Preface Executive Briefing

Week of April 4, 2026

Inside the Rise of Chinese AI

Has Chinese AI Caught Up?

Export controls vs. architectural efficiency vs. talent reversal. China may still trail in frontier chips, but it is moving fast through efficiency, open-weight ecosystems, and talent concentration.

See Key Takeaways
Explore the briefing
US
Leading
AI Tug of War
CN
Catching Up

At a Glance

Executive Summary

The global AI landscape has entered a phase of structural divergence. China is navigating U.S. export restrictions through architectural efficiency and talent concentration.

Hardware Constraints Remain Real

U.S. export restrictions have limited access to frontier chips. China relies on stockpiled older NVIDIA chips and emerging domestic alternatives like Huawei Ascend, though significant gaps in process node and interconnect remain.

7nm vs 3nm

Process node gap

Software Efficiency is China's Strongest Accelerant

Chinese labs have pioneered architectural breakthroughs: sparse attention, ultra-sparse MoE, knowledge distillation, and interleaved reasoning. Models are approximately 78% cheaper with competitive performance in reasoning and math.

78%

Cost reduction vs US

Talent and Ecosystem Dynamics Could Reshape the Race

The brain drain is reversing. Push factors from the U.S. (visa friction, scrutiny) and pull factors toward China (K-Visa, funded labs, national priority) are creating a talent flow shift with long-term implications.

7 of 12

Meta AI hires are Chinese

Based on analysis of U.S. export controls, Chinese AI lab innovations, and talent migration patterns through early 2026.

Round 1

Hardware and Manufacturing

The silicon foundation of AI capability. While China is catching up with domestic alternatives, the U.S. still dominates the most advanced chips.

U.S. Tech Export Restriction Timeline

Oct 2022restriction

First Restrictions

U.S. implements initial tech export restrictions on advanced logic chips, ICs, and semiconductor manufacturing equipment to China.

Oct 2023restriction

Expanded Controls

Additional controls on advanced AI chips and manufacturing equipment, closing loopholes in initial restrictions.

May 2025policy

AI Diffusion Rule

Countries categorized into trusted allies and countries of concern, limiting access to advanced AI hardware and models.

Jan 2025milestone

DeepSeek R1 Released

Chinese lab DeepSeek releases R1 model, demonstrating significant progress despite hardware constraints.

Dec 2025policy

H200 Sales Reopened

Trump administration greenlights export of NVIDIA H200 chips to China with 25% revenue-sharing requirement.

Chip Comparison: The Gap Remains Significant

Huawei Ascend 910C

China

Process NodeSMIC 7nm (DUV)
Compute~1,100 TFLOPS
Bandwidth3.2 TB/s
Interconnect392 GB/s

Interconnect bottleneck

Stability issues in large clusters

NVIDIA H20

US (Restricted)

Process NodeTSMC 5nm (EUV)
Compute296 TFLOPS
Bandwidth4.0 TB/s
Interconnect900 GB/s (NVLink)

Export controlled

25% revenue share required

NVIDIA Vera Rubin

US (Frontier)

Process NodeTSMC 3nm (EUV)
Compute50,000 TFLOPS
Bandwidth22 TB/s
Interconnect3,600 GB/s (NVLink v5)

Completely restricted

Blackwell/Rubin withheld

The Software Stack Challenge

CUDA (NVIDIA)

Industry-standard platform with mature ecosystem. The CUDA Moat creates significant switching costs for developers.

CANN (Huawei)

Compute Architecture for Neural Networks. Direct CUDA competitor but considered immature with limited ecosystem support.

MUSA (Moore Threads)

Translates CUDA code pragmatically but incurs 20-40% performance penalties. Compatibility workaround approach.

Round 1 Verdict: Hardware

U.S. maintains clear advantage at the frontier

Status:United States

Round 2

Model Innovation and Efficiency

The true strength of Chinese companies is rooted in their innovative software solutions and training methods. Efficiency, not brute-force scale, is becoming the competitive edge.

78%
Cost Reduction
vs US models
+29%
Math Index
Higher scores
+71.5%
AIME Scores
Math benchmark
+22.9%
LiveCodeBench
Coding capability
Knowledge Distillation
DeepSeek

Training smaller student models to mimic nuanced decision-making of larger teacher models.

->Much cheaper than traditional training
Sparse Attention
DeepSeek V3.2

Paying attention to only specific tokens vs. all. Uses a lightning indexer to prune the attention space.

->Reduces long-context FLOPs by over 50%
Manifold-Constrained Hyper-Connections
DeepSeek

Mathematical guardrails via Sinkhorn-Knopp algorithm force information mixing matrices to remain doubly stochastic.

->Enables reliable model scaling
Ultra-Sparse MoE
Alibaba Qwen3

80B total parameters, only 3B (3.75%) activated per inference step. Hybrid attention: 75% linear + 25% standard.

