🧔♂️ A friendly human may check it before it goes live. More news here
China’s AI unlikely to catch up with US in coming years: researcher
Top Chinese AI scientists say China is unlikely to overtake the US in AI over the next three to five years due to limited access to advanced chips and computing resources.
Speaking at an industry conference in Beijing, Alibaba’s Qwen team technical lead Lin Junyang estimated there is less than a 20% chance for any Chinese firm to surpass US companies like Google DeepMind and OpenAI in the near term.
Lin said US firms benefit from far greater computational power, while Chinese companies are operating at capacity just to meet current demand.
Tang Jie, chief AI scientist at Zhipu AI, echoed concerns that the US maintains a lead, noting some American AI models remain unreleased to the public.
Tencent’s new chief AI scientist, Yao Shunyu, predicted a Chinese firm could lead the field in three to five years if challenges in chip manufacturing and foundational research are overcome.
🔗 Source: South China Morning Post
🧠 Food for thought
Implications, context, and why it matters.
China’s compute gap under tighter US rules
- The US holds about 70% of AI compute (the capacity to train and run AI models), while China has 15%, a fivefold gap that could reach tenfold for GPU clusters 1.
- China hit 788 EFLOPS at FP16 in June 2025, with IDC projecting 1,037 EFLOPS by year‑end 1.
- The US AI Diffusion Framework (Jan 2025) blocks training of frontier AI models (systems needing ≥1e26 FLOPs) outside Tier 1 countries (a small set of approved jurisdictions) 2.
- Since 2023, Chinese models have trailed US frontier systems by seven months, with a four to fourteen month range 3. Leading Chinese models are open‑weight (parameters are publicly released), while US systems stay closed‑weight (proprietary) 3.
- Domestic GPU self‑sufficiency rose to 34% in 2024 from below 10% in 2020, and could reach 82% by 2027 4. Firms are at capacity, which limits closing the frontier gap.
Tier 2 providers target spillover AI work
- The framework lets Tier 2 countries (jurisdictions permitted limited imports) import GPUs, bans frontier training, and allows inference (running trained models) plus smaller training 2.
- Alibaba plus ByteDance are pursuing Nvidia’s China‑compliant H20 chips (AI accelerators designed to meet export rules) and tracking B30A, signaling demand for approved accelerators 5.
- Tier 2 hubs like Singapore can offer H20 for inference or sub‑frontier training, with the Validated End User structure favoring approved overseas providers 2.
- Locations with strong links to China, like Singapore, could see more demand as firms seek compliant access to compute 25.
Stay updated on the go with our mobile app.
Get latest insights with smoother, more personalized experience through TIA mobile app.