US Justice and Commerce Departments Confirm Probe Into Singapore’s Megaspeed

US Justice and Commerce Departments Confirm Probe Into Singapore’s Megaspeed

The US Department of Justice and the Department of Commerce confirmed on 17 November that they are investigating Megaspeed, a Singapore-based data centre and GPU compute leasing company, after reporting suggested the probe had been halted in 2025. The case is intertwined with the resignation of Nvidia chief executive Jensen Huang, turning a low-profile compliance review into a public test of how AI compute is intermediated across borders. The two agencies' rare on-record confirmation formally

Reporting first brought the Megaspeed inquiry into public view in 2025, when it was described as having been set aside. On 17 November 2026 the Justice Department and the Commerce Department broke their silence. In their response the agencies said they do not typically comment on ongoing investigations but that it is in the public interest to provide the facts and address public concern, then confirmed the inquiry remains active. On-record confirmation of an open case is unusual for US federal agencies.

Megaspeed is registered in Singapore and operates data centre capacity and GPU leasing, serving customers across Southeast Asia. Public records do not disclose its funding, rack count or client list. The company is intertwined with Huang's resignation: multiple outlets discuss the two together, but neither Nvidia nor Megaspeed has publicly explained the link. Megaspeed has not issued a

Three threads converge. First, the rules keep moving: Washington tightened AI chip export criteria in October 2022, October 2023 and January 2025, when the AI diffusion rule placed Singapore in tier two; that rule was rescinded in May 2025, forcing companies to reset compliance baselines repeatedly inside 18 months. Any intermediary can find room for interpretation in those gaps.

Second, Singapore's role is unusual. Nvidia billed about $23.7 billion to the city-state in FY2025, roughly 18% of $130.5 billion in revenue, while the company has said most of those shipments went to other destinations. The gap between billing location and physical destination is exactly where intermediaries are

Liu Gong believes the real target here is not a single Singapore company but the entire category of compute intermediaries. The prediction: within six months, the US Bureau of Industry and Security is likely to extend filing requirements to GPU leasing, sub-leasing and cross-border scheduling, rather than continuing to rely on chip model numbers as the primary hook. Within twelve months, at least one Asia-Pacific GPU cloud provider will likely be required to submit end-user lists to regulators. The reason is not Megaspeed's size but its position in the seam where billing location and physical destination diverge.

From a technical standpoint, export controls measure hardware, not time slices. The October 2022 rules drew lines around interconnect bandwidth, the October 2023 rules switched to total processing performance and performance density, and the January 2025 AI diffusion rule sorted countries into tiers. All three describe how powerful a chip is; renting compute sells how long a chip is available. In a containerised multi-tenant environment, one accelerator can serve customers in three jurisdictions within 24 hours. The serial number does not change; the end use already has. The analogy is inspecting a container's weight after the cargo has been repackaged as parcels. In industrial machine vision, buyers pay for stable inference latency and yield, not for a specific board, and leasing arrangements are inherently cross-border. An audit that only checks purchase invoices checks nothing; it has to attach to call logs, identity authentication and workload profiles.

On the business side, price spreads and billing structure combine to create motive. Nvidia's FY2025 revenue was $130.5 billion, of which about $23.7 billion was billed to Singapore, roughly 18%, against local data centre IT load of around 1.1 gigawatts. Billing value and physical capacity plainly do not match, and the company itself says most shipments move on to other markets. Over the same period, H100 rental rates fell from around $8 per GPU-hour in 2023 to $2-3 in 2025, a decline of more than 60%, while customer due diligence, export classification and legal opinions only became more expensive. Intermediaries with compressed margins have the strongest incentive to take orders that ask fewer questions and pay faster. The first effect of higher compliance cost is usually not the elimination of large players but the washing out of small sub-leasers with no legal team.

For the industry, this changes valuation models, not just compliance checklists. Huang's resignation puts board governance, related-party disclosure and personal liability on the table at once, and two federal agencies have now confirmed an open case on the record. The market will apply a governance discount to compute intermediaries: the same halls and card pools will be worth more when held by an operator with owned facilities and auditable due diligence than by a pure reseller. Looking at the 2025 Singapore server-transshipment fraud prosecution, once the judicial path is established, civil settlements and criminal charges tend to arrive together, and insurers and banks will reprice collateral backed by compute assets.

For Nvidia, the real risk is not the size of any penalty but a redefinition of the billing-location convention. If regulators require reporting by actual destination and end user, the meaning of that 18% Singapore figure changes, and the regional revenue structure in quarterly filings would need restating. That moves valuation more than any single investigation.

  • Nvidia's FY2025 revenue was $130.5 billion, with $115.2 billion from data centres, or 88%; about $23.7 billion was billed to Singapore, roughly 18% of group revenue.
  • Singapore's data centre IT load is roughly 1.1 gigawatts, with a regulatory roadmap targeting about 300 megawatts of new capacity by 2026; neighbouring Johor in Malaysia has more than 1 gigawatt in publicly disclosed construction and planning.
  • H100 rental rates fell from around $8 per GPU-hour at the 2023 peak to about $2-3 in 2025, a drop of more than 60%, while compliance and due diligence costs only rose.
  • US AI chip export rules were rewritten in October 2022, October 2023 and January 2025; the AI diffusion rule

The near-term effect runs through contracts. Within three months, GPU cloud and leasing providers across Asia-Pacific are likely to rewrite standard terms to include end-user clauses, restrictions on sub-leasing, audit rights and termination triggers; lenders financing GPU-backed assets are likely to raise haircuts. Within a year, export control is likely to extend from chips to compute services, with registration or licensing regimes for intermediaries in Singapore and Malaysia, modest upward pressure on regional compute prices and further concentration.

For ordinary users, the first three months feel like compute getting no cheaper: fewer cut-rate cross-border pools, possible small price

Reactions run along three lines. Among competitors, suppliers of Nvidia alternatives are shifting their Southeast Asia pitch toward auditable supply chains, using compliance transparency as a differentiator; regional data centre operators have kept public statements restrained, with activity concentrated in customer due diligence and contract terms. AMD and Huawei have not commented publicly on the case.

Among specialists, export-control lawyers focus on the cost-benefit of voluntary disclosure: self-reporting can reduce penalties but hands internal investigation records to prosecutors. Governance researchers read Huang's resignation as the product of pressure for board independence,

Questions worth exploring:

  • How export control extends from governing chips to governing compute services, and where the filing threshold lands
  • Where the governance boundary sits between a founder-CEO's resignation and affiliated suppliers
  • How far Singapore's transshipment model can run now that a criminal precedent exists

Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.

This item comes from The Verge AI (published 2026-11-17 16:19:03). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.


Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.

Republishing: feel free to quote or repost with a link back to the original article on Wenfeng Radar. · Business & partnership: see our media kit
Daily briefing + Sunday review · Unsubscribe anytime by replying · Your email stays private

Leave a Reply

Your email address will not be published. Required fields are marked *

中文EN
Contact · Privacy · Disclaimer · AI Disclosure · Weekly Review · Media & Business
Email: wenfengleida@wenfengleida.cn