Chip Export Rules Are Redrawing the Data Center Map

On January 15, 2026, the Bureau of Industry and Security quietly rewrote one of the rules that determines which GPUs a data center is legally allowed to run — and where. For most of the industry’s history, site selection has been a power-water-fiber-tax exercise. In 2026, a fourth variable has forced its way onto that checklist: export control jurisdiction. Which chips you can actually install, how fast you can get them, and what compliance burden comes attached now depends as much on geography and politics as on your utility interconnection date.

The rule that changed in January

BIS moved license review for Nvidia’s H200 and AMD’s MI325X exports to China and Macau from a presumption of denial to case-by-case review, provided the chips fall under a Total Processing Performance threshold of 21,000 and total DRAM bandwidth below 6,500 GB/s. Four conditions apply to every license: the exporter cannot divert supply away from US customers, Chinese buyers must run documented Know Your Customer and export-compliance screening, an independent US lab must verify chip performance and security, and foundry capacity can’t be redirected from domestic orders to fill the new demand. A companion proclamation signed the day before imposed a 25% duty on advanced computing chips crossing US territory under those same thresholds, with a volume cap pegged to US shipment levels.

The demand on the other side of that door is real. ByteDance reportedly lined up orders exceeding $14 billion for 2026, with Alibaba, Tencent, and Baidu also signaling interest, and Nvidia prepared an initial shipment of roughly 82,000 H200 units. China’s own customs authority responded by blocking H200 imports without public explanation and warning domestic firms off buying them except where necessary — a reminder that Washington loosening the valve doesn’t guarantee Beijing leaves it open. Blackwell (B200) and the upcoming Rubin generation remain off the table for China entirely, and the gap between a US product launch and any eventual China-eligible version is running 18 to 24 months. For a buyer modeling multi-year capacity in or near that market, that lag is now as load-bearing an assumption as a transformer delivery date.

Washington’s carve-outs for the Gulf

While China-bound chips got a narrower, fee-laden door, the Gulf got a different one entirely. In July, Commerce moved the UAE out of the licensing tier it shared with China and Yemen and into the same category as India, South Korea, and European allies — a reclassification the Wall Street Journal tied to the UAE’s support during the Iran war. The practical effect: G42, the Emirati AI holding company chaired by Tahnoun bin Zayed Al Nahyan, now has unrestricted access to buy advanced Nvidia and AMD processors, and hyperscalers building data centers in the UAE no longer need individual export licenses to ship chips there. That followed a November 2025 approval letting the UAE and Saudi Arabia’s respective “AI champions” each buy tens of billions of dollars in Nvidia GB300 systems or equivalent hardware. G42’s CEO said in January that the first shipments — a mix of Nvidia, AMD, and Cerebras silicon — were expected within months, enough to light up the first 200 MW of a planned 1 GW Stargate UAE cluster.

The lesson for buyers isn’t “the Gulf is now easy.” It’s that export eligibility is being negotiated bilaterally and can move fast in either direction, tied to geopolitics that have nothing to do with your lease terms. A jurisdiction that was effectively unbuildable eighteen months ago can become a frontier market overnight, and the reverse is just as true. Treating a country’s export status as a fixed input rather than a variable to monitor is how a capacity plan quietly goes stale.

The compliance layer nobody priced into the deal

Export policy isn’t just about which chips cross a border on day one — it’s increasingly about who gets to touch them remotely afterward. The House passed the Remote Access Security Act 369-22 in January, aiming to close a loophole where restricted parties access controlled GPUs through cloud rental rather than direct import. Investigators cited a Shanghai-based firm, INF Tech, that allegedly reached 2,300 banned Nvidia GPUs by renting servers in Indonesia, including 32 GB200 servers worth roughly $100 million. If that bill becomes law, cloud GPU providers and colocation operators hosting foreign tenants inherit a new layer of customer-screening obligations on top of the ones already required for export licensing. A separate bill, the AI Overwatch Act, cleared a key House committee in January and would give Congress a 30-day veto window over semiconductor export licenses — adding a legislative wildcard on top of an already case-by-case regulatory process.

For anyone leasing GPU capacity from a neocloud or colocation provider rather than buying hardware outright, this is no longer someone else’s compliance problem. Know Your Customer obligations, third-party testing requirements, and potential remote-access restrictions all flow downstream into contract terms, onboarding timelines, and which tenants a provider will even accept. It’s worth asking any GPU-as-a-service vendor directly how they screen customers today and how they’d handle a Remote Access Security Act-style rule if it passes — the answer tells you a lot about how exposed your capacity commitment is to a policy shift you don’t control.

The supply chain is getting routed through Washington too

It isn’t only finished GPUs that now pass through a US checkpoint. Under the January rule, Taiwan-made chips headed for China must first ship to the US for national security screening before any onward export — adding a hop, and a delay, to a supply chain that already runs through TSMC capacity reportedly running at roughly a third of AI-driven demand. Samsung and SK hynix, meanwhile, had to swap their old Validated End User status for new annual licenses just to keep operating chipmaking tools at their China facilities (Samsung’s Xi’an plant, SK hynix’s Wuxi and Dalian sites), with no guarantee the approval carries into next year. None of that shows up on a colocation term sheet, but it shapes whether the memory and logic going into your racks arrives on schedule, and from where.

What this means for where you build

Three practical implications follow for anyone sizing a build-to-suit campus, a colocation contract, or a GPU-cloud commitment in 2026. First, the United States remains the only place to get the current chip generation on day one with no licensing friction — that’s a real advantage buyers should weigh against higher US power and land costs, not assume away. Second, if your roadmap depends on a restricted or semi-restricted market, model the 18-to-24-month generational lag explicitly rather than assuming “it’ll loosen up eventually.” That lag compounds with the broader GPU and memory supply crunch already stretching lead times industry-wide, so a chip that’s merely scarce in the US can be functionally unavailable on your preferred timeline somewhere else. Third, treat export jurisdiction the way you’d treat a power queue or a water permit: something to check at contract signing, re-check before each capacity tranche, and build a contingency plan around, because the rules that applied when you signed a 10-year colocation agreement may not be the rules in year three.

None of this argues for panic. The Gulf deals show that chips still move quickly to approved partners once the politics align, and the case-by-case China pathway, however narrow, is wider than the blanket denial regime it replaced. The nuance is that “export policy” has graduated from a legal footnote to an operational input, on the same tier as interconnection queues and transformer lead times — and it’s one that can change with a single proclamation rather than a multi-year infrastructure build cycle.

If you’re underwriting a capacity deal or a build in 2026, chip export jurisdiction belongs on the same diligence checklist as power availability and permitting timelines — not as a footnote your legal team reviews after the term sheet is signed. For a structured way to run that diligence across your next deal, check out TechInfraHub’s data center buyer’s toolkit.

Written by

Raajeev Ratra

Data Center Infrastructure Expert | 15+ Years in DC Design, Operations & Project Management

Raajeev is a seasoned data center professional with hands-on experience in hyperscale facilities, colocation design, power & cooling infrastructure, and global DC operations. He shares practical insights to help engineers and IT leaders build better infrastructure.

Connect on LinkedIn →