Observed Signal · Sep 21, 2026 · Policy Update · Source: Hello China Tech · Impact: 4/5 · Sentiment: Neutral
Infrastructure Market: China's AI Gap Splits Three Ways: Code, Video, Copying
The article analyzes the AI capability gap between China and the US across three dimensions: coding, video generation, and model distillation. On Terminal-Bench 4.0, the best Chinese model (GLM-5.3) scores 41.8% versus OpenAI's GPT-6 Astra at 58.2%, but on DeepSWE, Kimi K3 matches Codex at 68%. In video generation, Chinese models dominate top rankings, while in image editing they lag. The article also highlights Nvidia's use of DeepSeek's R1 for training and Anthropic's allegations against Chinese labs for extracting Claude capabilities. It questions the role of compute, platform ownership, and licensing in shaping these gaps.
The article discusses major AI model capabilities and competitive dynamics between US and Chinese labs, impacting the AI/AdTech industry. It highlights performance benchmarks and training data sourcing, which are crucial for understanding AI infrastructure. But it is a roundup of multiple benchmarks and reports, not a single corporate event.
Key Takeaways & Evidence Grounding
- As of September 19, 2026, the best Chinese model on Terminal-Bench 4.0 (GLM-5.3) scores 41.8%, while OpenAI's GPT-6 Astra scores 58.2%.
- On the DeepSWE benchmark, Kimi K3 and OpenAI's Codex both score 68%.
- Chinese models hold five of the top ten spots on LMArena's image-to-video leaderboard.
- Nvidia generated 5 million training samples using DeepSeek's R1 model to train its Nemotron models, as published in July 2025.
- Anthropic reported that seven China-based AI labs, including Alibaba and Moonshot, conducted large-scale extraction of Claude capabilities, attributing over 151 million exchanges to Alibaba over three months.
Connected Companies & Entities
8 Entities mappedAnthropic
Foundation model company selling AI assistants and model APIs.
“Anthropic published a report alleging that seven China-based AI labs ... had conducted large-scale campaigns to extract Claude's capabilitie...”
NVIDIA
Accelerated computing company spanning AI software, cloud and gaming.
“In July 2025, Nvidia published a blog post describing how it generated 5 million training samples using DeepSeek's R1 model to train its own...”
DeepSeek
LLM developer offering AI chat and API access.
“Nvidia generated 5 million training samples using DeepSeek's R1 model to train its own Nemotron models....”
Alibaba Group
Chinese commerce and digital platform group spanning retail, ads and SaaS.
“Anthropic attributed over 151 million exchanges to Alibaba alone over a three-month period....”
ByteDance
Consumer platforms, advertising, commerce and collaboration software group.
“Seedream 5.0 Pro, ByteDance's best entry, ranks ninth on image editing....”
Z.ai (also marketed as Zhipu AI / 智谱AI)
Chinese LLM developer selling APIs, AI agents and MaaS.
“The best Chinese entry is Zhipu's GLM-5.3 at 41.8%....”
OpenAI
Foundation model company selling AI software, APIs and subscriptions.
“The top score on Terminal-Bench 4.0 belongs to OpenAI's GPT-6 Astra at 58.2%....”
Moonshot AI
Chinese AI company delivering Kimi models, agents, enterprise software and APIs.
“Anthropic alleged that seven China-based AI labs, including Alibaba and Moonshot, had conducted large-scale campaigns....”
Ontology Mapping & Concepts
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.