Chinese technology titans are circumventing stringent US semiconductor export restrictions by relocating their most advanced artificial intelligence model training operations to data centers in Singapore, Malaysia and other Southeast Asian nations, creating a new battleground in the escalating US-China technology rivalry.

Alibaba and ByteDance are among the prominent Chinese firms routing training jobs for their latest large language models through foreign-owned data centers equipped with high-performance Nvidia GPUs that cannot be legally exported to mainland China, according to industry sources with direct knowledge of the arrangements. The offshore strategy has accelerated since April 2025, when the Trump administration tightened restrictions on Nvidia's H20 chip, a downgraded accelerator designed specifically for the Chinese market.

Legal Loophole Fuels Southeast Asian Data Center Boom

The maneuver exploits a significant gap in US export control regulations. While Washington prohibits the direct sale of cutting-edge AI chips to Chinese entities, current rules do not prevent non-Chinese data center operators in approved countries from purchasing those processors and reselling access as cloud services to Chinese clients.

A Singapore-based data center executive described the arrangement as straightforward business logic. Chinese customers seeking state-of-the-art hardware and regulatory compliance find offshore clusters an obvious choice, the executive explained. These facilities typically deploy Nvidia accelerators near the top of the performance spectrum, including H100 and A100 families, rather than the constrained China-specific variants now facing heightened scrutiny.

The Biden administration attempted to close this loophole through a proposed "diffusion rule" that would have treated overseas leasing of restricted compute to Chinese customers as export control violations. However, the Trump administration rescinded that framework in May 2025 before implementation, creating renewed opportunities for Chinese firms to access advanced hardware through third-party arrangements.

Data center capacity has surged throughout Southeast Asia in response to Chinese demand. Singapore and Malaysia have emerged as primary hubs, with facilities expanding rapidly to accommodate high-density racks of Nvidia GPUs connected through low-latency networking infrastructure. Chinese companies typically sign long-term lease agreements with local or international operators who retain legal ownership of the hardware, maintaining a structure that keeps arrangements technically compliant with existing export regulations.

Billion-Dollar Investments Signal Strategic Shift

The scale of Chinese investment in offshore compute access reflects the strategic importance companies place on maintaining AI competitiveness despite geopolitical constraints. ByteDance, owner of TikTok, reportedly planned to spend up to $7 billion accessing Nvidia chips through overseas servers in 2025, though the company has disputed specific figures as inaccurate.

Multiple sources indicate ByteDance co-founder Zhang Yiming has been negotiating with data center operators across Southeast Asia and other regions to secure access to Nvidia's Blackwell chips, the latest generation of AI accelerators. The planned investment would position ByteDance as one of Nvidia's largest global customers, comparable in scale to major American technology companies.

Industry analysis suggests ByteDance ordered approximately 230,000 Nvidia Hopper chips in 2024, ranking as the second-largest global purchaser behind Microsoft, which acquired 485,000 units. This substantial procurement occurred despite export restrictions, achieved through a combination of stockpiling degraded H20 chips permitted for sale to China and accessing more powerful processors through overseas data center agreements.

Alibaba and Tencent have similarly expanded their offshore training operations. Over the past year, Alibaba's Qwen and ByteDance's Doubao models have climbed into the top tier of global large language model benchmarks, with testing indicating performance rivaling Western systems developed by OpenAI and Anthropic.

Elaborate Workarounds Include Data Smuggling

Some Chinese AI companies have resorted to even more elaborate methods to train models abroad while adhering to data sovereignty requirements. In one documented case, four Chinese technology workers flew from Beijing to Kuala Lumpur, each carrying 15 hard drives containing 80 terabytes of training data, totaling approximately 4.8 petabytes of information.

The meticulously planned operation required months of preparation. Engineers chose physical data transport because online transfers would attract attention and consume excessive time. Dividing hard drives among multiple passengers helped avoid triggering alarms with Malaysian customs and immigration officials. Upon arrival, personnel proceeded to a Malaysian data center where their company had rented 300 Nvidia AI servers to process the data and construct AI models.

