China's AI labs are renting Nvidia's banned chips from data centres in Southeast Asia

Key points

• Moonshot AI, maker of the Kimi K3 model, was accused by White House official Michael Kratsios in July of accessing Nvidia GB300 chips through a data centre in Thailand.
• US export controls ban the physical sale of advanced Nvidia chips to China but do not currently cover remote access to those chips via overseas cloud providers.
• ByteDance reportedly accessed Nvidia compute in Malaysia through Singapore-based cloud provider Aolani, first reported by The Wall Street Journal in March.
• A proposed US law called the Remote Access Security Act (RASA) passed the House of Representatives in January 2025 but has not yet cleared the Senate.
• DC Byte data shows 31 data centres of 100 megawatts or more are planned across Malaysia, Indonesia and Thailand, compared to just two today.

---

Imagine banning a country from buying a particular weapon, only to find it can rent the exact same weapon from a neighbour. That is, roughly, the situation Washington now faces with its AI chip export controls.

Nvidia, the California company whose processors power most advanced AI systems worldwide, sits at the centre of US efforts to slow China's AI progress. Its latest GB300 chips cannot legally be shipped to Chinese buyers. No chips, no training powerful AI models. Except there is a gap.

How are Chinese companies getting around the ban?

US rules control who can own the chips physically. They say nothing about who can use them remotely over the internet.

Data centres, large warehouses packed with thousands of Nvidia chips, have been built across Thailand, Malaysia and Japan. Chinese AI companies can buy time on those machines and train their models without ever touching the hardware. The chips never cross into China. No law is broken.

Cassia King, senior researcher on the Compute Policy team at the Institute for AI Policy and Strategy, confirmed to CNBC that Moonshot's reported use of a Thai facility was legal "so long as Moonshot isn't actually buying and owning the physical hardware directly."

Moonshot's Kimi K3, released in July, is one of a wave of strong recent Chinese AI releases. As we reported on 10 August, Chinese models are cheaper, widely adopted, and closing the gap with American frontier labs faster than most people expected. DeepSeek and Alibaba have both launched new systems that scored well on performance benchmarks, the industry's standard way of measuring model capability. Critics of current US policy argue that remote access to top-tier chips is a key reason those models keep improving.

Michelle Nie, a visiting fellow in technology and national security at the Center for a New American Security, called the loophole "threatening US national security," telling CNBC the whole point of the controls was to "deny China the ability to train frontier AI using advanced US chips."

What is the US trying to do about it?

The Remote Access Security Act (RASA) would stretch export controls to cover cloud-based access to critical chips, not just physical ownership. It passed the House in January 2025 but has stalled in the Senate.

Even if RASA passes, it only hands the government authority to write new rules. The Bureau of Industry and Security (BIS), the Commerce Department body that enforces export controls, would still need to draft those rules separately. King told CNBC that process could theoretically move in "a matter of days" with White House backing, but warned the hard part is writing rules that actually work: deciding which chips are covered, which users are blocked, and how cloud providers verify who their customers really are. Cloud companies would carry much of that verification burden, and industry pushback is expected.

What does this mean for ordinary people?

Directly, not much yet. But the broader race shapes which AI tools reach consumers and at what cost. If Chinese labs can train world-class models cheaply using rented compute, competition intensifies and AI service prices could fall. The security concern runs separately: powerful AI trained outside US oversight could eventually be used in ways touching everything from cyber security to financial markets.

ByteDance, whose own compute ambitions we covered on 7 August, was reportedly working with Singapore-based cloud provider Aolani to access Nvidia chips in Malaysia, a story The Wall Street Journal broke in March. Aolani told CNBC its customers have no ownership or physical access to its chips and that all services comply with applicable regulations. With 31 large data centres planned across the region against just two today, this debate will only grow louder.

Common questions

Is what these Chinese companies are doing actually illegal?

No, under current US law. Export controls apply to the physical chips, not to renting computing time on chips located outside China. CNBC reported that a senior researcher confirmed the arrangement is legal as things currently stand.

Would the Remote Access Security Act fix this immediately?

Not on its own. RASA would give the US government the power to regulate remote chip access, but regulators would still need to write separate rules spelling out who is banned, which hardware is covered, and how companies must verify their customers' identities before those rules bite.