying. For analysis of how data center operators are navigating these dynamics, explore Introl's coverage of GPU deployment, power infrastructure, and enterprise AI strategy.
