Vol.37 · June 29, 2026
dera news AI Weekly Vol.37 | 2026-06-29 - This Week's AI News
🤖 dera news AI Weekly Vol.37
2026-06-29
This week's AI world in one sentence? OpenAI launched its first custom AI chip 'Jalapeño,' igniting the in-house AI infrastructure development race. Developed with Broadcom in just nine months, this inference-focused chip aims for gigawatt-scale deployment by late 2026, potentially transforming AI cost structures.
📊 What You Need to Know This Week
This week, developments surrounding the foundational infrastructure of the AI industry captured significant attention. OpenAI's announcement of 'Jalapeño,' its first custom AI chip developed in collaboration with Broadcom and specialized for large language model (LLM) inference, signals a clear intent to reduce reliance on NVIDIA and optimize computing resources in-house. This move directly impacts the cost reduction and performance enhancement of AI services, marking a crucial step towards diversifying future AI infrastructure.
Parallel to these efforts in strengthening proprietary infrastructure, AI talent mobility remains dynamic. Top researchers who contributed to Google's flagship AI model, Gemini, have successively moved to competitors like Anthropic and OpenAI, highlighting the intensifying talent acquisition battle within the AI sector. Furthermore, OpenAI's decision to limit preview access to its next-generation model, GPT-5.6 Sol, at the request of the U.S. government, suggests that high-performance AI technology is increasingly being treated as a critical national security issue, ushering in an era where its use may be regulated.
Meanwhile, the presence of the Asian AI ecosystem is growing. Alibaba's Qwen team open-sourced "language world models" that simulate agent environments, while China's Zhipu AI released GLM 5.2, an open-source model that rivals the performance of cutting-edge U.S. models on certain benchmarks at a significantly lower cost. These developments are filling gaps left by some U.S. labs constrained by export regulations, expanding the diversity of the AI ecosystem and the potential of open source. Japan's Sakana AI also introduced "Fugu," a multi-agent orchestration system that intelligently links multiple AI models, offering a new approach to mitigate single-vendor dependency risks.
Beneath the surface of technological competition, ethical and legal challenges are emerging. Anthropic's accusation against Alibaba's Qwen AI lab for large-scale, illicit extraction of its AI models, and its subsequent letter to the U.S. Congress, underscores the urgent need for intellectual property protection and the maintenance of a fair competitive environment in AI. In this context, OpenAI's expansion of its "Daybreak" cybersecurity program, which leverages AI to discover and fix vulnerabilities, represents a vital step towards the safe societal implementation of AI.
What we're watching closely is AI infrastructure self-development and the rise of open-source models. As technological competition between the U.S. and China intensifies, OpenAI's in-house chip development promotes supply chain diversification, while high-performance open-source models from Chinese players expand options for AI utilization that are not dependent on specific vendors.
💡 This Week's Actions
1. Explore the Potential for In-house AI Infrastructure Optimization (1 hour) OpenAI's venture into custom chip development signals a growing focus on reducing AI usage costs and boosting performance. Begin considering whether your current AI workloads are optimally served by existing infrastructure, or if custom solutions or alternative providers could improve cost-efficiency and performance in the long term. → OpenAI and Broadcom unveil LLM-optimized inference chip
2. Consider Leveraging Open-Source AI Models (2 hours) High-performance, low-cost open-source models like China's Zhipu AI GLM 5.2 are emerging. These models, operable on your own servers without vendor dependency, offer potential cost savings and flexibility, especially for automating agentic tasks. → China's Zhipu is closing in on top U.S. AI models with Anthropic and OpenAI held back → Qwen-AgentWorld: Alibaba open-sources 'language world models' that simulate agent environments across 7 domains
3. Experiment with Multi-Agent Systems for Operational Efficiency (1.5 hours) Sakana AI's Fugu is a system that links multiple AI models to automatically handle complex tasks. By providing advanced AI capabilities through a single API, it presents a significant opportunity for small to medium-sized businesses to streamline operations, particularly for multi-stage tasks like coding or information organization. → Sakana AI Launches Sakana Fugu: An Orchestration Model That Routes Tasks Across a Swappable Pool of Frontier LLMs
📰 This Week's AI Articles (9 stories)
1️⃣ OpenAI and Broadcom unveil LLM-optimized inference chip
🏷️ AI Models, AI Infrastructure What happened? OpenAI, in collaboration with semiconductor giant Broadcom, announced 'Jalapeño,' its first custom AI accelerator chip specifically designed for large language model (LLM) inference. Developed in a record nine months, it plans gigawatt-scale deployment with data center partners like Microsoft by late 2026. Our take OpenAI's move to develop its own AI infrastructure reduces its reliance on NVIDIA and will significantly impact AI's cost structure and supply chain. This could lead to greater diversity in AI service provision and potentially cut processing costs by up to 50%. 📎 Read more
2️⃣ Qwen-AgentWorld: Alibaba open-sources 'language world models' that simulate agent environments across 7 domains
🏷️ AI Models, Open Source What happened? Alibaba's Qwen team open-sourced Qwen-AgentWorld-35B-A3B and -397B-A17B, "language world models" that simulate agent environments across 7 domains (MCP, Search, Terminal, SWE, Android, Web, OS). Trained on over 10 million trajectories, they can be used as agent foundations or reinforcement learning simulators. Our take The fact that Chinese players are pushing the frontier of agent technology and offering it open-source is noteworthy. Their ability to simulate diverse environments will significantly influence future AI agent development and open new application possibilities. 📎 Read more
