June 27 is a day of AI infrastructure and governance stories bracketing the usual model-release cadence: OpenAI and Broadcom unveiled the Jalapeño inference accelerator chip just as the White House reportedly asked OpenAI to delay its next frontier model launch, underscoring how compute and politics are now intertwined. On the release side, Liquid AI shipped a tiny 230M-parameter LFM2.5 model for edge use, Krea open-weighted a 2-second enterprise image generator, and Vercel pushed AI SDK 7 for agent orchestration. Anthropic is retooling Claude for the workplace with an always-on Slack coworker and a framework for human-agent teams, while Lilian Weng's deep dive on scaling laws and a new Axios report on AI data centers' water demands both point to the same theme: the industry is shifting focus from raw model capability to the messier economics, infrastructure, and governance of deploying AI at scale. For engineers, the throughline is that the interesting problems are moving from 'can we build a bigger model' to 'how do we run, govern, and pay for the ones we have.'

• OpenAI and Broadcom launch Jalapeño, a performance-per-watt optimized LLM inference chip.
• White House reportedly asks OpenAI to delay its next frontier model deployment.
• Liquid AI ships a 230M-parameter foundation model built for edge devices.
• Anthropic retires Claude-in-Slack for an always-on agentic Slack coworker.
• Water is emerging as the next AI data center flashpoint alongside energy.