Key Takeaways

  • The founder launched MiroMind in California while distributing development across Singapore, Beijing and Shanghai.
  • Semiconductor controls, uneven computing capacity and data-transfer rules complicate MiroMind’s multinational structure.
  • AI businesses spanning the US and China increasingly need localized infrastructure, data and governance arrangements.

The founder launched California-based AI startup MiroMind last year with an unusually distributed operating model. Its teams in Singapore, Beijing and Shanghai are working across jurisdictions that offer different pools of engineering talent, computing infrastructure and data, but also increasingly divergent technology regulations.

The setup captures a wider challenge facing leaders trying to connect Chinese AI expertise with US capital and infrastructure. MiroMind may be incorporated in California, yet corporate location alone does not resolve questions about where models are trained, which employees can access sensitive technology, or whether data can move between development centers. Investment controls add another layer.

Compute is the most immediate constraint. Advanced AI development depends on large clusters of accelerators, high-capacity data centers and substantial capital spending. According to Boston Consulting Group, US technology companies invested more than $400 billion in capital expenditure during 2025, compared with $63 billion in China. US data-center capacity exceeded 50 GW, while China had 31 GW.

That gap shapes where a startup such as MiroMind can efficiently train and deploy large models. Nearly 75% of advanced AI-computing clusters were located in the United States in 2025, compared with 14% in China. Access, however, is not purely a matter of choosing an American cloud region. Restrictions covering NVIDIA’s advanced accelerators can affect procurement, employee access and the transfer of technical knowledge to teams connected with China.

China is still spending heavily. Rhodium Group projects the country’s AI infrastructure capital expenditure will reach RMB932 billion ($139 billion) in 2026. Yet chip shortages and financing constraints remain binding limitations. China’s AI infrastructure spending also fell 8.1% in Q4 2025 amid export controls on advanced semiconductors, even as worldwide spending is projected to reach $487 billion in 2026, approximately 53% higher year over year (source).

More domestic spending does not immediately recreate the most advanced supply chain. Data-center buildings, electricity and conventional servers can be added relatively quickly. Leading accelerators, networking equipment, software tooling and experienced cluster operators are harder to assemble at scale. Larger Chinese companies such as Alibaba may have comparatively greater access to overseas data centers. MiroMind has less room to absorb duplicated infrastructure and compliance costs.

Data presents a different kind of border. China’s Personal Information Protection Law and Data Security Law regulate the handling and export of personal and potentially important data. The country’s 2024 cross-border data-flow provisions and 2025 certification regime add procedural layers, particularly for businesses processing sensitive information. In Europe, the EU AI Act creates another governance environment for companies serving or operating in the region.

Research from the Atlantic Council notes that Chinese rules covering personal, important, health and genomic data can push AI developers toward localized datasets, domestic training or federated analysis. That matters beyond regulated healthcare projects. Model developers often need to document the origin, permitted use and geographic movement of training and evaluation data.

Can one development organization still function coherently when its compute, code access and datasets are divided by jurisdiction? Possibly, but the architecture changes. MiroMind may need separate development environments, regional access controls and carefully defined interfaces between its teams. Some model components could be trained locally, while less sensitive research is shared through controlled repositories.

That said, distributed operations also offer advantages. Singapore can serve as a regional engineering and governance hub, while Beijing and Shanghai provide access to China’s deep technical workforce. California places MiroMind closer to US computing capacity and the broader AI investment ecosystem. The trade-off is operational friction.

For the founder, MiroMind is therefore more than another AI venture. It is a test of whether a company can preserve the benefits of multinational research while technology blocs pull further apart. The likely result is not a fully unified global development stack, but a more compartmentalized business that treats compute, data rights and employee access as core design decisions from the start.