🧠 Food for thought

Implications, context, and why it matters.

ByteDance uses a gap in export rules for AI training

• Chinese companies such as ByteDance train advanced AI models in Southeast Asian data centers to get access to high-end Nvidia GPUs 1.
• U.S. export controls stop direct sales to China. Renting compute from foreign-owned facilities overseas is still allowed, and a draft rule that would have treated overseas leasing as an indirect violation was later withdrawn 1.
• After offshore training, model weights can move back to China for inference. More inference workloads there now run on locally made chips from companies like Huawei 1.
• ByteDance reportedly plans to spend $7 billion in 2025 to secure access to Nvidia chips outside China 2.

A $7 trillion buildout is forming around AI infrastructure

• AI hardware needs are driving a worldwide construction boom. One estimate puts required capital at $7 trillion over the next decade across the AI value chain, including power, compute, and data centers 3.
• Nvidia is putting money into the plumbing that keeps its products moving, with billions going into its supply chain. That includes a $4 billion commitment to optical-component suppliers Lumentum and Coherent, makers of parts used in networking gear that connects data centers 4.
• The company is also backing new AI-focused cloud providers and working with asset managers (large investment firms) on $100 billion programs to build “AI factories” that run on its chips 5.
• Governments are adding another pull on supply. India, for example, is rolling out billion-dollar “sovereign AI” missions (government-backed programs to build domestic AI infrastructure) aimed at building local capacity, which can lift orders for Nvidia hardware 6.