->Performance of 32B model at less than 10% cost

Leading Chinese Models

Kimi K2.5
Moonshot AI
Reasoning and Agents
10% cheaper than Claude Opus 4.5
DeepSeek V3.2
DeepSeek
Cost Efficiency
Aggressive cost reduction via distillation
Qwen3
Alibaba
Ultra-Sparse MoE
3.75% activation ratio
GLM-5
Z.ai
Agentic Capability
Highest agentic index in open weight
MiniMax M2.5
MiniMax
Cost Efficiency
High cost-efficiency audio generation

Round 2 Verdict: Models

China has caught up in meaningful areas

Status:Strategic Parity

Round 3

The Talent War

The overlooked but key driving force. The traditional brain drain of Chinese talent heading to Silicon Valley is now reversing.

Push Factors

Factors Pushing Talent Out

U.S. policy changes and geopolitical pressure are creating friction for Chinese-origin researchers working in American AI labs.

  • Visa and immigration hurdles: stricter H-1B policies and increased processing delays
  • Geopolitical scrutiny: increased pressure on Chinese scientists
  • Research restrictions: limitations on collaboration and publication
  • DoD contractor restrictions and IP theft accusations

Pull Factors

Factors Pulling Talent In

China's national AI priority and well-funded labs are creating compelling alternatives for researchers considering their options.

  • K-Visa program: for global STEM talents to start businesses or work without upfront employer sponsor
  • National priority: science and technology treated as top national priority
  • Well-funded labs: generous government-backed capital and research funding
  • Prestige and speed: high-intensity work culture enabling rapid iteration

"Our Chinese" vs. "Their Chinese"? The distinction is becoming increasingly blurred as talent flows reshape the global AI landscape.

The talent race is not settled, but China is becoming increasingly attractive to top researchers.

Round 3 Verdict: Talent

Open question with high strategic importance

Status:Contested / Dynamic

Synthesis

So, Has Chinese AI Caught Up?

The answer is nuanced. The race is no longer about raw power, but about architectural efficiency, ecosystem strategy, and talent dynamics.

Hardware
Not Yet

The U.S. maintains a clear lead in frontier chip design and manufacturing. China faces significant gaps in process node, interconnect speed, and ecosystem maturity.

-NVIDIA Rubin: 50,000 TFLOPS vs Huawei 910C: approx. 1,100 TFLOPS
-TSMC 3nm EUV vs SMIC 7nm DUV
-Software stack moat remains significant
Model Capability
Parity / Near-Yes

In several domains, Chinese models have reached parity or near-parity. Particularly strong in reasoning, cost-efficiency, and architectural innovation.

-78% cost reduction vs US models
-+29% higher Math Index scores
-Leading in image/video generation niches
Commercialization
Increasingly Competitive

China's open-weight strategy and aggressive pricing are driving global adoption. IPO activity signals market confidence.

-Open-weight models driving ecosystem spread
-Zhipu and MiniMax HK IPOs (Jan 2026)
-Strong in practical deployment scenarios
Talent and Momentum
Highly Consequential

The talent race is not settled but moving. Push factors from U.S. and pull factors toward China are creating a structural shift.

-7 of 12 Meta AI hires are Chinese
-K-Visa vs H-1B friction
-National priority status in China

Overall Assessment

Chinese AI has not uniformly "caught up"; the reality is more nuanced. While the U.S. maintains leadership in frontier hardware, China has achieved strategic parity in model innovation through architectural efficiency. The talent dynamics are shifting, and the open-weight ecosystem strategy is driving global adoption.

Hardware: US LeadsModels: ParityTalent: Contested

Key Takeaways

Key Takeaways

What the rise of Chinese AI means for business leaders, strategists, and technology decision-makers.

01

Cost-Efficiency Over Brute-Force Performance

Chinese model APIs can run at approximately 20% of the cost of comparable US models, enabled by distillation and sparse attention that reduce compute per token. Knowledge distillation reduces training costs; sparse attention cuts inference FLOPs by 50%; ultra-sparse MoE delivers 32B performance at 10% cost.

02

Winning Market Niches

In image and video generation, Chinese providers increasingly lead quality rankings. Kling's models rank #1 and #2 on major public leaderboards. ByteDance Seedance 2.0 launched in February 2026. Tencent Hunyuan is advancing 3D generation capabilities.

03

Open-Weight Ecosystem Play

China's open-weight strategy is an adoption play: make models easy to download and build on, so the ecosystem spreads globally. Qwen, DeepSeek, and GLM models are freely available, driving global adoption in emerging markets and creating switching costs through ecosystem lock-in.

FAQ

The rescission reverts the regulatory landscape to the Oct 2022 and Oct 2023 controls, focusing on hardware performance thresholds rather than worldwide licensing of model weights.