H200s Now Allowed to China — For Now. Malaysia's Stability Just Became a Feature.
The BIS reversed course on advanced AI chips in January 2026, letting H200-class exports through on a case-by-case basis. Congress wants it reversed by summer. For Chinese enterprises building AI businesses overseas, the lesson isn't about the policy — it's about not needing it.
A policy that is already mid-reversal
On January 15, 2026, the US Bureau of Industry and Security quietly changed posture on advanced AI chips bound for China. Previously, export license applications for products like the Nvidia H200 and AMD MI325X were evaluated under a presumption of denial. Under the January rule, applications are reviewed case-by-case — a small phrase, an enormous shift. Analysts estimate the change could cap H200 exports to China at roughly 1 million units, with a comparable headroom for H100s. The net installed AI compute in China could rise by 250% relative to a domestic-only baseline.
That news landed in Shenzhen and Hangzhou boardrooms as a signal that the worst of the export squeeze was easing. Then, within weeks, the AI Overwatch Act was introduced in the House, passing committee, and giving Congress a 30-day review window on any AI export license — an effective legislative veto on the executive branch’s new permissiveness.
Nobody we speak to in KL or Singapore believes the January rule will remain untouched by the end of Q3 2026.
The cost of building on regulatory sand
For a Chinese enterprise doing overseas AI work, the last eighteen months have been an education in jurisdictional fragility. Sanctions-dependent business models — smuggled GPUs, grey-market cloud reselling, back-channel model access — all share a single structural flaw: their unit economics are one Federal Register notice away from breaking.
The companies that have quietly continued operating through every policy swing have one thing in common: they don’t touch controlled goods in the first place. They buy Mac minis, which aren’t controlled. They host in Malaysia or Singapore, which aren’t the target jurisdiction. They run their compute on CPUs and consumer-grade accelerators that no one is trying to restrict. They use these not because they’re optimal for training frontier models — they aren’t — but because their legal surface area is small and their policy risk is near zero.
What Malaysia buys you that Shenzhen cannot
Three specific things:
- Jurisdictional distance from US export control. A Malaysian Sdn Bhd running AI inference workloads on commodity hardware is a non-event to BIS. Every chip in the stack is legally free to move.
- Access to the enterprise tier of Anthropic, OpenAI, and Google. These vendors operate in Malaysia. They do not operate, contractually, inside mainland China. The difference is everything.
- A neutral billing address for overseas SaaS and ad platforms. Google Ads, Meta, Stripe, Shopify — all of which apply elevated scrutiny to mainland-issued cards and IPs — treat a Malaysian commercial entity as a normal corporate customer.
Our position
The BIS January rule and the AI Overwatch Act will spend the rest of 2026 grinding against each other in Congress. We have no prediction on the outcome, and no need for one. MalakaToken’s architecture — physical Mac mini nodes in a Tier-III KL facility, native Malaysian commercial IPs, Sdn Bhd billing, and enterprise-tier AI service subscriptions in our customers’ legal names — doesn’t move when that fight moves.
If your AI overseas strategy requires a specific export license to remain viable, your strategy was always fragile. Malaysia is the other answer: a route that doesn’t care which way Washington swings this quarter.
Sources: BIS Final Rule effective January 15 2026 (Morgan Lewis summary); AI Overwatch Act reporting (The Register, CFR). MalakaToken operates commodity compute in a non-sanctioned jurisdiction and is not a party to any export-controlled transaction.