Moonshot AI reportedly accessed restricted Nvidia chips in China despite US ban

The real story isn't a leak — it's that US export controls carry a structural enforcement failure baked into their economics. A single H100 routed through a third-country distributor commands a $200K–$400K margin. Compliance costs are a rounding error against that spread, so the gray channel won't die; it will simply get more layered. Technically, Moonshot's Kimi training stack remains architecturally welded to CUDA and NVLink interconnect bandwidth. Huawei's Ascend 910B still trails in HCCS throughput and operator-library maturity. Domestic substitution works at inference scale; at 10K-GPU training clusters, the lock-in persists. The regulator-filed documents signal institutional procurement, not a one-off. That directly pressures Nvidia's data-center revenue, where China still represents roughly 20%. Expect Washington to escalate from chip-level to architecture-level controls — targeting EDA tooling and CoWoS advanced packaging rather than just die shipments. 12–24 month outlook: decoupling enters its deep-water phase, but a 2–3 year generational gap won't compress on policy alone. The real swing variable isn't who bans whom — it's whether SMIC's 5nm yield crosses the mass-production threshold before 2026. Until then, every 'domestic' AI lab is quietly running on American silicon.