Test report DSG-9957 · Rev D · tested November 18, 2026
Supply Chain & PolicyDevice under test
U.S. Closes Overseas Data Center Loophole; Fed Flags $1 Trillion AI Exposure
New BIS rules block Chinese firms from reaching U.S. AI chips via overseas data centers, tightening Samsung's foundry vetting and SK hynix's HBM outlook as the Fed flags $1 trillion in bank exposure.
- Read
- 3 min
- Words
- 615
- Node
- 3nm
- Operator
- Elena Vasquez
Spec summary
- BIS rules released December 18 restrict Chinese access to U.S. AI chips via overseas data centers above 12 million TPP.
- Existing data centers must obtain a U.S. license within 90 days of the rules taking effect.
- GPUs accessed by China through indirect routes account for about 20% of global usage, per industry estimates.
- New AI infrastructure debt exceeded $680 billion January–October, versus $130 billion in all of 2023.
- The Fed estimates about $1 trillion in bank risk exposure and over $3 trillion in obligations at seven large tech firms.

Roughly 20% of the world's GPU usage flows through indirect routes accessed by China — and as of December 18, the U.S. Commerce Department's Bureau of Industry and Security (BIS) has moved to shut that channel down. New export control rules extend restrictions beyond direct chip shipments to China, covering U.S.-made AI chips reached through overseas data centers. For Samsung Electronics and SK hynix, that changes both the compliance load and the demand picture.
What do the new rules actually require?
The controls apply to data centers providing AI computing services above a combined 12 million total processing performance (TPP) to companies in China and other restricted countries. Supplying advanced chips to those facilities now requires U.S. government approval. Data centers already delivering such services must obtain a license within 90 days of the rules taking effect.
Washington also tightened licensing for advanced chip exports to five countries: Singapore, Malaysia, Thailand, Indonesia and Brazil. The conclusion behind the move: Chinese companies had been parking compute in Southeast Asia and elsewhere to sidestep existing restrictions.
How does this hit Samsung's foundry?
Samsung's foundry business falls directly in scope. The rules strengthen requirements for overseas contract chipmakers and packaging firms to verify the location of design companies and their ultimate parent companies when producing advanced chips. They also narrow the avenues through which Chinese-affiliated chip designers can prove their products sit outside export controls.
In practice, Samsung will spend more time and money vetting the ownership structure and chip performance of Chinese-affiliated customers, and some products may need U.S. government sign-off. Some orders could be disrupted.
What about SK hynix and HBM?
SK hynix's exposure runs through its high-bandwidth memory (HBM) business. If Chinese companies can't easily use Nvidia GPUs at overseas data centers, those data centers will install fewer GPUs — and fewer GPUs means less HBM demand. SK hynix is a major HBM supplier to Nvidia; Samsung's HBM business faces the same squeeze.
Why is the Fed involved?
The Federal Reserve's financial stability report, released December 13, adds a second layer of risk. New debt raised for advanced computing infrastructure, including data centers, exceeded $680 billion from January through October this year — more than five times the $130 billion raised in all of 2023. Risk exposure at large U.S. banks, including loans and credit to tech companies and data center developers, stands near $1 trillion.
The Fed also flagged that future lease payments, contractual obligations and financial guarantees owed by seven large tech companies — among them Nvidia, Microsoft and Oracle — exceed $3 trillion.
What changes for the chipmakers' bottom line?
Even without an immediate hit, the business environment for Samsung and SK hynix looks tougher over the medium to long term. If data center operators struggling to raise funds scale back or delay investment, GPU and server purchases decline, dragging orders for HBM and server DRAM with them. Both companies are pouring enormous sums into expanding HBM capacity, so slower-than-expected demand growth would weigh on pricing power and profitability.
Some observers expect the near-term earnings impact to stay limited. Samsung said at its earnings presentation last month that its entire 2027 HBM supply is already sold, and SK hynix has locked in substantial volumes through long-term supply contracts with major customers. Skeptics counter that even with LTAs in place, a rapid downturn could push some buyers to forfeit advance payments rather than take delivery.
Either way, the loophole that carried a fifth of global GPU usage is closing, and Korean chipmakers no longer get to treat compliance and financing risk as someone else's problem.
via en.sedaily.com (Source)
More from Elena Vasquez
Show full bio
Senior reporter covering people, launches and day-to-day practice in hardware layer of artificial intelligence at Die Signal.
675 articles
Same lot · LOT-362E
- DSG-541820nmUS Rule Closes Cloud Loophole in AI Chip Export Controls
- DSG-445745nmU.S. Rule Hits Rented AI Compute: Five Countries, 90 Days
- DSG-150745nmUS Tightens AI Chip Exports After Nvidia Smuggling Scandal
- DSG-35695nmUS Proposes License Regime for Nvidia Chip Exports to Five Countries
- DSG-85277nmU.S. Export Controls Now Target Who Uses AI Compute, Not Just Where Chips Go