Nvidia chip curbs turn Singapore into AI hub for China

The United States’ export controls aimed at restricting China’s access to Nvidia’s advanced artificial intelligence chips are unlikely to stop Chinese firms from upgrading their AI models, according to some data‑center experts.

When the Trump administration announced last December that it would approve exports of Nvidia’s H200 AI chips to China, it was widely expected that Beijing would welcome the move. Instead, Chinese authorities asked major technology companies to first consider using domestically developed chips.

Industry experts say Chinese technology firms can easily access Nvidia’s high‑end chips in Southeast Asian data centers, either by buying computing power or installing their own server cabinets, allowing them to train AI models overseas before deploying them in China.

Gary Wojtaszek, a director at GDS Holdings, a major developer and operator of high‑performance data centers in China, told Asia Times in an interview in Dublin that US chip export controls are limited in scope because they restrict chip shipments but not the movement of AI data or trained models.

“The US chip export rules are pretty nebulous. In China, you can’t import chips, but you can export your data to train your model. Then you import it back,” Wojtaszek said.“Water always finds the level playing field. People may try to control or slow a trend, but it will eventually find a natural balance.”

He said China would ultimately overcome such restrictions either through domestic technological development or by sourcing computing resources elsewhere, noting that enforcing global chip controls is inherently difficult because hardware can easily change hands across borders.

Wojtaszek said his global perspective is shaped by his exposure to the data‑center industry worldwide.

• In August 2011, he joined CyrusOne Inc, a real estate investment trust that builds and manages carrier‑neutral hyperscale data centers globally, as its president and chief executive officer.
• In October 2017, CyrusOne announced a US$100 million equity investment to acquire an approximately 10% stake in GDS Holdings, and since June 2018, Wojtaszek has served as a member of GDS’s board of directors. He said he then became a good friend ofGDS founder and Chief Executive William Wei Huang.
• In February 2020, he resigned from CyrusOne, but he continues to serve GDS.
• In 2022, GDS established an international subsidiary, GDS International (GDSI), in Singapore to develop data centers across Southeast Asia and other overseas markets. The unit was rebranded as DayOne Data Centers on January 1, 2025. After several rounds of fundraising, GDS retains a 25% stake in DayOne.

DayOne operates data centers across key markets including Singapore, Johor in Malaysia, Batam in Indonesia, Greater Bangkok, Hong Kong, Tokyo and Finland.

“I’ve been on the board of DayOne Data Centers since its establishment,”Wojtaszek said.“I’ve been friends with William and the team for many years.”

He said he also sits on the board of Nxtra in India, a data center firm in Dublin and several in the US, and advises a data center company in Brazil.

Commenting on the global AI competition, Wojtaszek said the race has so far been dominated by the US and China. He said the US has cutting‑edge graphics processing units (GPUs) and large language model (LLM) algorithms,but China is catching up quickly and also holds advantages in energy production needed to power AI infrastructure.

He said China has far larger data pools for AI training, while the US faces constraints, including stricter data restrictions and privacy requirements.

“DeepSeek is ingenious. They did something interesting at significantly lower cost and with lower power,” Wojtaszek said, adding that China will continue to reinvent other people’s inventions and make improvements to them.

AI Diffusion Rule

In October 2022, the Biden administration banned the export of Nvidia’s H100 and A100 chips to China. In October 2023, it expanded the controls to include the H800 and A800 chips.

In early 2025,the Biden administration introduced the so‑called AI Diffusion Rule to control the global distribution of advanced AI chips and computing power. The rule created a three‑tier framework that categorized countries according to their access to US AI technologies.

• Tier 1: 18 US allies, such as Japan and most European countries, can freely import Nvidia’s most advanced AI chips.
• Tier 2:120 countries, including Singapore, Malaysia, India and the United Arab Emirates, face tighter controls and must obtain individual licenses or validated end‑user (VEU) authorization to access advanced AI chips.
• Tier 3: Countries including China, Russia, Iran and North Korea are considered high risk for diverting or misusing advanced AI chips, particularly for military purposes.

Although the rule was scheduled to take effect last May, the Trump administration withdrew it after concluding that its scope was too broad and it would be difficult to enforce.

Washington now allows exports of Nvidia’s H200 chips to China, but not the most advanced Blackwell GPUs. Meanwhile, Beijing wants domestic data centers to prioritize local alternatives such as Huawei Technologies’ AI processors. As a result, many Chinese firms seek access to Nvidia chips in overseas locations, including Singapore and Johor in Malaysia.

The Wall Street Journal reported last June that four Chinese engineers flew from Beijing to Kuala Lumpur carrying hard drives loaded with AI training data. They reportedly rented about 300 Nvidia AI servers at a Malaysian data center to process the data. The newspaper said the move did not appear to violate any law but illustrated a loophole in US chip export controls.

The Financial Times reported last November that major Chinese companies, including Alibaba and ByteDance, have been training AI models in Southeast Asian data centers to gain access to Nvidia’s high‑end chips.

In other words, whether Nvidia’s most advanced chips can be shipped directly to China may not significantly affect the company’s overall performance. Nvidia’s revenue surged 65% year‑on‑year to US$215.9 billion for the financial year ended January 25, 2026. About one‑fifth of the company’s revenue came from Singapore, making it Nvidia’s second‑largest market after the US.

Meanwhile, new uncertainty has emerged. On January 14, the US House of Representatives passed the Remote Access Security Act, a bipartisan bill that modernizes the Export Control Reform Act by expanding federal authority to restrict foreign adversaries’ ability to access sensitive technologies remotely through cloud computing services.

If the act becomes law, using US AI chips or computing services via overseas data centers could fall under the same export‑control rules that govern hardware shipments.

Demand from Hong Kong

Apart from the impact of US chip controls, another reason Southeast Asian countries have seen rising demand for AI data‑center services is that they are benefiting from a shift in demand away from Hong Kong.

Doug Adams, chief executive and president of NTT Global Data Centers, the world’s third‑largest data‑center operator, told Asia Times in an interview in London that many customers in the region now prefer to place their computing infrastructure in Southeast Asia rather than Hong Kong.

“Hong Kong is a challenging market. I think that most customers in that region, given the choice, prefer to be in Singapore and other parts of that area,” Adams said. “Johor Bahru is becoming a big market. Malaysia and Indonesia are both growing. Thailand is growing quite a bit. All these markets are outgrowing Hong Kong.”

Adams said a shift in global telecommunications infrastructure may have contributed to this trend.

In recent years, several US technology companies and telecom operators have scaled back plans to connect new undersea fiber‑optic cables directly to Hong Kong. After Washington raised national‑security concerns about data flows to mainland China, projects such as the Pacific Light Cable Network were forced to reroute connections away from Hong Kong toward alternative hubs including Taiwan, the Philippines and Singapore.

In addition to connectivity issues, Hong Kong has faced challenges in adopting generative AI services. Several major Western LLMs, including OpenAI’s ChatGPT, Google’s Gemini and Anthropic’s Claude, are not officially available for direct consumer use in Hong Kong.

Read: Nvidia halts H200 production as China backs Huawei AI chips

Follow Jeff Pao on Twitter at@jeffpao3