The Chinese firm had previously used the same process to train models through a Malaysian data center, with a Singapore-registered subsidiary signing rental agreements. When Singapore authorities intensified scrutiny of AI technology exports, the Malaysian data center operator requested the Chinese client register locally to reduce regulatory exposure.

Domestic Alternatives and Strategic Exceptions

Not all Chinese AI developers are pursuing offshore strategies. DeepSeek, a Shanghai-based firm known for producing high-quality, cost-efficient models, continues training operations domestically using a substantial stockpile of Nvidia accelerators accumulated before the most stringent export bans took effect.

DeepSeek has also established close partnerships with domestic chipmaker Huawei to optimize both hardware and software stacks for future training runs. Reports indicate Huawei has stationed a team of engineers at DeepSeek's Hangzhou headquarters, viewing the collaboration as strategic for accelerating adoption of its AI-focused semiconductors across Chinese training clusters.

The company achieved international attention in January 2025 by releasing its R1 model, reportedly trained for just $5.6 million using approximately 2,000 Nvidia H800 GPUs. The achievement demonstrated that architectural innovations and training efficiency could partially compensate for hardware limitations, producing results competitive with far more expensive Western models.

Persistent Preference for Nvidia Despite Domestic Alternatives

Despite significant government pressure and investment in domestic semiconductor alternatives, Chinese AI developers overwhelmingly prefer Nvidia hardware, even severely performance-degraded variants, and pursue elaborate strategies to access it. Industry analysis indicates Chinese companies purchased approximately 1 million Nvidia H20 chips in 2024, compared to estimated shipments of 450,000 Huawei Ascend 910B chips.

Only a handful of state-backed companies in China have adopted Huawei chips for model training, including iFlytek, SenseTime and China Mobile. Many Chinese firms reportedly view Huawei's offerings as lagging behind foreign counterparts in key performance dimensions, despite steady improvements.

In anticipation of US restrictions on Nvidia's H20 chips, ByteDance, Alibaba and Tencent collectively rushed to spend $16 billion stockpiling approximately 1.3 million to 1.6 million H20 units before the ban took effect. The massive procurement underscored the perceived superiority of even downgraded Nvidia products over domestic alternatives.

Chinese technology companies have also scoured black markets across Asia and e-commerce platforms to acquire banned Nvidia chips for as much as double their normal prices. Some buyers have resorted to purchasing Nvidia's RTX gaming chips as substitutes, despite those products not being designed for AI workloads, while others smuggled hard drives containing training data out of the country to run on servers in permitted jurisdictions.

Geopolitical Implications and Policy Responses

The offshore training phenomenon highlights fundamental challenges in implementing effective technology export controls in an interconnected global economy. While the United States maintains approximately a tenfold advantage over China in total compute capacity, Chinese firms have demonstrated remarkable adaptability in circumventing restrictions through legal loopholes, stockpiling, smuggling and architectural innovation.

Former Google CEO Eric Schmidt publicly acknowledged surprise at Chinese AI progress despite chip restrictions. In May 2024, Schmidt confidently asserted the United States maintained a two-to-three year AI lead. By November, after observing advances from Alibaba, Tencent and DeepSeek, Schmidt revised his assessment, stating he was shocked that chip restrictions had failed to keep Chinese capabilities back.

US policymakers face difficult tradeoffs in responding to offshore training strategies. Expanding restrictions to cover cloud service leasing in third countries could damage relationships with Southeast Asian allies and harm American cloud providers seeking to serve legitimate international customers. However, allowing the loophole to persist enables Chinese access to cutting-edge hardware that export controls explicitly aim to deny.

The Commerce Department has acknowledged enforcement challenges. Recent congressional investigations found the department significantly underfunded and ineffective at preventing export control violations. Despite incremental improvements in enforcement, the volume of smuggling and creative workarounds indicates substantial implementation gaps remain.

