Weekly AI Paper Briefing — 2026-06-28

I’ve put together the most important AI research papers and industry trends from the past week, based solely on recently released information.

<!-- /headline -->

China's narrowing AI gap, the rise of local models, and the chip bottleneck.

<!-- /headline -->

---

1. Zhipu's GLM 5.2 Model — China's rising AI competitiveness

• Key takeaway: The GLM 5.2 model released by China's Zhipu is being seen as a sign of China's technical progress in its competition with U.S. AI labs. The model is delivering performance that challenges existing assumptions about the state of Chinese AI technology.
• Key contribution: It highlights the intensifying AI development race between China and the U.S. and signals a reshaping of the global AI industry landscape.

---

2. Local model performance improvements — The rise of Qwen 27B and Gemma 31B

• Key takeaway: A new trend is emerging in the performance comparison between dense models and Mixture-of-Experts (MoE) models. Dense models like Qwen 27B and Gemma 31B are proving to be slower but highly accurate, while MoE models like Gemma 26B and Qwen 35B are faster but show higher error rates.
• Key contribution: It clarifies the performance-versus-speed trade-off when deploying local AI models, providing clear criteria for practical implementation.

---

3. Semiconductor packaging technology as an AI bottleneck

• Key takeaway: Advanced chip packaging technology has emerged as a critical factor determining AI computing power. With Taiwan's TSMC handling the majority of this technology, U.S. reliance on Taiwan for the AI industry is at an all-time high.
• Key contribution: It highlights geopolitical risks in AI infrastructure and the importance of diversifying supply chains.

---

4. Hyperscience and Reveal win 2026 AI Breakthrough Awards

• Key takeaway: Hyperscience won "IDP Platform of the Year" for the second consecutive year, and Infragistics' Reveal platform won "Data Visualization Solution of the Year." These companies are focused on providing conversational AI analytics and automation solutions for enterprise environments.
• Key contribution: It boosts industry confidence through the successful real-world application and validation of enterprise-grade AI solutions.

---

5. Expanding AI infrastructure investment and energy bottlenecks

• Key takeaway: Goldman Sachs predicts that AI-related spending will reach $800 billion by the end of 2026, up from an annualized $650 billion the previous year. At the same time, the market continues to face structural issues where 90% of AI inference costs are covered by subsidies.
• Key contribution: It highlights both the massive capital influx into the AI industry and the challenges regarding its economic sustainability.

---

Research trend analysis for the week

1. Intensifying geopolitical competition: The emergence of China's Zhipu GLM 5.2 model is leading to a multipolar AI development landscape, moving away from a U.S.-centric one. This suggests a decentralization of global AI technology, with intensified competition to build regional AI ecosystems.

2. Maturation of edge computing: Improved performance in local models (e.g., Qwen 27B, Gemma 31B) makes the shift from centralized cloud-based AI to distributed edge AI more tangible. This opens up possibilities for achieving lower latency, better privacy, and improved cost efficiency simultaneously.

3. Structural supply chain bottlenecks: The concentration of semiconductor packaging technology in Taiwan and the limits of energy infrastructure are acting as physical constraints on AI industry expansion. This has moved beyond simple technical advancement to become a sustainability issue for the entire industry.