OpenAI Unveils Jalapeño Chip; NVIDIA Boosts MoE Training, ZAI Tops Open Models

Key Takeaways

1. OpenAI announced its first self-designed AI chip "Jalapeño," co-developed with Broadcom, aimed at scaling its full-stack platform and democratizing AI.
2. NVIDIA released NeMo AutoModel powered by Transformers v5, boosting MoE training throughput by 3.4-3.7x, and dominates the latest TOP500 supercomputing list.
3. ZAI (GLM-5.2) has reportedly overtaken DeepSeek as the world's leading open model, according to industry commentary.
4. Hugging Face continues to expand its data ecosystem with HIW-500, the largest open-source humanoid robot teleoperation dataset, and reports 60K public robot datasets on its platform.
5. New agent tools emerged: Hugging Face's CUGA open-source agent harness, xAI's Grok Build plugins (Firecrawl, MongoDB), and Anthropic's Claude Tag for Slack collaboration.
6. Ethan Mollick highlights organizational risks of Claude Tag, including knowledge lock-in, security, and cost control, urging strategic AI deployment.

1. AI Infrastructure and Hardware

1. OpenAI announced its first self-designed AI chip, Jalapeño, developed in collaboration with Broadcom. The chip is optimized for LLM workloads powering ChatGPT, Codex, API, and future agent products. OpenAI stated that chips are the foundation of the AI economy and that custom silicon helps scale its platform and broaden AI accessibility. — via 1
2. NVIDIA released NeMo AutoModel, based on Hugging Face Transformers v5, which delivers 3.4-3.7x training throughput improvements for Mixture-of-Experts (MoE) models through Expert Parallelism, DeepEP, and TransformerEngine. — via 1 2
3. NVIDIA dominates the latest TOP500 supercomputing list, powering 81% of all systems and 89% of new systems. Its AI training throughput is 2x that of all other platforms combined, and inference throughput is 3x. Additionally, NVIDIA holds the top 8 most energy-efficient systems on the Green500 list. — via 1

2. Open Models and Data Ecosystem

1. According to swyx, ZAI (GLM-5.2) from Zhipu AI has overtaken DeepSeek as the top open model after its January IPO, and is returning to San Francisco for the AI Engineer World's Fair. — via 1
2. Hugging Face reported that its platform now hosts 60,000 public robot datasets, up from 1,000 in early 2025, with streaming speeds up to ~1326 MB/s. It also released HIW-500, the largest open-source humanoid robot teleoperation dataset, featuring 500+ hours, 23K+ episodes, 10+ TB of data covering 10+ household tasks across 12 homes. — via 1 2
3. Hugging Face emphasizes its role as the best place to store data (public and private), especially for robotics and video AI. — via 1

3. Agent and Developer Tools

1. Hugging Face open-sourced CUGA, an agent harness that allows developers to focus on building agents rather than infrastructure. — via 1
2. xAI's Grok Build Marketplace launched official plugins for Firecrawl (web search, crawl, interaction via CLI) and MongoDB (query data, optimize indexes, manage databases). — via 1 2
3. Anthropic introduced Claude Tag, allowing teams to add Claude to Slack channels, tag it, and delegate tasks. However, Ethan Mollick warned of risks: Claude accumulates tacit knowledge, creating lock-in and security vulnerabilities; its token-based billing can lead to unpredictable costs. Mollick argues that organizations need both general AI exploration and dedicated AI build labs. — via 1 2 3
4. Runway released Seedance 4K, Seedance Mini, and Kling 3.0 Turbo with a 30% discount, and added an ad localization feature that generates multi-language versions from a single image. — via 1 2

4. Video and Creative AI

1. NVIDIA's FastVideo achieves 1.8-second end-to-end generation of a 5-second 480P video on a single RTX 5090, using the FastWan-QAD model. — via 1
2. Runway's latest Seedance models and ad localization feature (as detailed above) highlight the rapid progress in AI-driven video creation and localization. — via 1 2