BIS Export Policy Shift: H200 and MI325X Move to Case-by-Case Review for China

TL;DR

The Bureau of Industry and Security (BIS) published a final rule on January 15, 2026 shifting export license review for NVIDIA H200 and AMD MI325X chips from presumption of denial to case-by-case review for China—subject to a 25% tariff, 50% volume cap, third-party testing, and strict KYC requirements. Separately, Samsung and SK hynix received annual licenses replacing expired Validated End User (VEU) status, introducing annual uncertainty for chipmaking tool shipments. The House passed the Remote Access Security Act (369-22) to close the cloud GPU rental loophole. These policy shifts create a complex compliance environment while data suggests export controls have widened—not narrowed—the US-China AI compute gap.

Presumption of denial.

For three years, those three words defined US policy on advanced AI chip exports to China. Any license application for high-performance semiconductors faced automatic rejection unless applicants could demonstrate exceptional circumstances. The policy blocked billions in potential sales while forcing Chinese AI companies to scramble for alternatives.

On January 15, 2026, that changed.

The Bureau of Industry and Security published a final rule shifting review policy for certain advanced computing semiconductors—specifically NVIDIA's H200 and AMD's MI325X—from presumption of denial to case-by-case review. The policy applies only to exports from the United States to China and Macau, and comes loaded with conditions designed to protect domestic supply chains.

For data center operators, GPU customers, and AI infrastructure planners, the shift creates new complexity. Some chips may now flow to China under specific conditions, while cloud-based access faces tightening restrictions through separate legislation. Understanding what's actually changed—and what hasn't—requires unpacking multiple interlocking policies.

The January 2026 Rule Change

BIS's final rule establishes specific thresholds for case-by-case review eligibility: chips with Total Processing Performance (TPP) less than 21,000 and total DRAM bandwidth less than 6,500 GB/s qualify. This includes the NVIDIA H200 and AMD MI325X specifically named in the rule.

To qualify for case-by-case review, license applicants must demonstrate four conditions:

1. US Supply Protection: Exporting to China won't reduce global semiconductor production capacity currently available to US customers
2. Know Your Customer: Chinese purchasers have adopted export compliance procedures including customer screening
3. Third-Party Testing: Products undergo independent testing in the United States to verify performance and security
4. Foundry Capacity: Production won't divert foundry capacity from US customers

The policy applies only to exports from the United States. Applications for reexport or transfer licenses remain subject to presumption of denial. Chips manufactured in Taiwan must first ship to the US for national security screening before potential onward export to China.

The 25% Tariff Structure

On January 14, 2026—one day before the BIS rule took effect—President Trump signed a proclamation imposing a 25% duty rate on advanced computing chips with the same performance thresholds. The tariff applies to semiconductors produced outside the United States that pass through US territory before export to customers abroad.

President Trump framed the structure as the government "taking 25% of sales". Wall Street analysts characterized it as a "protection fee" providing long-term regulatory certainty—NVIDIA shares rose 4.2% and AMD gained 3.8% following the announcement.

The combined effect: any H200 or MI325X shipped to China faces 25% revenue capture plus mandatory third-party testing costs, while volume cannot exceed half of what ships to US customers.

NVIDIA's Market Response

NVIDIA welcomed the policy shift: "We applaud President Trump's decision to allow America's chip industry to compete to support high paying jobs and manufacturing in America. Offering H200 to approved commercial customers, vetted by the Department of Commerce, strikes a thoughtful balance that is great for America."

Chinese demand materialized immediately. ByteDance prepared orders totaling upwards of $14 billion for 2026. Alibaba and other tech giants contacted NVIDIA for large H200 orders.

The demand surge created supply chain challenges. NVIDIA faces constrained advanced-node capacity at TSMC, which reported supply about three times short of AI-driven demand. The company had largely transitioned H200 production capacity to next-generation Blackwell designs, but signaled it would take new H200 orders starting in 2026.

NVIDIA reportedly prepared an initial shipment of 82,000 H200 GPUs for China following approval.

China's Counter-Response

Despite US export approval, China blocked H200 imports at customs. Chinese authorities informed customs agents that NVIDIA's H200 chips were not permitted to enter the country—without providing explanation.

The decision disrupted supply chains linked to the chips. Suppliers providing H200 components paused production after operating continuously to prepare for post-approval shipments.

Government officials separately warned domestic technology companies not to purchase H200 chips unless necessary. The warning lacked official justification, creating uncertainty about whether the block represents temporary negotiating posture or permanent policy.

The standoff illustrates that US export approval doesn't guarantee market access. Chinese industrial policy may prioritize domestic alternatives—primarily Huawei's Ascend chips—even when US products become available.

Samsung and SK Hynix Annual Licenses

Separately from the H200/MI325X rule, the US Commerce Department approved annual export licenses for Samsung Electronics and SK Hynix to ship chipmaking equipment to their China manufacturing facilities throughout 2026.

The approvals arrived one day before the expiration of Validated End User (VEU) status on December 31, 2025. Under VEU, Samsung's Xi'an plant and SK hynix's facilities in Wuxi and Dalian could receive US-controlled semiconductor tools without individual export licenses for each shipment.