Industry Innovation and Long-Term Trajectories

While offshore strategies provide Chinese firms short-term access to advanced hardware, they introduce operational complications and costs compared to domestic training. Companies must navigate complex legal structures, manage cross-border data transfers within sovereignty constraints, and accept higher infrastructure expenses in premium Southeast Asian markets.

These frictions have driven parallel investment in domestic semiconductor development and algorithmic efficiency. Huawei's Ascend chip series continues advancing, with the company projected to sell over 1 million accelerator units in 2025. While still trailing Nvidia's latest offerings by approximately four years in raw performance, Ascend processors provide an increasingly viable fallback option for Chinese firms facing supply uncertainties.

Chinese companies have also excelled at algorithmic innovation to maximize performance from constrained hardware. Techniques such as Mixture-of-Experts architectures, Multi-Head Latent Attention and novel training optimizations have enabled Chinese models to achieve competitive benchmark results despite using less powerful chips and smaller training budgets than Western counterparts.

DeepSeek's achievement training a frontier model for under $6 million exemplified this efficiency-focused approach. The accomplishment suggested that architectural sophistication and training methodology could partially overcome brute-force compute advantages, potentially altering the economics of AI development globally.

Domestic Chip Mandate Reshapes Market Dynamics

Beijing has intensified pressure on Chinese technology firms to adopt domestically produced semiconductors. Recent regulations mandate that any new data center receiving government funding must rely exclusively on locally developed chips. Early-stage projects must remove foreign chips or cancel purchase plans, while facilities already exceeding 30 percent completion face case-by-case reviews.

The mandate largely excludes Nvidia and other foreign chipmakers from lucrative government-backed infrastructure projects, despite advanced models under US controls continuing to appear in China through informal channels. The policy aims to accelerate development of domestic semiconductor capabilities while reducing strategic dependence on foreign technology suppliers vulnerable to geopolitical disruption.

However, the regulation creates tensions between nationalist industrial policy and commercial pragmatism. Private Chinese firms generally prefer foreign chips for superior performance and established software ecosystems, while state-owned entities face political pressure to support domestic alternatives even when technically inferior.

Technological Competition

The offshore training phenomenon represents an interim phase in the broader US-China technological rivalry. As AI training requirements scale from clusters of thousands to hundreds of thousands of chips, hardware quality gaps may impose increasing constraints on Chinese capabilities. A cluster of 100,000 Nvidia B200 chips might require a Chinese equivalent of 300,000 Ascend 910C processors, substantially increasing energy consumption and engineering complexity.

Compounding hardware disadvantages is Nvidia's sophisticated software and networking ecosystem, crucial for orchestrating massive distributed training runs and currently unmatched by Chinese alternatives. While Chinese firms have made impressive strides in model architecture and training efficiency, the cumulative effect of multiple technical gaps could widen over time if export controls successfully limit access to cutting-edge manufacturing equipment and design tools.

Conversely, continued policy inconsistency and enforcement failures could enable Chinese firms to maintain approximate parity through hybrid strategies combining offshore access, stockpiled hardware, domestic alternatives and algorithmic innovation. The rescission of the diffusion rule and delayed enforcement actions suggest political and economic pressures complicate maintaining coherent, sustained technology restrictions.

The ultimate trajectory depends on multiple factors: whether the United States and allies can coordinate effective multilateral export controls, how quickly Chinese domestic semiconductor capabilities mature, whether architectural innovations continue compensating for hardware limitations, and the political sustainability of costly technology restrictions amid competing economic interests.

For Southeast Asian nations, the competition creates both opportunities and risks. Data center investment brings economic benefits and positions countries as crucial infrastructure hubs in the global AI ecosystem. However, being caught between competing American and Chinese interests creates diplomatic challenges and potential exposure to economic pressure from both powers seeking to advance their strategic technology objectives.