
Four days after Jensen’s exit, Washington moved the export license from the shipping dock into your server room. Change who you serve, and you’ve exported. Outside the US, every rack now comes with a landlord.
Sovereign AI | theCUBE Research | Amit Eyal Govrin | November 2026
The lease that didn’t survive a phone call
In October, the Financial Times reported that Tencent had signed the biggest overseas lease in its history. Five years. About 100,000 advanced AI chips, spread across several Oracle data centers in Southeast Asia. Roughly $7 billion, with Tencent paying about 30% upfront [6].
On November 4, the FT reported the lease was dead. US officials raised concerns, and Oracle walked [6].
No court. No ruling. No rule. A conversation.
Hold that story, because it’s the whole edition.
Last week I closed the Jensen edition with this: if someone in another capital can locate your hardware, read its logs and switch it off, you don’t own a data center. You own a very expensive lease.
I thought that line would age slowly. It aged in four days.
On Monday, November 16, the Commerce Department’s Bureau of Industry and Security released a 132-page interim final rule called “Addressing Remote Access to Advanced Semiconductors” [1][4]. It takes effect when it lands in the Federal Register, expected Thursday, November 19 [1]. Same day Trump sits down with Xi in Shenzhen [3]. Nobody planned the seating chart. Probably.
The Oracle phone call now has paperwork. The license used to stop at the shipping dock. Now it follows the chip into your server room and checks who’s renting it.
The neocloud nobody had heard of
Credit first, because the problem Washington is chasing is real.
Meet Megaspeed. Founded in Singapore in 2023, it became the largest buyer of Nvidia chips in Southeast Asia in under three years. Customs records from Malaysia and Indonesia show at least $4.6 billion in Nvidia imports, about 136,000 chips, more than half of them Blackwell [8]. One of its key tenants was Alibaba, which, per Bloomberg, reportedly reached those chips in Malaysia through a Singaporean shell company controlled by a Cayman Islands entity [4].
Megaspeed denies any wrongdoing and says inspections by US and Malaysian authorities and by Nvidia found no violations [9]. Then a former Megaspeed executive named Alice Huang became the center of the scandal that ended Jensen Huang’s run at Nvidia [4]. That’s how a neocloud nobody had heard of ended up rewriting export policy.
It isn’t one company. In March, the Wall Street Journal reported ByteDance was lining up about 36,000 Blackwell chips in Malaysia through a local cloud called Aolani, a build that could cost over $2.5 billion [24]. The Institute for Progress projects that China’s remote access to US chips in overseas data centers will nearly triple by the end of 2027, to about 2.8 million H100 equivalents [10]. On November 12, the day Jensen resigned, Reuters reported Chinese AI firms were racing to lock in overseas Nvidia capacity before the door closed [7]. They read the room correctly.
And the rule shows more of its work than most. It publishes a threshold. It reviews licenses case by case instead of banning outright. It opens a 60-day comment period [1][3]. If you’re going to build a fence, telling people where it is counts for something.
Then you read the lease.
Read the lease
Three clauses do the work.
The tenant clause. A license is required when a data center provides more than 12 million “aggregate TPP” of compute to companies headquartered in China, Macau or the other Country Group D:5 destinations, or whose ultimate parent is [1]. Linked data centers count together, including ones that are only digitally linked [1].
The sublet clause. It applies to chips already installed. A data center that starts serving restricted customers after the chips arrive needs a license too, because BIS treats that change as an in-country transfer [2]. Read that twice. Your tenant list is now an export.
The screening clause. Exporters must get the data center’s know-your-customer and ownership information. A data center that refuses or can’t provide it raises a red flag that presumptively establishes knowledge [1]. And mixing US chips into a foreign-built server doesn’t wash them [1].
One more detail about who held the pen. BIS said the rule was written by and at the direction of Treasury and the National Security Council. CFR’s Chris McGuire, who called it a major step forward, couldn’t resist: “I have never seen a BIS rule that was not written by BIS.” [3]
Four tenants, one landlord
Here’s the part nobody said out loud on Monday.
The rule names five countries with “a higher-than-normal risk of diversion”: Indonesia, Malaysia, Singapore, Thailand and Brazil. Shipping advanced chips there now needs a license. Unless the recipient is headquartered in the US, in which case a monthly report will do [3].
Same building. Same chips. Two lanes. The national champion fills in an application. The American hyperscaler fills in a form.
Malaysia. A year ago, YTL switched on a 600MW Nvidia-powered data center in Johor as the anchor of Malaysia’s sovereign AI. Prime Minister Anwar Ibrahim: “Building our own AI is essential” [11]. Last Thursday, Malaysian data center shares fell, led by YTL Power, on the first Nvidia headlines [12]. On Tuesday, the trade minister said Kuala Lumpur won’t negotiate with Washington and expects operators to comply [5]. Building your own AI is still essential. The application is still due.
