Export Controls Backfired: China Achieves Frontier AI on Huawei Silicon While US Open-Source Stalls

Key Takeaways

• GLM-5 trained entirely on Huawei Ascend chips without NVIDIA dependency achieved 77.8% SWE-bench—outperforming GPT-5.2 (75.4%) and Gemini 3 Pro (74.2%)
• Zhipu AI (GLM developer) raised $558M in Hong Kong IPO while on US Entity List, demonstrating sanctioned entities can access non-US capital
• Chinese open-source models (Qwen 3.5 Apache 2.0, DeepSeek V4, GLM-5) now dominate the permissive licensing tier while US open-source comes from a 30-person startup (Arcee)
• Export controls intended to maintain US competitive advantage inadvertently created conditions where Chinese-origin models are now more accessible and permissively licensed than US alternatives
• Hardware pluralism is the emerging infrastructure reality: Huawei Ascend for training, Apple Neural Engine + Qualcomm NPU for inference, AMD investing in world models

The Huawei Ascend Proof Point

Zhipu AI's GLM-5, developed at Tsinghua University by a lab on the US Entity List since January 2025, trained entirely on Huawei Ascend chips with zero NVIDIA dependency. The performance: 77.8% on SWE-bench Verified—beating GPT-5.2 (75.4%) and Gemini 3 Pro (74.2%), trailing only Claude Opus 4.5 (80.9%) and Claude Sonnet 4.6 (79.6%).

GLM-5 is not a proof-of-concept. Zhipu raised $558M in a Hong Kong IPO in January 2026 while on the US Entity List. This is production-scale commercial deployment from a sanctioned entity that has solved its NVIDIA dependency problem.

The significance is structural: Huawei's Ascend 910B and 910C chips are not yet equivalent to NVIDIA's Blackwell generation in raw FLOPS, but GLM-5 demonstrates that a sufficiently optimized training pipeline on domestic silicon can achieve frontier-tier results. Once the bottleneck is software and methodology rather than hardware FLOPS, export controls lose their leverage.

The Open-Source Licensing Asymmetry

Arcee AI's Trinity (400B, Apache 2.0, $20M training) was explicitly positioned as a US response to Chinese AI infrastructure dependency. CEO Mark McQuade's stated mission: "Arcee exists because the US needs a permanently open, Apache-licensed, frontier-grade alternative."

This reveals the competitive dynamic: the US open-source response to Chinese AI comes from a 30-person startup, not from the dominant labs. Meta's Llama 4 Maverick carries commercial use restrictions that create legal ambiguity for enterprise deployment. Alibaba's Qwen 3.5 ships under Apache 2.0 with 201 language support and 60% cost reduction versus its predecessor. DeepSeek V4 extends context to 1M+ tokens via its Engram memory architecture.

Three Chinese open-source models in February 2026 alone offer cleaner licensing, broader language coverage, and more aggressive cost reduction than the incumbent US open-source offering. The irony is precise: US export controls intended to preserve competitive advantage have created conditions where the most permissively licensed, globally accessible frontier models are increasingly Chinese-origin, while US frontier labs compete primarily in the closed-source, API-gated tier.

The February 2026 Model Distribution

Of seven frontier models launched:

• Four US: Claude Sonnet 4.6, GPT-5.3, Grok 4.20, Gemini 3 Pro — all closed-source, API-gated
• Three Chinese: Qwen 3.5 (Apache 2.0), GLM-5 (open), DeepSeek V4 (open) — all open-source or permissively licensed

The US labs dominate closed-source; Chinese labs dominate open-source. Arcee Trinity is the only US Apache 2.0 frontier model, from a 30-person startup without the ecosystem depth of Alibaba or Tsinghua.

For sovereignty-sensitive enterprise buyers (EU companies concerned about GDPR, government agencies, healthcare systems), Chinese open-source under Apache 2.0 is increasingly attractive: cleaner license, no US cloud dependency, locally deployable. Qwen 3.5's 201-language support makes it particularly attractive for Southeast Asia, Africa, and Latin America markets where US closed-source models have limited language coverage.

Hardware Pluralism Is Now the Structural Reality

GLM-5 on Huawei Ascend is the most consequential data point for export control strategy. Once a major lab demonstrates frontier capability on domestic silicon, the hardware moat closes permanently. Huawei can now sell Ascend chips with proven GLM-5 results as reference deployment. Other compute-constrained actors—Iranian researchers, Russian labs, sanctioned entities—have a viable hardware path. The 6-18 month compute gap assumption that justified export controls no longer holds for models in the 400-700B parameter range.

The infrastructure shift toward hardware pluralism is accelerating on multiple fronts:

• Huawei Ascend for training (demonstrated frontier capability)
• Apple Neural Engine + Qualcomm NPU for inference (ExecuTorch 12+ backends)
• AMD investing in Runway for world model compute

What This Means for ML Engineers

For teams evaluating open-source models:

1. Qwen 3.5 Is Now a Production-Viable Llama 4 Alternative – Apache 2.0 licensing, 201-language support, 60% cost reduction, globally deployable
2. Evaluate Arcee Trinity Alongside Llama 4 – US-origin Apache 2.0 frontier capability exists; force the licensing and geopolitical risk question to surface explicitly
3. For Sovereignty-Sensitive Deployments, Chinese Open-Source Creates Legal Complexity – Qwen 3.5 licensing is cleaner than Llama 4's, but geopolitical risk of Chinese-origin on-premise deployment is a first-time legal question that corporate legal teams are not yet equipped to assess
4. Assume Hardware Pluralism in Your Infrastructure Planning – Do not design for NVIDIA-only training or inference pipelines. Build deployment flexibility that can work across Apple Neural Engine, Qualcomm NPU, MediaTek APU, and AMD architectures