3️⃣ Sakana AI Launches Sakana Fugu: An Orchestration Model That Routes Tasks Across a Swappable Pool of Frontier LLMs
🏷️ AI Models, Multimodal AI What happened? Japanese startup Sakana AI launched 'Fugu,' a multi-agent orchestration system that intelligently links multiple cutting-edge LLMs and routes tasks via a single API. Fugu automatically assigns complex tasks to the optimal AI, also positioned as a hedge against single-vendor dependency risks. Our take Fugu has the potential to reduce AI single-vendor dependency risks and allow SMBs to leverage advanced AI features without complexity. It could be a powerful tool for operational efficiency, especially for businesses with multi-stage workflows. 📎 Read more
4️⃣ OpenAI expands Daybreak: full GPT-5.5-Cyber release and 'Patch the Planet' open-source security program
🏷️ AI Models What happened? OpenAI expanded its 'Daybreak' cybersecurity program, combining AI models like 'GPT-5.5-Cyber' and 'Codex Security' to efficiently discover and fix software vulnerabilities. It has already scanned over 30,000 codebases and automatically fixed more than 500,000 vulnerabilities. Our take Efforts by AI developers to enhance the security of AI itself are crucial for its safe societal implementation. For developers and businesses, AI-powered security measures will be an important challenge moving forward. 📎 Read more
5️⃣ Anthropic Economic Index 'Cadences' maps how 49% of jobs already offload a quarter of tasks to Claude
🏷️ AI Models What happened? Anthropic's latest report, "Economic Index 'Cadences'," analyzed Claude AI assistant usage. Data shows a shift from work-related tasks to personal content like emotional support, investment advice, and agent design, particularly on weekends and evenings. Our take This report highlights how AI is not merely a business tool but is deeply integrating into personal learning, skill development, and new idea generation. Companies should consider a multifaceted view of employee AI usage, beyond just operational efficiency, to include skill enhancement and new business opportunities. 📎 Read more
6️⃣ China's Zhipu is closing in on top U.S. AI models with Anthropic and OpenAI held back
🏷️ AI Models, Open Source What happened? Chinese AI startup Zhipu AI released its latest open-source model, 'GLM 5.2,' generating significant buzz in Silicon Valley. This model reportedly rivals Anthropic's cutting-edge models on certain benchmarks, offering similar performance at about one-fifth the cost, and is particularly strong for "agentic tasks." Our take As U.S. export restrictions continue, Chinese open-source models are rapidly advancing, providing high-performance, low-cost alternatives. This is a crucial development for reducing vendor lock-in risks in AI usage and fostering a diverse AI ecosystem. 📎 Read more
7️⃣ Anthropic accuses Alibaba of campaign to 'brazenly' and 'illicitly' extract AI capabilities
🏷️ AI Models What happened? Leading AI developer Anthropic accused operators linked to China's Alibaba Qwen AI lab of using thousands of fraudulent accounts to "brazenly and unlawfully extract" capabilities from its AI models. Anthropic called it the largest "distillation attack" to date and sent a letter to the U.S. Congress. Our take While AI model "distillation" is technically possible, when done illicitly, it threatens intellectual property protection and fair competition. This incident underscores that ethical and legal discussions and countermeasures are urgently needed as AI technological competition intensifies. 📎 Read more
8️⃣ OpenAI unveils GPT-5.6 Sol, Terra and Luna — but limits launch to ~20 'trusted partners' at US government request
🏷️ AI Models What happened? OpenAI began a limited preview of its next-generation AI models, the GPT-5.6 series (flagship Sol, balanced Terra, fast/cheap Luna). However, at the request of the U.S. government, preview access to the flagship Sol model is restricted to "approximately 20 trusted partners," marking the first time OpenAI has self-limited a flagship model's public release. Our take OpenAI's self-imposed restriction on its flagship model's release highlights the current reality of AI technology as a critical national security issue. It suggests an era where the use of high-performance AI will be regulated, significantly impacting future AI development and deployment. 📎 Read more
9️⃣ AI researchers continue to leave Google for its rivals
🏷️ AI Models What happened? Top-tier researchers from Google's AI division are reportedly leaving for competitors like Anthropic and OpenAI. This includes Jonas Adler and Alexander Pritzel, who played key roles in developing Google's flagship AI model, Gemini, with four senior researchers departing in the past week alone. Our take The competition for top talent in the AI industry is intensifying. With OpenAI and Anthropic potentially heading towards initial public offerings, this talent migration is likely to continue, significantly influencing each company's technological development roadmap. 📎 Read more
📚 Editor's Note
This week, movements surrounding the "foundations" of AI were particularly prominent. OpenAI's foray into in-house chip development to enhance AI infrastructure autonomy, coupled with a notable outflow of top talent from Google, indicates an intensifying competition among major players. This shows that AI technology's evolution is not just about improving model performance but also unfolds across multi-layered aspects such as supporting hardware, human talent, and national strategies.
Furthermore, the rise of open-source models from Chinese players and the issues raised regarding illicit AI model extraction suggest that technological competition spans diverse dimensions, including supply chains, intellectual property, and ethics. The future of AI will not be monopolized by a single company or nation but will be shaped by a complex interplay involving various actors. We will continue to monitor how these structural changes in the industry will impact future business and society.
Until next week, we'll continue to deliver useful insights and food for thought. dera news editorial team
📬 About this newsletter