Brazil. In August, Brazil split about $444 million between two AI supercomputers, one with Huawei, the other expected to go to Nvidia. The stated logic: “The strategy is not to depend on a single company, technology or country.” [13] On Monday, Brazil made the list, and said it was closely monitoring developments [4]. The hedge worked. The American half now needs a permission slip. The other half comes with a different landlord, and I covered that one’s financing terms in the Egypt edition.
The UAE. The clause I keep rereading. In July, BIS cleared G42 and Core42 to receive advanced chips license-free [14]. That status expires on April 6, 2027, unless they become “U.S. companies” or win BIS approval through its advisory opinion process [14][16]. Amazon, Apple, Google, Meta, Microsoft, OpenAI, Oracle and xAI got the broader approval [15]. The national champion got a lease with an expiry date. The renewal clause asks it to change passports.
Singapore. One of the seven original signers of the Pax Silica Declaration [17]. Signed a fresh AI supply-chain framework with Washington last weekend [4]. Named a diversion risk on Monday [3]. Ally on Saturday. Tenant on Monday.
Four tenants, one pattern. The fastest lane to “sovereign” AI in Johor, Abu Dhabi or Rio Grande do Norte runs through an American headquarters.
The landlord always had a key
None of this started on Monday. We’ve watched the switch get flipped before, one layer up.
In July 2025, India’s Nayara Energy, the country’s second-largest private refiner, found itself locked out of Outlook, Teams and its own cloud data. EU sanctions had landed, and Microsoft suspended service despite fully paid-up licenses [18]. Nayara said there was no legal reason under US or Indian law, sued, and moved its email to a local Indian provider while it waited [19].
That same year, after US sanctions on the International Criminal Court, its chief prosecutor lost his Microsoft email and moved to Proton Mail [20]. Microsoft has since disputed parts of how that happened [21]. The lesson survives the dispute. Those were cutoffs at the software layer. Monday’s rule reaches the silicon.
Europe saw it coming. In December 2025, the European Council on Foreign Relations sketched a deliberately hypothetical “November 2026” in which export controls require individual licenses for AI chips and EU data center orders go on hold [22]. The calendar landed. The target, for now, didn’t. Europe is in the friendly lane this round. The mechanics don’t care which lane you’re in.
And the next clause is already drafted. The Chip Security Act would require a way to track where advanced chips physically are [23]. Nvidia representatives reportedly told one lawmaker it’s technically possible [23]. The Remote Access Security Act passed the House 369-22 and would make remote access itself a controlled activity [2].
In Part 1, I asked Washington to put it in the statute before the button ships. It put the lease in writing. The button is in committee.
The five pillars, under pressure
I’m grading the dependency, not the company. The buyer is anyone outside the US building on US silicon: a ministry in Kuala Lumpur, a bank in Frankfurt, a neocloud in São Paulo.
01 Territorial
The question. Where do your data and compute physically reside?
The story. The racks are in Johor. The concrete is local. But who they may serve is decided by where your customers’ ultimate parents are headquartered [1].
Under pressure. A tenant you signed last year turns out to sit under a Cayman holding company with a parent in Hangzhou.
Verdict. Partial. The building is yours. The guest list isn’t.
02 Operational
The question. Who runs it, who holds the keys, and who can pull the plug?
The story. Your operators now carry KYC duty for a foreign government. Refusing to share ownership data counts against you by presumption, and existing operators get a 90-day clock [1].
Under pressure. Your chip supplier asks for an ultimate-parent map of your customers before the next shipment. You’ve never built one.
Verdict. Not Demonstrated. Your ops team now moonlights as Washington’s front desk.
03 Technological
The question. Can you audit it, fork it, self-host it?
The story. Credit where it’s earned. Open weights still travel, and nothing in the rule touches the model layer. But the silicon under them carries a passport, and a foreign-built server doesn’t change it [1].
Under pressure. You swap accelerators. A second US vendor carries the same passport. A Chinese one carries a different one.
Verdict. Partial. Your weights are portable. Your silicon isn’t.
04 Legal
The question. Which jurisdiction governs, and can anyone enforce it?
The story. My headline finding in “Sovereign or Not” was that no hyperscaler has a practical answer to the CLOUD Act. That was a clawback on your data. Monday added one on your compute.
Under pressure. Your largest customer gets acquired mid-contract. New parent, new license question, same racks.
Verdict. Not Demonstrated. One clawback for the data. One for the compute.
05 Financial
The question. Can you leave on your timeline, at a survivable cost, without someone else’s decision setting your margin?
The story. The pillar that rules them all. Tencent agreed to pay 30% upfront on a lease that didn’t survive a phone call [6]. Local operators in five countries now finance against a case-by-case license while US rivals down the road file a monthly report [3]. The Philadelphia Semiconductor Index fell as much as 5.5% on Monday [3].
Under pressure. Month 14 of a 60-month capacity contract, your provider’s license stalls. Who eats the prepaid capacity?
Verdict. Not Demonstrated. The rent was never the problem. The landlord’s mood is.
