Trump-Xi pre-summit watch: H20 setbacks and Ascend gains reshape China's Nvidia calculus

The pivot isn't volume—it's necessity. H20's stalled commercial rollout, squeezed between export friction and a performance ceiling, has inadvertently widened the window for Ascend 910-series penetration in government and enterprise inference workloads. The decisive variable isn't silicon; it's the developer migration curve from CUDA to CANN. That curve, over the next 12 months, determines whether ecosystem bifurcation becomes structurally irreversible. On the technical stack, HBM3e supply and CoWoS advanced-packaging capacity remain shared chokepoints. The 2.5D packaging gap between TSMC (Taiwan, China) and SMIC cannot be closed by policy alone within this cycle. Compliance costs keep compounding for Nvidia with each regulatory iteration, while Huawei's self-sufficient chain still faces hard constraints in yield and HBM self-sufficiency. Competitively, AMD and Intel are positioning as 'compliance-friendly' alternatives but lack ecosystem depth. Domestic players like Cambricon and Moore Threads are accelerating inference-side fill-in. Outlook: the next 18 months will likely crystallize a dual-track architecture—imported silicon for training, domestic for inference. Software-stack fragmentation will slide from reversible to structural. The summit will produce managed decoupling, not reopening.