Cambricon Q1 Revenue Surges 150%: US Export Controls Are Clearing the Field for Chinese Chip Makers

Core Judgment

Two seemingly independent events are telling the same story:

DeepSeek V4 released on April 24, announcing full training and deployment based on Huawei Ascend chips — not relying on any Nvidia GPU
Cambricon Q1 revenue year-over-year growth exceeding 150%, Chinese CSPs (cloud service providers) have moved from testing phase to full-scale deployment

The effect of US export controls is backfiring: rather than stopping China’s AI development, it has created a protected monopoly market for Chinese chip makers.

What Happened

Event One: DeepSeek V4 × Ascend 950

The DeepSeek V4 series (1.6T parameter V4-Pro and 284B parameter V4-Flash) is based on Huawei Ascend 950 chips from the training phase. This is the first frontier large model designed from the source to adapt to domestic chips.

Key data:

Ascend 950’s FP4 compute is 2.87x that of Nvidia H20
First token latency as low as 20ms
Huawei announced complete adaptation of the full Ascend supernode series within hours

Event Two: Cambricon Performance Explosion

According to the latest financial data, Cambricon Q1 revenue grew over 150% year-over-year. The growth driver is not government subsidies, but commercial orders — Chinese cloud service providers are moving from testing to large-scale deployment.

Event Three: vLLM 0.20.0 MegaMoE Optimization

On April 29, the vLLM project released version 0.20.0, introducing MegaMoE optimization. Combined with DeepSeek V4 Pro’s MoE architecture, significant performance improvements were achieved on GB200. This means DeepSeek V4 can run efficiently on both Nvidia and Huawei platforms.

Data: Competitiveness of Domestic Chips is Being Quantified

Chip	FP4 Compute Comparison	Ecosystem Maturity	Representative Customers
Huawei Ascend 950	2.87x H20	High (CANN + MindSpore)	DeepSeek, iFlytek
Cambricon MLU	~1.5x H20	Medium (Neuware)	Multiple CSPs
Nvidia H20 (China special)	Baseline	Highest (CUDA)	Restricted
Nvidia H100/B200	Far exceeds	Highest	Banned

Note a key detail: Nvidia’s H100/B200 is banned for China, H20 is a significantly degraded special version. This means Chinese companies cannot buy the best Nvidia chips even if they want to — this is not “choosing domestic,” but “forced domestic.”

Why This is a Turning Point

For the past two years, the质疑 facing Chinese AI chips was: “usable, but not good.” DeepSeek V4’s performance on Ascend is breaking this narrative:

Three marks of moving from “usable” to “good”:

Native training: Not training first then migrating, but running on Ascend from day one of training
Performance exceeding: FP4 compute exceeds Nvidia H20, and H20 itself is a degraded version — meaning within the range of available chips, domestic chips are no longer “backup options”
Ecosystem completion: Rapid adaptation of open-source inference frameworks like vLLM and MegaMoE lowers deployment barriers

The “Backfire Effect” of US Export Controls

This is essentially a classic import substitution story, but accelerated 10x:

Normal path: Domestic chips catch up → 10-15 years → gradual substitution
Current path: Imports blocked → Market forced open → Domestic chips get real customer feedback → 2-3 years rapid iteration

Cambricon’s 150% revenue growth is not from government orders, but from commercial customers’ active procurement. This means domestic chips have crossed the “usable” threshold and entered the “good and cost-effective” stage.

Industry Impact

For Model Companies

Controllable compute costs: No longer subject to Nvidia’s pricing and supply
Technical sovereignty: Training infrastructure no longer depends on external supply chains
Reduced compliance risk: Using domestic chips involves no US export control compliance issues

For Chip Companies

Deterministic market demand: Chinese AI companies’ compute demand is real and sustained
Accelerated feedback loop: Large-scale deployment brings real-scenario feedback, driving chip iteration
Increased capital confidence: Cambricon’s performance provides a valuation anchor for the entire track

For Global Landscape

Accelerated dual-tracking: Global AI infrastructure is splitting into two tracks: “Nvidia ecosystem” and “domestic ecosystem”
Third-party market competition: Southeast Asia, Middle East, Latin America will become competitive battlegrounds for the two ecosystems

Risk Warnings

Performance gap remains: Ascend 950 targets the degraded H20, not H100/B200. There is still a gap in absolute compute
Ecosystem barriers: CUDA’s global developer ecosystem cannot be replicated in the short term
Supply chain risk: Domestic chip manufacturing itself still depends on external equipment (lithography machines, etc.)

Action Recommendations

AI companies operating in China: Re-evaluate compute strategy, Ascend+Cambricon is already a trustworthy alternative
Chip investors: Cambricon’s revenue growth is an industry signal — the entire domestic AI chip track is accelerating
Overseas developers: Pay attention to vLLM + MegaMoE inference optimization for MoE models, this is key for cross-platform deployment