共有:

Silicon Sovereignty

DeepSeek Teams Up With Huawei —
Where Chinese AI Gets Its Silicon

China's leading models have caught up on quality and price. The next round of the contest is being fought one layer down. DeepSeek's move to build an inference stack that skips NVIDIA — and runs on Huawei silicon instead — adds a new column to every enterprise buyer's model-selection spreadsheet.

AI Navigate Editorial·2026.07.14·6 min read

BEFORE DeepSeek model CUDA / cuDNN NVIDIA H100 / H200 2026 shift AFTER DeepSeek model CANN / MindSpore Huawei Ascend 910C Fully domestic stack in China
FIG. The model may stay the same, but the four layers below get swapped end-to-end
01

The News

After catching up on quality,
DeepSeek moves under the hood

On July 14, 2026, Nikkei xTECH first reported DeepSeek's partnership with Huawei to shift its entire inference stack onto Ascend silicon.

The core of the story is simple. DeepSeek is signaling that its next generation of training and inference will run on Huawei Ascend 910C chips paired with CANN / MindSpore, rather than NVIDIA GPUs. Japan's Nikkei xTECH broke the story first, with Reuters, Bloomberg, and SCMP following soon after.

The backdrop is the third phase of the US "AI Diffusion Rule" enforced in October 2025 by the Bureau of Industry and Security. Exports of H100 and H200 to China are effectively frozen, and even the China-specific H20 has grown scarce. For Chinese AI labs, dependence on the NVIDIA stack has become the single biggest supply risk — and DeepSeek is the first frontier player racing out of it.

02

By the Numbers

It's not one chip — it's
four layers replaced at once

Ascend 910C
Huawei's flagship inference chip
CANN 8.0
CUDA-equivalent wrapper
MindSpore
PyTorch-equivalent framework
Atlas 900
Cluster-building enclosure

NVIDIA's stack is dominant not because of the GPU alone, but because four layers — CUDA, cuDNN, PyTorch, and the deployment stack — work seamlessly together. DeepSeek and Huawei are replacing all four with Ascend / CANN / MindSpore / Atlas. They know that swapping just the chip won't work — and that's why the swap covers the whole tower.


03

Why It Matters

Why this matters now

It looks like a China-only story, but every enterprise buyer just gained a new column on their vendor spreadsheet.

01

From model-vs-model to stack-vs-stack

DeepSeek vs Claude used to be a quality-and-cost comparison. Going forward, buyers also need to ask "which country's chips power this inference?" — a new axis for audit, procurement, and export-control reviews.

02

Continuous with Lindy's Claude-to-DeepSeek migration

On June 27, US startup Lindy announced a company-wide switch from Claude to DeepSeek. Companies chasing cost efficiency are already migrating — and now that cost efficiency will be delivered by Huawei Ascend.

03

NVIDIA's China revenue shrinks further

NVIDIA's data-center GPU revenue from China is already about 60% below its 2022 peak. If a marquee customer like DeepSeek moves fully to Ascend, expect that number to keep bending down.

04

Who's Affected

Who feels it, and how

Engineers

If you only touch the API, no immediate change. But teams planning self-hosted DeepSeek should track the maturity of Ascend-native inference engines (e.g., vLLM's CANN backend) before committing.

Executives & Procurement

Country-of-origin for the silicon underneath your AI is now a real audit item. Finance, defense, and public-sector work will increasingly demand a line-of-sight not just to the model, but to which chips run the inference.

PMs & Deployment Leads

"DeepSeek because it's cheaper" no longer sails through legal review by default. Confirm you have a vendor path with "served from Japan/EU/US region" written into the contract.

05

The Counterpoint

The optimist case has three cracks

"NVIDIA-free" isn't decided yet. Three unknowns remain.

①The stack maturity gap. CANN is roughly eight years younger than CUDA, and real-world evidence at frontier-model scale is still thin. ②Supply. Ascend 910C is manufactured at SMIC on a 7-nm-equivalent process; production capacity remains a fraction of NVIDIA H200's volume. ③Ecosystem gravity. Non-Chinese developers have little motivation to port to Ascend, so day-zero OSS support may not follow. Reaching parity on both quality and availability will take another few years.


06

What to Do Next

Recommended actions

This weekThis quarter
Inventory every current DeepSeek route (API / self-hosted)Add a "silicon layer" column to vendor selection sheets
Add Ascend-ready vLLM / SGLang to your watch listRedesign DeepSeek vs Claude / GPT usage by SLO tier
Ask Japanese-region vendors (Sakura, IIJ, etc.) for alternate pathsAdd "model residency country" clauses to procurement contracts

DeepSeek and Huawei's move makes "should we use a Chinese model?" harder to answer with a single yes/no. Splitting the decision by SLO tier across four dimensions — quality, price, export controls, availability — will be the operational standard over the next 12 months. Keep an eye on the DeepSeek company site for further releases expected this month.