High-voltage electrical substation with steel structures and power lines at dawn

Daily Brief: OpenAI pledges grid-flexible data centres, Meituan's LongCat-2.5, Abu Dhabi deals

OpenAI's Grid Protectors plan for flexible-power data centres, Meituan's 1.6T LongCat-2.5 trained on Chinese chips, Abu Dhabi's OpenAI deals, and AWS agent workflows.

OpenAI wants its data centres to give power back when the grid is strained

OpenAI has launched the Grid Protectors Program, a commitment that its future compute campuses will cut their electricity draw when local grids are under stress, and in some places feed power from on-site generation or batteries. The pilot is the planned Effingham County campus in Georgia, where OpenAI says it will pay the full infrastructure and service costs and its Georgia Power deal includes up to 1,000 MW of flexible demand response, roughly the average demand of 700,000 homes. Why it matters: with US data-centre capacity projected past 100 GW by 2030, the political fight over who pays for AI's power is getting louder. Flexible load is the industry's answer, and if utilities accept it, campuses could connect faster without forcing as many costly peak-capacity upgrades onto household bills.

What else matters

Meituan ships LongCat-2.5, trained entirely on Chinese chips. Meituan's new LongCat-2.5 is a 1.6-trillion-parameter native multimodal model aimed at long-horizon agent and coding work, and the company says training ran fully on domestic compute. It has a 1M-token context and is already live via API and on Vercel's AI Gateway.

Abu Dhabi signs four deals with OpenAI. After a meeting with Sam Altman, Abu Dhabi entities signed agreements spanning government, energy and investment, per the Abu Dhabi Media Office. ADNOC becomes the first energy company to deploy OpenAI's frontier models, and a "Machine-Readable Government" pilot turns rules into traceable, AI-checkable logic with officials keeping final say.

AWS lets agents call Step Functions workflows as tools. Per the AWS weekly roundup, Step Functions workflows can now be exposed to agents over MCP through Bedrock AgentCore Gateway. The agent decides when to run a process, but the steps, such as approvals before a refund, run in a fixed order the model can't improvise. GPT-6.1 Terra also went GA on Bedrock.