DeepSeek and Huawei Open Ascend Software to Build an Alternative to Nvidia - Межа. Новини України.

Huawei opening the Ascend stack to DeepSeek is not an 'Nvidia killer' story—it's an ecosystem leverage play against a decade-long CUDA lock-in. Letting third-party models run natively on CANN shifts Huawei from selling silicon to selling a platform, mirroring Android's 2007 open-source gambit: get developers coding against your ISA, then make your hardware the only viable target. The real disruption hits inference first. Ascend 910-series FP16 throughput already reaches roughly 80% of H100, and open-sourcing the stack collapses adaptation costs from enterprise-grade to community-grade. Upstream, HBM and advanced packaging procurement will see its first meaningful non-Nvidia order diversion. But for training beyond 7B parameters, HCCL's distributed communication still lags NVLink by a generation. Nvidia's moat is not the die—it's two decades of operator libraries. On compliance, an open software layer may trigger additional US scrutiny on 'technology transfer,' while advanced-node foundry dependence on Taiwan, China and mainland lines keeps yield and capacity as core variables. Strategically, AMD's MI325X is most exposed, likely repositioning as the 'non-Huawei second option' for Western compliance markets, while Cambricon and Hygon face ecosystem cannibalization. 12-24 month call: inference markets will hit a 'de-Nvidia' inflection; training stays Nvidia-dominated. The global landscape shifts from 'one superpower' to 'one superpower, one challenger'—but CUDA migration costs mean that challenger needs at least three years to solidify its footing.