More than 50%. That was the reported reduction in Nvidia’s authorized Asian AI-chip client roster in July 2026. It measured which customers remained on a supplier’s list, not how many GPUs were still running.

Key takeaways

  • Nvidia reportedly cut its authorized Asian AI-chip client roster by more than 50% in July 2026.
  • In August 2023, Nvidia reportedly asked some cloud-provider startups to identify their customers.
  • On November 6, Nvidia reportedly told cloud-compute clients to sell more than 51% of their services to non-Chinese companies.
  • Megaspeed, a Singapore-based operator, purchased nearly $2 billion in Nvidia technology.
  • The U.S. Commerce Department loosened AI-chip controls for the UAE in July, allowing G42 and certain U.S. companies to export chips there without a license.

An authorized customer is not the same as a running GPU

Nvidia reportedly cut the roster while intensifying due diligence in Singapore, Malaysia and Japan. The distinction matters. A cloud operator can own working chips and still face a different question when it seeks its next allocation or supplier designation.

In August 2023, Nvidia reportedly asked some cloud-provider startups to name their customers. Nvidia also ensured some upstart providers received early new-chip shipments, according to the same report. The reporters did not establish that Nvidia conditioned those shipments on customer names. This November, Nvidia reportedly required Southeast Asian applicants for Nvidia Cloud Partner status to enroll in a program tracking their AI chips’ locations.

Nvidia’s partner condition reaches past the chip sale into where the hardware operates. Its separate customer checks reach toward who consumes the capacity. For tenants renting AI-chip capacity, the service may look locationless while Nvidia insists on a map.

A country reading cannot name the tenant

Reuters reported in December 2025 that Nvidia had privately demonstrated unreleased technology capable of indicating the country in which a chip operated. The supplied reporting does not establish that Nvidia deployed that technology in its new partner program, or explain how the program tracks locations. A country reading would not, by itself, identify the company renting the chip’s output.

Nvidia’s other reported test goes after that missing field. On November 6, sources said the company had verbally told cloud-compute clients they must sell more than 51% of their services to non-Chinese companies. That is a reported commercial instruction, not a demonstrated universal legal rule. The account does not explain how Nvidia would count those services.

Location, inventory, ownership and end-customer records answer different questions in a compute chain of custody. An operator can record a chip in Singapore without showing who controls its capacity. Nor does the supplied research contain a primary export-control document imposing one universal tracking or audit requirement. Nvidia’s reported partner condition and the government’s legal requirements should not be collapsed into one rulebook.

Megaspeed lost a partner amid unanswered questions

Megaspeed, a Singapore-based operator, purchased nearly $2 billion in Nvidia technology. U.S. officials were reported to be investigating whether it had helped Chinese companies evade export curbs. Later, sources reported inconsistencies in its chip inventory and scrutiny of alleged Chinese ownership. Those are reported concerns, not findings of diversion.

Megaspeed appeared in five raw coverage items in the first 38 days of 2026’s fourth quarter, against one in all of the second quarter. Those items measure attention, not five verified incidents. One commercial development carries more weight: Bain’s data-center unit reportedly cut ties with Megaspeed while the investigation was underway. The supplied reporting does not establish the contract terms or prove why Bain made the decision.

Bloomberg reported finding no evidence that Megaspeed’s Nvidia chips had been diverted to China. Separately, Nvidia said it had seen no substantiation or received tips about smuggling through data centers outside China. Neither statement resolves the inventory questions. Both limit what can be inferred from them: a supplier or landlord may demand a legible account before anyone has proved misconduct.

A government can open access while a supplier narrows its list

The U.S. Department of Commerce loosened controls for the UAE in July, allowing G42 and certain U.S. companies to export AI chips there without a license. G42 nevertheless reportedly explored selling a majority stake to U.S. companies to safeguard access beyond April 2027. Ownership could matter to future supply even after a licensing barrier eased. That prospective sale is not evidence that Nvidia imposed its Southeast Asian tracking condition on G42.

Commerce also withdrew a proposed rule tightening AI-chip exports after circulating a draft for agency feedback. Nvidia’s reported partner test sits on a different timetable: a supplier can revise its commercial eligibility checks without waiting for a universal export rule. Operators must keep purchase records, deployed inventory and customer accounts intelligible across both timetables. A one-time assurance can age faster than a data center.

Frequently asked questions

What records would be sufficient for a cloud operator to qualify for a future Nvidia shipment or partner designation?

The supplied reporting does not specify a required document set, audit standard, or retention period. It indicates that location, inventory, ownership, and end-customer records are distinct categories Nvidia may scrutinize.

How would Nvidia calculate the reported requirement that more than 51% of services go to non-Chinese companies?

The account does not explain the measurement method. It does not say whether Nvidia would count revenue, compute hours, contracts, affiliated entities, or another basis.

Did Nvidia make naming end customers a formal condition for receiving chip shipments?

Not on the supplied evidence. Nvidia reportedly asked some startups to name customers in August 2023 and ensured some received early new-chip shipments, but the reporting did not establish that the shipments were conditioned on providing those names.

What would G42’s prospective majority sale to U.S. companies change after April 2027?

G42 reportedly explored such a sale to safeguard future access, but the supplied reporting gives no deal terms, completion date, or assurance that a sale would secure supply. It also does not show that Nvidia applied its Southeast Asian tracking condition to G42.

Reported milestones in cloud-chip oversight

  • August 2023 — Nvidia reportedly asked some cloud-provider startups to identify their customers.
  • December 2025 — Reuters reported that Nvidia had privately demonstrated unreleased technology capable of indicating the country where a chip operated.
  • July 2026 — Nvidia reportedly reduced its authorized Asian AI-chip client roster by more than 50%.
  • November 6, 2026 — Sources said Nvidia had verbally told cloud-compute clients that more than 51% of their services must be sold to non-Chinese companies.
  • November 2026 — Nvidia reportedly required Southeast Asian applicants for Nvidia Cloud Partner status to enroll in a program tracking AI-chip locations.

More than 50% was a cut to Nvidia’s authorized Asian client roster, not a measured loss of working GPUs. The machines can remain in their racks. What the number puts in question is who can credibly account for them when the next shipment is allocated.