Nvidia put Singapore-based Megaspeed on a no-sell list in June 2025. The U.S. Commerce Department had begun investigating Megaspeed the year before, yet Commerce had not added it to the Entity List by November 2026. The supplier closed a gate that the government had left open.

Key takeaways

  • A U.S. rule announced in January 2025 gave 18 key allies full access to AI chips while requiring licenses for most other countries, including China.
  • Nvidia placed Singapore-based Megaspeed on its no-sell list in June 2025, after the Commerce Department began probing the company in 2024.
  • Nvidia reportedly told some cloud-compute clients in 2026 that more than 51% of their services must be sold to non-Chinese companies.
  • Commerce permitted G42 and U.S. companies including Apple, Meta and xAI to export AI chips to the UAE without a license in July 2026.
  • A Reuters review identified more than $3.2 billion in Chinese compute projects built or planned around Nvidia GPUs restricted by U.S. controls.

AI chip export controls began with a question a customs officer could answer: where may this hardware be shipped? Cloud rental poses a second question after delivery: who can use the compute, from where, and under whose control? Suppliers and cloud operators are becoming the working perimeter of the rule, checking location, customer identity and corporate ownership while retaining the ability to restrict future supply or access. A shipping license cannot do that work once a server is running.

A shipment rule cannot govern a rental after the shipment

The original design assigned countries to access tiers. A U.S. rule announced in January 2025 gave 18 key allies full access to AI chips while requiring licenses for most other countries, including China. That system could govern a box leaving a warehouse. It could not, by itself, establish who later rented an hour on a GPU inside the box.

U.S. authorities reportedly placed location trackers in selected high-risk shipments, including Dell and Supermicro servers, to detect diversion to China. A tracker addresses the hardware’s route. A remote customer presents a different problem: the server can remain in its approved country while delivering compute across a network. An investigation traced a Chinese startup’s access to advanced Nvidia chips at an Indosat data center in Jakarta, illustrating why a server’s address and a user’s address cannot be treated as the same fact.

By December 2025, Nvidia had privately demonstrated technology that could indicate the country in which its chips were operating. In November 2026, Bloomberg reported that Nvidia had turned location checks into a condition of its Cloud Partner program: Southeast Asian applicants had to enroll in a chip-location-tracking program to qualify.

The U.S. also issued a rule limiting remote access by China and other countries to U.S. AI chips and servers in Singapore, Malaysia, Thailand, Indonesia and Brazil. That rule reaches the rental rather than only the shipment. To apply it, an operator needs a compute chain of custody: an account of which asset is in which facility, who controls it, and who receives its capacity. A serial number alone cannot identify a tenant.

Megaspeed shows what a supplier can do before Washington agrees

Commerce says it began probing Megaspeed in 2024 over possible export-rule violations; Nvidia placed the company on its no-sell list in June 2025. Reporting describes Megaspeed as having bought nearly $2 billion of Nvidia technology while U.S. officials examined whether it had helped Chinese companies evade restrictions. A Justice Department investigation focused on Megaspeed could also involve Nvidia, according to Semafor. None of that establishes that Megaspeed violated export controls.

Nvidia’s decision and a government designation serve different purposes. A supplier can stop selling to an account because it cannot satisfy its own risk threshold. The Entity List requires a federal decision with wider consequences. Scott Bessent said Treasury, Defense and State had favored adding Megaspeed to that list for months. Nvidia had stopped selling to Megaspeed while Commerce, which controls the list, had not added the company.

Megaspeed says it runs a legitimate and lawful business, says former executive Alice Huang left in August 2025, and has opened an internal investigation. Reported concerns about its inventory, ownership and possible smuggling remain allegations.

Cloud operators must make the customer behind the customer legible

Nvidia was asking some cloud-provider startups for their customers’ names as early as 2023. At the time, the company was also helping some young providers obtain early shipments of new chips. Nvidia was learning about downstream demand through a commercial relationship, before remote access became an explicit focus of the U.S. rule.

Customer names answer only part of the question. A cloud operator may know which company signed a rental agreement without knowing whether that company’s ownership, users or onward sales fit the conditions attached to preferred access. Nvidia reportedly told cloud-compute clients in 2026 that more than 51% of their services must be sold to non-Chinese companies. That reported verbal instruction turns a provider’s revenue mix into a supplier concern; it should not be mistaken for a published, universal requirement on every Nvidia customer.