The executive TCO
Three line items, none of which show up on the invoice:
- The license premium. Lenders price uncertainty. Your cost of capital just learned a new acronym.
- The two-lane tax. Your American competitor files a report. You file an application. That gap is a margin line [3].
- Stranded capacity. Every tenant you can no longer serve is capacity you still pay for [1].
Same baseline-and-burst as always, applied to the building. Baseline on capacity whose license chain you’ve priced. Burst to the rest.
Own the core. Stay movable at the edges. This week, the edge is the building.
What to do Monday
- Map every customer’s ultimate parent. Not the invoicing entity. The Cayman one above it, and the one above that.
- Ask your GPU provider in writing: Are you above the threshold? Are you licensed? What happens to my capacity if you aren’t?
- Add a license-change clause to every capacity contract: notice, a migration window and a refund on prepaid capacity. Ask Tencent why.
- Run the 72-hour “prohibited, not down” drill on your compute. Assume the racks are fine and you’re not allowed to use them.
- Write it into the risk register. Put a number on it.
You don’t own a data center
Sovereignty is not binary. It’s control with an acceptable risk sidecar. Most of the world will keep building on US silicon, and for most builds I’d sign off on it. Hybrid is fine. Mislabeling isn’t.
Round one took out a CEO. Round two rewrote the lease. Round three is a chip that reports its own address.
If someone in another capital decides who your racks are allowed to serve, you don’t own a data center. You own an expensive lease.
Don’t get caught with your pants down.
Amit
Amit Eyal Govrin is lead analyst for sovereign AI at theCUBE Research and CEO of Agentcy Labs, which designs and builds sovereign AI platforms for enterprises and governments. His research for theCUBE Research is conducted under theCUBE Research’s editorial standards and is not sponsored or directed by Agentcy Labs. Agentcy Labs may hold commercial relationships with vendors covered in this research.
References
[1] Nixon Peabody, “BIS targets remote access to advanced chips through overseas data centers,” Nov 17, 2026. Link
[2] Kharon, “Understanding Remote Access Controls: The AI Remote Access Rule and the Remote Access Security Act,” Nov 2026. Link
[3] The Next Web, “US curbs China’s remote access to AI chips through overseas data centres,” Nov 16, 2026. Link
[4] Free Malaysia Today (Bloomberg), “US tightens AI chip curbs on data centres ahead of Trump’s visit,” Nov 17, 2026. Link
[5] Quartz, “U.S. requires export licenses for AI chips bound for Southeast Asia data centers,” Nov 17, 2026. Link
[6] The Next Web, “Oracle cancels $7B Tencent AI chip lease under US pressure, FT reports,” Nov 4, 2026. Link
[7] The Edge Malaysia (Reuters), “Chinese AI firms rush to secure Nvidia chip capacity as spy scandal stirs fear of US curbs,” Nov 12, 2026. Link
[8] Thairath English, report on Megaspeed and its Nvidia imports, Nov 2026. Link
[9] Open Magazine, “Quick Take | What Is Megaspeed? Seven Things to Know About the Nvidia Chip Buyer in a $4.8 Billion AI Hardware Trail,” Nov 2026. Link
[10] ForkLog, “China’s AI Computing Power to Double by 2027 with U.S. Chips,” Oct 2026. Link
[11] Converge Digest, “Malaysia Accelerates AI Sovereignty with 600MW Data Center,” Dec 4, 2025. Link
[12] The Edge Malaysia, report on Malaysian data center shares falling, led by YTL Power, Nov 12, 2026. Link
[13] The Next Web, “Brazil is building two AI supercomputers, one Chinese and one American,” Aug 21, 2026. Link
[14] Akin Gump (via Mondaq), “BIS Upgrades the UAE from Country Group D to Country Group A Under the EAR,” Jul 2026. Link
[15] Global Sanctions, “US eases UAE export controls,” Jul 2026. Link
[16] Morrison Foerster, “U.S.-UAE Cooperation Deepens, as BIS Eases Export Control Restrictions on UAE,” Jul 20, 2026. Link
[17] ITIF, Pax Silica timeline, Nov 3, 2026. Link
[18] Business Recorder (Reuters), report on Nayara Energy’s legal proceedings against Microsoft, Jul 2025. Link
[19] Outlook Business, “Microsoft Suspends Services to Russia-Backed Nayara, Firm Turns to India’s Rediff,” Jul 2025. Link
[20] TechCrunch, “European leaders worry they’re too reliant on U.S. tech,” Jun 21, 2025. Link
[21] The Register, “Microsoft throws spox under the bus after Parliament testimony on ICC email kerfuffle,” Feb 18, 2026. Link
[22] European Council on Foreign Relations, “Get over your X: A European plan to escape American technology,” Dec 10, 2025. Link
[23] MIT Technology Review, report on the Chip Security Act and Nvidia chip tracking, Nov 12, 2026. Link
[24] MarketScreener (Wall Street Journal), “ByteDance deploys Nvidia AI infrastructure outside of China,” Mar 2026. Link

