--- title: "Huawei opens the software behind its AI chips, but keeps the silicon for itself" date: 2026-11-08 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-huawei-ascend-cloud language: en --- # Huawei opens the software behind its AI chips, but keeps the silicon for itself > Huawei pledged to fully open-source the CANN software stack behind its Ascend AI processors by end-2025, betting that a free toolchain — not cheaper hardware — will build China's alternative AI ecosystem. For two years, the hardest part of buying a Huawei AI accelerator was not the chip. It was everything you had to run on top of it. Developers complained that the software was a walled garden with a steep learning curve and few examples. Huawei's answer, delivered in the second half of 2025, was to tear part of that wall down. ## The decision At an Ascend industry summit in Beijing on 5 August 2025, Huawei laid out a four-point response to customer pressure that had built up since DeepSeek's open release strained Ascend 910B and 910C capacity. The headline commitment: open the CANN compiler and its virtual instruction-set interface, and open-source essentially all remaining software, with the CANN open-sourcing for Ascend 910B/C to be completed by 31 December 2025. Huawei also promised that future open-sourcing would happen in step with product launches, rather than months later. Two related pledges followed. The Mind series of application-enablement toolkits and its toolchain would be fully open-source by the same year-end. And the openPangu foundation large model (大模型) would be fully open-sourced. ## What CANN actually is CANN — Compute Architecture for Neural Networks — is the layer that lets a model trained in PyTorch or another framework actually execute on Ascend silicon. Open-sourcing it means outside developers can write custom operators, debug performance and port workloads without waiting for Huawei. At Huawei's Connect 2025 conference in September, Ascend computing president Zhang Dixuan announced a CANN Technical Steering Committee and said full operators would land on the GitCode community by the end of September, with domain acceleration libraries, graph computing, Ascend C and MindIE following in December. ## The twist: hardware stays commercial The catch is structural. Huawei explicitly said it "persists in Ascend hardware monetization" — the chips and systems remain products to be sold, while the software becomes free. This is close to the inverse of a typical open-hardware play. Huawei is betting that a liberated toolchain pulls developers and partners onto Ascend, and that those relationships, not the software license, are what lock in future hardware sales. Put plainly, the company is trading the price of its software for the loyalty of its users: give the layer away free, and the cost of switching accelerators later — rewriting operators, retraining teams, re-porting workloads — becomes the real lock-in. ## Why it matters for the cloud The move strengthens Huawei Cloud's (华为云) pitch: a vertically integrated stack from silicon to service, now with an open programmer surface. By Huawei's 2025 annual report, the Ascend ecosystem counted 4 million developers and more than 9,800 partners, who together built over 26,000 industry solutions. Huawei Cloud said it had helped more than 30 industries and 500 scenarios move toward intelligent upgrade. According to Li An, chief scientist at BrainNet (脑机网), opening the software layer is the cheaper half of building an ecosystem; the harder half is filling it with developers who stay. ## Key takeaways - Huawei committed on 5 August 2025 to open-source CANN (compiler plus virtual ISA) and most Ascend software by 31 December 2025. - The Mind toolkits and the openPangu foundation model were also pledged to full open-source by year-end 2025. - Ascend hardware remains a commercial product; only the software is free. - By Huawei's 2025 report, the Ascend side had 4 million developers, 9,800-plus partners and 26,000-plus solutions. - At WAIC 2026, Huawei said the CANN open community already held 67 projects, 12.44 million lines of code and 3,500 monthly active developers. ## What readers can do now - **Audit your stack for lock-in.** If you run inference on Ascend today, the open CANN release means you can now customize operators instead of waiting on vendor updates. - **Pilot before you commit.** Spin up a small Ascend instance on Huawei Cloud (云计算) and port one workload to measure real throughput against your current Nvidia or AMD baseline. - **Track the December deliverables.** Verify that the promised December code drops actually land on GitCode before planning a multi-year dependency. - **Read the source before you trust it.** Clone the CANN repository and skim the operator implementations for the one model family you care about, so you know what is genuinely open versus what still depends on closed Huawei components. - **Join the committee and contribute.** If you maintain a workload on Ascend, follow the CANN Technical Steering Committee and file a custom operator or a performance fix — early contributors shape the interfaces everyone else inherits. - **Benchmark the open path.** Port a single inference graph to the open MindIE path and compare it against your current Nvidia or AMD baseline on throughput per dollar, not just raw latency. ## Why it matters For a developer used to a rival accelerator stack, the practical worry was always that Ascend was a dead end: a chip you could buy but not fully program. Opening CANN answers that fear directly, because the compiler and virtual instruction set are now something you can read, patch and build against, so the workload you write is no longer hostage to Huawei's release calendar. For the ordinary reader, the significance is strategic rather than technical. A toolchain anyone can extend tends to attract the tutorials, drivers and community fixes that make a platform feel normal to use — the missing ingredient Chinese AI hardware has long struggled to grow. The bet is that freedom at the software layer quietly converts into dependence at the silicon layer: cheap to adopt, expensive to leave. ## Honest limitations The timeline and ecosystem numbers come from Huawei's own keynotes and annual report; we have not independently audited developer counts or solution tallies. "Fully open-source" is a Huawei commitment with a defined deadline, not a completed state at the time of writing — readers should confirm the December 2025 code drops shipped. Hardware pricing and availability were not disclosed and are excluded. We did not evaluate whether open CANN code is production-ready for third-party chips. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-huawei-ascend-cloud Free to quote with attribution and a link to the original.