$4.6 Billion, 136,000 GPUs: What Megaspeed's Record Shows

Affiliated companies of Megaspeed International imported at least US$4.6 billion of Nvidia equipment between 2023 and November 2025, including no fewer than 136,000 GPUs. Bloomberg's estimate, relayed by Lianhe Zaobao, states that more than half of those units were Blackwell-series parts. Nobody has proven the company broke a law. The figures are estimates, not confirmed totals, and every promise in this story carries that caveat.

What does the company actually do?

Reuters describes Megaspeed as a provider of computing capacity hosted in data centers across Southeast Asia. Its customers buy large volumes of compute, including AI workloads. In plain terms, it is a landlord for GPU time: clients rent slices of accelerators by the hour, month or contract. Specialist GPU clouds (neoclouds, vendors outside the big three hyperscalers) have multiplied since the 2023 training boom, so the model itself is ordinary.

What stands out is the combination. A thin public profile, a premium hardware list and contested ownership, each unremarkable alone, explain why reporters and Washington staffers now read its filings line by line. The customer list is not public. Neither Reuters nor Bloomberg has named a client, which matters: a cloud leasing to anonymous buyers cannot easily show regulators who touched which machine.

What do the filings establish?

Business filings reported by Reuters give the documented timeline:

• 2015: incorporated in Singapore as 7Road International.
• March 2023: renamed Megaspeed International.
• 2023 to November 2025: at least US$4.6 billion of Nvidia equipment imported by affiliated companies, at least 136,000 GPUs, over half Blackwell-series (Bloomberg estimate).

The import window begins in the same year as the rename, but the record does not say who instructed the change or whether the business model shifted. Reuters reported the filing dates and stopped there.

Three details in the Bloomberg figure deserve scrutiny. The wording covers affiliated companies, not Megaspeed alone, and the record does not list which entities qualify or how the link was established. The figures are stated as minimums, and Bloomberg's method is not spelled out in the relayed version. The window ends in November 2025, a full year ago, so the data describes the past, not the current fleet. Dividing US$4.6 billion by 136,000 GPUs yields roughly US$33,800 per unit, inside a plausible range for data-center accelerators; consistency is not validation, since both figures share one underlying estimate.

What hardware does it advertise?

Reuters reports the company's website lists two systems: the GB200 NVL72 (Blackwell-generation rack-scale hardware for large training and inference jobs, the class US export rules aim to keep from Chinese buyers) and the HGX H100 (the older Hopper-generation workhorse of the 2023-2024 build-out). Listings prove advertised access, not unit counts. The majority-Blackwell share implies serious supply access, which supports the picture of a well-funded buyer and nothing more.

Who runs it, and what is alleged?

The record is messy. Reports describe Alice Huang as founder, CEO, managing director, senior executive or former executive; her role is unconfirmed, and the Economic Times ran a piece on precisely that confusion. Claims she founded the company in 2023 conflict with the 2015 incorporation. Ownership is equally unsettled: some accounts name Tan Yong Pong, also known as James Tan; other accounts conflict.

Diversion allegations are suspicion, not findings. No charge, ruling or regulator finding appears in the reporting. Two meanings of diversion apply: physical shipment to a restricted destination, or remote access where foreign customers rent compute without hardware moving. Reporting does not say which, if either, regulators are examining. Megaspeed has reportedly said inspections and reviews found no evidence of illicit transfers; The Week reported its denial. The full wording, date and inspectors are not established. The only verified quotes are the company's own: its site calls Nvidia a "core business partner" (CNBC) and its LinkedIn page claims "privileged access to cutting-edge products, services, and cloud computing capabilities" (Business Times). Nvidia has issued no detailed statement.

Why does the market care?

Nvidia shares fell 6.4% on a day the S&P 500 rose 1.9% on governance news, a reminder of the sensitivity. Channel risk is the second question: findings against any partner could tighten customer checks across every neocloud, slowing shipments and raising compliance costs. Management faces questions on partner vetting at the November 17 earnings call. The Bloomberg figure also reframes Southeast Asian GPU hosting as a multibillion-dollar sub-market, not a rounding error, regardless of what any probe finds. The allegations are serious. They remain allegations.