Sources: ByteDance plans to develop an AI model trained primarily with Huawei Ascend 910B chips, which it already uses for less intensive inference tasks
TikTok's Chinese parent ByteDance plans to develop an AI model trained primarily with chips from compatriot Huawei Technologies (HWT.UL) …
Context & Ripple Effects
This report marks an early move by ByteDance to use Huawei hardware beyond lower-intensity inference and into model training. Later coverage points to a broader sourcing shift: ByteDance was reported to be directing much of its 2025 China AI-chip budget to domestic suppliers, including Huawei, in a larger domestic-chip procurement push.
The hardware strategy has since widened rather than become exclusive: ByteDance has reportedly pursued its own CPUs and an InnoStar-designed inference chip, while also considering GPUs from Iluvatar CoreX and Baidu's Kunlunxin. That makes Ascend training part of a diversified effort to build AI capacity around multiple Chinese compute options.
First-order effects
- ByteDance would validate Ascend 910B for a more demanding workload than the inference tasks it was already handling, making Huawei a prospective training-compute supplier as well as an inference supplier.
- Huawei gains a high-profile potential workload for its AI accelerator platform, while ByteDance takes on the engineering work of adapting training software and models to that hardware.
Second-order effects
- A successful deployment would give ByteDance more leverage to distribute workloads across chip platforms instead of relying on a single supplier or architecture; its later in-house CPU effort reinforces that direction.
- Other Chinese AI-chip vendors face a higher bar: they must compete not only on chip availability but on training performance, software compatibility, and the ability to support large-scale deployments.
Third-order effects
- If large model developers keep combining domestic accelerators, bespoke chips, and alternative GPU suppliers, China's AI-compute market could become more heterogeneous, with software portability and systems integration becoming strategic constraints.
- The pattern could also deepen the separation between AI stacks built around Chinese hardware and those centered on other accelerator ecosystems, though the corpus does not establish that ByteDance will fully replace any one platform.
The trend: This is one data point in the shift from single-vendor AI compute toward heterogeneous, domestically anchored hardware stacks for training and inference.