The Jensen Huang Affair and the $36 Billion Business Behind Nvidia’s AI Boom

This story was originally published on MartechAI.com

In Nvidia’s fiscal 2025 annual report, Singapore accounted for 18 per cent of global revenue by billing address. Shipments to Singapore represented less than 2 per cent. Both numbers came from the same footnote. The gap helps explain how the AI chip trade works. A company can pay for chips in one country, install them in another and rent access to customers in a third. 

On November 8, Nvidia’s board opened an independent investigation after reports that its founder and CEO, Jensen Huang, had been targeted by a foreign intelligence effort. In a memo to employees, Huang wrote, “I was unfaithful to my wife,” and pledged to cooperate. He did not name the other person. News reports linked the affair to Alice Huang, a former senior executive at Megaspeed International, a Singapore-based cloud provider and Nvidia customer.

The commercial connection gives the story consequences beyond the executive suite. In the quarter ended July 26, 2026, Nvidia reported $89 billion in data centre revenue, about 93 per cent of its $96.2 billion in total revenue. Some of its buyers rent out the computing power those systems produce. That puts an intermediary between the company that makes the chips and the people who ultimately use them.

What the board is investigating

On October 29, The New York Times reported that US intelligence officials believed Alice Huang had passed information to Beijing. The Wall Street Journal reported a similar US assessment. China’s foreign ministry rejected the allegation. Alice Huang did not respond to the Times when it first reported the relationship, according to Reuters.

Nvidia initially described Jensen Huang’s contacts with Alice Huang as routine interactions with a customer in a statement to ABC News. On November 8, the board said it had “no reason to believe” national security or company secrets had been compromised. Independent directors, assisted by outside counsel, will oversee its inquiry. The board expressed confidence in Huang and his leadership team. The public record leaves open what, if anything, was shared and whether the relationship described in the reports intersected with company decisions. 

Singapore drew a distinction of its own. On October 31, its Ministry of Home Affairs said authorities had conducted no separate investigation into Alice Huang’s alleged intelligence ties, disputing part of the Journal’s account. Singapore police confirmed an ongoing investigation into Megaspeed. The company has denied wrongdoing and did not immediately respond to Reuters on November 9.

The scrutiny of Megaspeed’s business goes back further. In October 2025, US Senator Tom Cotton asked the Commerce Department to investigate the company after the Times reported that Nvidia chips had gone to Malaysia and Indonesia, where they appeared to serve customers in China remotely. Nvidia told Reuters at the time that it had made its own inquiries and had found no reason to believe its products were diverted. 

Following the chip and the customer

Nvidia’s filings show how easily a sales figure can be mistaken for a map of installed capacity. In the quarter ended July 27, 2025, 22 per cent of its revenue was billed to Singapore. US-based customers accounted for more than 99 per cent of the revenue billed there for data centre computing equipment subject to export controls, the company said. In its next quarterly filing, Nvidia switched to assigning sales by customer headquarters and restated earlier figures. The company still says the shipping destination and end customer may differ from the headquarters of its direct buyer. 

That distinction matters to regulators. A physical shipment produces freight and customs records. A customer can rent access to a cluster abroad while the hardware stays in place. In January, the US House passed the Remote Access Security Act, a bill to give regulators explicit authority to control certain forms of remote access. An inquiry into a shipment can follow serial numbers and delivery records. Tracing who rented the resulting capacity can require contracts and access logs held by a cloud operator.

Even a warehouse inspection has limits. In a separate indictment unsealed in March, US prosecutors alleged that intermediaries presented false paperwork naming a Southeast Asian buyer as the end user of AI servers destined for China. Ahead of an inspection, prosecutors say, a broker and a co-defendant used a hair dryer to move labels and serial number stickers onto dummy equipment. 

The rules governing physical sales can also change by chip and customer. In January, the US Commerce Department moved Nvidia’s H200 and certain comparable chips to case-by-case licensing review for China and Macau, subject to conditions. In August, Nvidia said it had shipped only a fraction of what its US H200 licences allowed, citing Chinese restrictions as well. The shipments accounted for less than 1 per cent of its quarterly data centre revenue. Nvidia took a $400 million charge for excess H200 inventory and purchase obligations in the first half of the fiscal year. 

Who pays for the capacity

Nvidia’s latest quarterly filing describes another reason to look closely at the cloud operators between chipmakers and end users. Under some agreements, cloud companies buy Nvidia systems and Nvidia commits to buying cloud services from them. The operators can redirect that capacity to third-party customers at more favourable rates. Nvidia put those commitments at $36 billion as of July 26 and said the amount falls as it or third parties use the capacity. 

That figure concerns Nvidia’s wider cloud partner business. It shows how a chipmaker can help underwrite the market for its own hardware. A cloud provider can finance a cluster before filling its order book, and the supplier has a stake in how quickly the capacity finds users. If regulators require more extensive checks on who ultimately uses those clusters, operators that depend on renting capacity quickly could face longer onboarding times and higher compliance costs.

There is a governance question inside that business model. Huang has led Nvidia since co-founding it in 1993. He also made the case to Washington for allowing certain chip sales to China. The board’s review can establish how the company handled any decisions touching the customer at the centre of the reports. Which executives could approve an allocation? Who could stop a sale if compliance staff raised a concern? A record of those decisions would show how much authority sat with one executive and what checks others could exercise.

What control means for India

On the day Nvidia announced its inquiry, it also announced a partnership with South Korean chip designer Rebellions and SK Telecom. The plan pairs Korean-designed processors with Nvidia’s networking and rack technology. SK Telecom intends to deploy the systems in its own data centre. It is a practical example of how a country can build more AI capacity at home while relying on a foreign supplier for parts of the system.

India faces the same distinction. The government says the IndiaAI Mission expanded shared capacity to more than 45,000 graphics processing units by June 2026. Separately, Yotta Data Services announced plans in February to deploy 20,736 Nvidia Blackwell Ultra GPUs in India. For a startup training a model for Indian languages or a business running a voice agent locally, usable capacity depends on which chips are available, how long they can be booked and how much work it takes to move a system if a supplier or cloud contract changes. 

How compute reaches the ad budget

The advertising industry offers a direct view of what those infrastructure choices buy. In February, Meta announced a multiyear infrastructure partnership with Nvidia. A month later, it said it was running hundreds of thousands of its own accelerators for inference across organic content and ads. In October, Meta’s engineers described how spare computing capacity lets the company score more candidate ads or route more requests to a more powerful ranking model. The sophistication of the system deciding which ad gets seen can vary with how much computing power the platform has available. 

A large media platform can spread the cost of running its models across its ad inventory. A martech vendor that pays for each model call has to recover that cost from a smaller client base. Agencies and brands usually encounter this infrastructure through software subscriptions. A retailer using one vendor to generate catalogue images and another to run a customer service agent may be buying from two companies that depend on the same cloud region or model provider. If that shared capacity tightens, the effects can show up as slower responses, fewer images produced in an hour or higher usage charges. The invoices tell the buyer little about whether the services share a point of failure.

The chairs of three US House committees have asked Huang for a closed-door briefing on November 18 and requested a response by November 11. They want to hear about his interactions with Alice Huang and Nvidia’s safeguards against espionage. 

For the wider AI industry, the records worth watching are approvals for any deals under review, the destinations of the equipment and evidence of who could use it. Those records would show how far Nvidia’s oversight extends from the sale of a server to the moment a cloud customer uses it to run a model.

Disclaimer: All data points and statistics are attributed to published research studies and verified market research. All quotes are either sourced directly or attributed to public statements.