H100

谷歌推出DiffusionGemma模型，采用扩散机制实现文本生成速度提升4倍，支持本地运行。

NVIDIA has released a 2.6B parameter open-source world model that can run on a single GPU, making it accessible for various applications.

The Parameter Golf challenge demonstrated the potential of AI-assisted research, including training optimization, quantization, test-time and evaluation strategies, and new modeling ideas.

Cohere launches Command A+, its fastest and most powerful model that combines high-performance agentic AI with efficient deployment running on as few as two H100s.

Andrej Karpathy points out that suggesting users 'boot up an 8XH100' in tutorials instantly blocks most learners, exposing a systemic neglect of hardware accessibility in AI education.

The article proposes the need for a coworking space dedicated to vibecoding with access to a local H100 GPU network.

Cognition, Mercor, Etched, and AnthropicAI are hosting a one-day hackathon in San Francisco with $100k total prize pool, including $50k for the winner, and all accepted teams receive 8x H100 GPUs, Anthropic credits, and Cognition API access.

Cursor and SpaceXAI announced a partnership to train a significantly larger model from scratch using Colossus 2's million H100-equivalent GPUs, with 10x total compute, expecting a major leap in model capability.

The GPU shortage crisis continues to worsen: H100 prices are higher than three years ago, major AI labs have locked up multi-year supply, leaving university researchers and individual developers struggling to access compute.