DeepSeek Pairs With Huawei to Challenge Nvidia's Grip on AI Compute

DeepSeek and Huawei didn't port a model to a new chip this week. They open-sourced a full set of performance-critical libraries designed for Huawei's Ascend accelerators: DeepGEMM-Ascend for matrix math, DeepEP-Ascend for expert routing on mixture-of-experts models, TileKernels, FlashMLA for multi-head latent attention, and DeepSelect for token selection. The five repos landed the same week as HUAWEI CONNECT, the company's annual compute conference, and they ship together rather than as ports of Nvidia-targeted libraries renamed for Ascend.

That is the actual story. CUDA, Nvidia's proprietary programming layer, is more than a compiler; it is the reason most AI software "just works" on Nvidia hardware and nowhere else. A port gets a model running on a new chip. Co-design treats the model, the runtime, and the silicon as one system, so the resulting stack can match the Nvidia baseline at the kernel level instead of after the fact. DeepSeek is now publishing evidence it is doing the second thing.

The release turns DeepSeek into the hinge between China's model labs and its domestic chip industry. Until now, Chinese AI developers had two rough options: run on Nvidia hardware, where export controls keep tightening, or port existing CUDA code to Ascend and accept the performance tax. DeepSeek's Ascend-native libraries, per a geopolitechs write-up, are the first time a top-tier Chinese lab has shipped a complete open-source toolchain for the alternative path. Other labs can build on it instead of rewriting from scratch.

This is the part of the story that Recode China AI's weekly digest captured as "DeepSeek and Huawei target Nvidia's CUDA moat." Moat is finance jargon. The concrete version: a complete stack, model down to metal, that does not depend on Nvidia.

Manus relaunched this week as Manus 2.0 with an always-on personal agent called Cue, a reported $4 billion valuation (roughly double the $2 billion acquisition offer that collapsed earlier), and the product claim that agents get their own phone numbers and wallets. The phone number is the substrate for voice and SMS. The wallet is the substrate for payments. The pairing is what turns an agent from a chatbot into something a person can leave running.

That pairing is the agent layer's version of the Ascend-native toolchain. ByteDance's Doubao, StepFun, Zhipu's Flash line, and a string of others are racing toward the same shape: an always-on personal agent wired to the consumer substrate. Manus 2.0 is the first Chinese product to ship all three pieces (model, identity, money rail) as a public relaunch rather than a research preview.

There is a falsifier that keeps the DeepSeek-Huawei story honest. The five repos are real and the code is public. Production-equivalent frontier training at scale on Ascend, the kind that would let a Chinese lab train a new top-tier model end-to-end without Nvidia, is not yet independently verified. The right read on this week's release: a toolchain claim, not a frontier-training claim. If the next model trained entirely on Ascend matches the Nvidia baseline on benchmark cost and speed, the moat cracks. Until then, the moat is under design, not broken.

At the Xi-Trump summit, Washington and Beijing announced a direct AI-incident communication channel as one of eight points of agreement. Within days, Trump rejected calls to cooperate with China on AI safety and pushed to rename "AI" as "superintelligence." The hotline exists on paper. Operating status is unconfirmed.

Kuaishou's Kling AI unveiled a new flagship video-generation model, tightening its chase of ByteDance's Seedance. Ant Group previewed Ling-3.1-flash, a compact model aimed at on-device agents.

The next test is whether a frontier model ships with Ascend-native training in its provenance. DeepSeek published a paper this week disclosing DSec, the training infrastructure behind V4.1, with Liang Wenfeng on the byline. If V4.1 was trained end-to-end on Ascend, the toolchain claim graduates to a frontier claim. If the paper's training footprint still leans on Nvidia, the toolchain is this week's lead, and the frontier claim waits.