G42 exposes the ownership side. The United Arab Emirates gained broader U.S. chip access in July 2026, and G42 was allowed to buy U.S. AI chips freely for at least nine months. G42 said it planned to become a U.S. company. By September, it was reportedly exploring a majority sale to U.S. companies to preserve access beyond April 2027. No GPU needed to move for corporate control to become relevant to its future supply. The distinction resembles the one in sovereign compute contracts: where capacity sits and who controls it are separate questions.

Trusted capacity costs more to operate than a GPU fleet

A cloud provider selling into this environment has to account for more than available chips. It needs asset records, evidence of facility location, customer attribution and a way to decide which accounts may receive restricted capacity. Those are operational tasks, not paperwork completed when a purchase order clears. Providers must keep the records useful as hardware moves, ownership changes and tenants come and go.

Nvidia reportedly cut its authorized Asian AI-chip client base by more than half while intensifying diligence in Singapore, Malaysia and Japan. It also reportedly offered some young cloud providers a financial backstop, renting unused GPUs in exchange for a share of their revenue.

Bloomberg’s Cloud Partner reporting has an important limit. It establishes a reported enrollment condition for certain Southeast Asian applicants; it does not disclose continuous-monitoring terms, which hardware is covered, who can see the location data, or what happens beyond loss of program eligibility if an applicant declines. Nor does it establish a means to disable an already deployed GPU remotely. The documented gate is access to a supplier program, not a demonstrated universal kill switch.

Political permission and incomplete visibility leave gaps in the perimeter

Commerce loosened UAE controls in July 2026, permitting G42 and U.S. companies including Apple, Meta and xAI to export AI chips there without a license. Commerce also withdrew a separate planned rule tightening chip exports after a draft circulated for interagency feedback. Governments can grant access to favored partners and abandon proposed restrictions; supplier diligence does not replace those choices.

Enforcement faces a separate evidentiary limit. In December 2025, Nvidia said it had not seen substantiation or received tips about chip smuggling through data centers outside China. A later Reuters review identified more than $3.2 billion in Chinese compute projects built or planned around Nvidia GPUs restricted by U.S. controls. Built and planned projects are not proof that every chip reached its intended site, nor that each project involved an unlawful shipment. They do show why a list of approved destinations cannot settle where restricted compute ultimately serves demand.

Frequently asked questions

What exact technology does Nvidia use to determine a chip’s location?

The reporting establishes only that Nvidia demonstrated technology in December 2025 that could indicate the country where chips were operating. It does not disclose the technical method, its precision, or whether it operates continuously.

Does Nvidia’s Cloud Partner tracking requirement apply outside Southeast Asia?

The reported condition applies to certain Southeast Asian applicants to Nvidia’s Cloud Partner program. The available reporting does not establish a worldwide requirement or identify all countries and applicants covered.

What happens if a Cloud Partner applicant declines location-tracking enrollment?

The documented consequence is loss of eligibility for the Cloud Partner program. The reporting does not disclose additional penalties, monitoring terms, or any ability to remotely disable already deployed GPUs.

What would let G42 retain broader U.S. chip access after April 2027?

G42 was reportedly exploring a majority sale to U.S. companies to preserve access beyond April 2027. The evidence does not specify a final ownership threshold, transaction structure, or a U.S. government decision governing post-April 2027 access.

How compute controls moved beyond shipment checks

  • 2024 — The U.S. Commerce Department began probing Megaspeed over possible export-rule violations.
  • January 2025 — A U.S. AI-chip rule gave 18 key allies full access while requiring licenses for most other countries, including China.
  • June 2025 — Nvidia placed Megaspeed on its no-sell list.
  • December 2025 — Nvidia privately demonstrated technology that could indicate the country in which its chips were operating.
  • July 2026 — Commerce loosened UAE controls, allowing G42 and U.S. companies including Apple, Meta and xAI to export AI chips there without a license.
  • September 2026 — G42 was reportedly exploring a majority sale to U.S. companies to preserve access beyond April 2027.
  • November 2026 — Bloomberg reported that Southeast Asian applicants to Nvidia’s Cloud Partner program had to enroll in chip-location tracking to qualify.

The border inspection still has a job: a crate containing servers must reach an authorized address. But once those servers enter a cloud facility, the crate keeps its customs stamp while customers arrive by login.