Under the updated policy, companies submit annual plans detailing chip equipment needed for the period. Approvals come on a yearly basis rather than individual licenses per shipment.

Samsung and SK hynix manufacture NAND and DRAM chips in China for servers, smartphones, and AI data centers. The annual license structure introduces new uncertainty—industry observers characterize it as a temporary measure rather than long-term policy shift. Whether approvals continue beyond 2026 remains unclear.

Remote Access Security Act: Closing the Cloud Loophole

On January 12, 2026, the US House of Representatives passed the Remote Access Security Act with overwhelming bipartisan support (369-22). The bill extends export controls to cloud computing services, addressing a loophole Chinese companies had exploited to access restricted hardware.

The loophole emerged publicly in late 2025 when reports revealed Chinese companies accessing NVIDIA's Blackwell chips through cloud rentals. Shanghai-based INF Tech allegedly gained access to 2,300 banned NVIDIA GPUs by renting servers in Indonesia—32 GB200 servers worth an estimated $100 million from an Indonesian telecom company.

The Remote Access Security Act defines remote access as access by foreign persons of concern—specifically those from Russia, Iran, North Korea, and China (including Hong Kong and Macau)—to US-controlled items through cloud infrastructure services.

If enacted, cloud providers and GPU rental intermediaries face stronger compliance obligations around customer screening, license requirements, and how remote access gets provisioned for controlled accelerators. Senator Ron Wyden has partnered on companion legislation in the Senate.

Export Control Effectiveness

The policy shifts arrive amid debate about export control effectiveness. Data suggests controls have widened the US-China AI compute gap rather than allowing China to catch up.

Council on Foreign Relations analysis found that even under aggressive assumptions where Huawei produces millions of AI chips in 2026, the company delivers only about 5% of NVIDIA's aggregate AI computing power. Under median-case assumptions, Huawei produces under 3%—falling to 1% by 2027.

US and allied firms control approximately 90% of global semiconductor manufacturing equipment and 92% of overall supply chain value. Companies including Alibaba Cloud, Baidu, and Tencent have postponed major training runs, requiring thousands of additional processors simply to maintain previous performance levels.

However, enforcement challenges persist. Gray markets have emerged, with distributors in Singapore and Malaysia reportedly diverting restricted hardware toward China through shell companies and falsified documents. While volumes remain too small to offset high-end chip shortfalls, sustained smuggling presents ongoing risks.

Next-Generation Restrictions

The policy creates a rolling window for technology access. NVIDIA's Blackwell B200 and upcoming Rubin architecture remain strictly prohibited for China export for at least 18-24 months after domestic launch.

This framework ensures China receives previous-generation technology while the US maintains capability advantages in frontier systems. The AI Overwatch Act, which passed a key House committee on January 22, 2026, introduces 30-day congressional veto power over any semiconductor export license—creating permanent supply chain uncertainty as Commerce Department licenses could be revoked legislatively.

Implications for Data Center Operators

The interlocking policies create a complex compliance environment:

For US-based operators:

• H200 and MI325X availability unchanged domestically
• Cloud GPU services face new customer screening obligations
• Remote Access Security Act may require infrastructure changes

For operators with China-serving capacity:

• Case-by-case review doesn't guarantee approval
• China's import block creates uncertainty regardless of US policy
• Annual license model for Samsung/SK hynix adds planning complexity

For global infrastructure planning:

• 50% volume cap prioritizes US customer access
• Third-party testing requirements add costs and timelines
• Congressional veto authority creates ongoing uncertainty

Compliance Framework

For organizations navigating the new environment, BIS has established specific compliance requirements:

License Application Requirements:

1. Demonstrate US supply won't be reduced
2. Verify customer has export compliance procedures
3. Complete independent third-party testing
4. Show production won't divert foundry capacity

Ongoing Obligations:

• Customer screening and KYC documentation
• Volume tracking against 50% cap
• Security verification maintenance
• Export compliance training for Chinese customers

Prohibited Activities:

• Reexport to third countries (presumption of denial)
• Cloud access provision to restricted parties
• Circumvention through affiliate structures
• Transfer without license modification

Key Takeaways

For Semiconductor Companies:

• H200/MI325X can reach China under specific conditions, but China's import block creates execution risk
• 25% tariff becomes cost of China market access
• Next-generation products remain restricted

For Data Center Operators:

• US supply priority protected through 50% volume cap
• Cloud GPU rental faces new compliance obligations
• Remote Access Security Act may require infrastructure review

For Supply Chain Planners:

• Samsung/SK hynix annual licenses introduce planning uncertainty
• Congressional veto authority creates ongoing policy risk
• Gray market enforcement remains inconsistent

For China-Serving Operations:

• US approval doesn't guarantee China access
• Domestic alternatives (Huawei Ascend) remain primary option
• Annual license reassessment affects long-term planning

About Introl

Introl deploys GPU infrastructure for AI data centers across complex regulatory environments. With 550 HPC-specialized field engineers operating in 257 global locations, Introl has deployed over 100,000 GPUs—navigating export compliance, supply chain constraints, and regional requirements. When policy complexity intersects hardware deployment, Introl ensures AI infrastructure reaches its destination.