Illustration for: DeepSeek Bets 160,000 Huawei Chips On Beating Export Curbs

DeepSeek Bets 160,000 Huawei Chips On Beating Export Curbs

DeepSeek CEO Liang Wenfeng expects Huawei to begin delivering AI training chips as early as Q4 2026, with DeepSeek planning to deploy at least 160,000 Huawei Ascend accelerators at a new Inner Mongolia data center.

By the Numbers

160,000+
Ascend chips planned
Q4 2026
Delivery target
~750,000 units
Huawei 2026 Ascend output
~510,000 (68%)
DeepSeek + ByteDance share
TC
By the AI Desk
Edited by Trace Cohen · Early-stage VC & angel · Founder, New York Venture Partners
2 min read
ShareXLinkedInEmail

THE RUNDOWN

1

160,000 Huawei Ascend accelerators at a single Inner Mongolia data center is a serious commitment to domestic silicon at exactly the scale needed for frontier-model training, not a hedge or backup plan alongside continued Nvidia purchases.

2

Huawei detailed its next-generation Ascend 960 family at its September 17 Huawei Connect conference, with the Ascend 960DT launching in Q1 2027 -- meaning DeepSeek's Q4 2026 chip deliveries likely rely on the current Ascend 950 generation, not the newer chip.

3

Huawei is targeting 750,000 total Ascend units shipped in 2026, with ByteDance and DeepSeek alone accounting for roughly 510,000 of that -- meaning two customers represent about two-thirds of Huawei's entire annual AI-chip output.

4

The same US export controls that pushed DeepSeek toward Huawei are simultaneously constraining Huawei's own ability to meet the demand those controls helped create, since manufacturing-equipment restrictions limit how much advanced chip capacity China can build domestically.

TC

The VC Read · Trace's Take

Trace Cohen

DeepSeek betting on the older Ascend 950 rather than waiting for the 960DT in Q1 2027 tells you supply certainty now beats performance parity later for Chinese labs under export controls -- a trade-off US labs simply don't have to make. The diligence item for anyone tracking the AI chip wars: Huawei's own production capacity, not DeepSeek's stated 160,000-chip target, since two customers already eating two-thirds of Huawei's annual output leaves almost no room for a manufacturing miss.

Analysis

DeepSeek CEO Liang Wenfeng told investors he expects Huawei to begin delivering AI training chips to DeepSeek as early as the fourth quarter of 2026, with DeepSeek planning to deploy at least 160,000 Huawei Ascend accelerators at a new data center it is building in Inner Mongolia, according to The Information. Years of tightening US export controls on Nvidia's most advanced data-center GPUs have progressively pushed Chinese AI labs toward domestic alternatives, and Huawei's Ascend line is the most credible option at scale. Pulse previously covered DeepSeek's parallel effort to build its own inference chip rather than depend entirely on Nvidia or Huawei -- this Ascend commitment shows that hedge hasn't stopped the company from also leaning harder on Huawei's silicon in the meantime.

The Chips Available Aren't The Newest Ones

Huawei detailed its next-generation Ascend 960 family at its Huawei Connect conference on September 17, with the flagship Ascend 960DT not launching until the first quarter of 2027. That timing means DeepSeek's fourth-quarter 2026 deliveries will most likely draw on the current Ascend 950 generation rather than the newer chip -- a real capability gap DeepSeek is accepting in exchange for chips it can actually get in volume, rather than waiting for Nvidia hardware it may never be allowed to buy at the scale it needs.

Two Customers, Two-Thirds Of Supply

Huawei is targeting roughly 750,000 total Ascend units shipped across 2026, with ByteDance and DeepSeek alone accounting for approximately 510,000 of that total -- meaning two customers represent close to two-thirds of Huawei's entire annual AI-chip output this year. That concentration cuts both ways: it gives DeepSeek and ByteDance priority access to China's most advanced domestic AI silicon, but it also means any Huawei production shortfall lands disproportionately on the two companies most dependent on it.

The Controls Constrain Both Sides

The irony underlying this entire shift is that the same US restrictions on advanced chipmaking equipment that pushed Chinese labs toward Huawei are also limiting Huawei's own ability to manufacture the Ascend 950 at the scale the resulting demand requires. DeepSeek's bet on 160,000 units assumes Huawei can actually deliver at that volume on the promised Q4 timeline -- a supply-chain assumption that has already slipped once this year across the broader Chinese chip sector.

What This Signals For The AI Chip Race

DeepSeek's willingness to commit at this scale to a chip generation it knows is one step behind Huawei's own roadmap suggests Chinese AI labs increasingly view supply certainty as more valuable than raw performance parity with Nvidia -- a strategic trade-off distinct from anything US labs currently have to make. Whether that trade pays off depends entirely on Huawei's execution against its own Q4 delivery promise, which is the number to watch, not DeepSeek's stated deployment target.

ShareXLinkedInEmail

Key Sources

2 sources

THE WIRE in your inbox— Tech, startup & VC news with Trace's take. Free, no spam.