AMD Briefly Tops $1 Trillion as Meta's Muse AI Agent Fuels CPU Demand Bet

Microsoft CEO Satya Nadella said the United States and China should establish shared norms on AI development and safety, including a direct leader-to-leader notification channel for cybersecurity attacks, as Chinese President Xi Jinping visits Washington to meet with President Trump. Speaking with CNBC's Deirdre Bosa at APEC, Nadella said both countries care about the same things: diffusing the technology so citizens get value and doing so safely, and that the more norms they set, with the US in particular able to lead, the better off the world is. Nadella also discussed Microsoft's new Copilot, which has about 30 million customers so far and rolls out in the coming weeks with three modes: chat, code, and Autopilot, an agentic mode connected to a user's apps, email, and workflows. He said trust will be Microsoft's biggest issue, pointing to Agent 365 as the company's fastest-adopting enterprise product for observing, governing, and securing AI agents, and warned that token costs mean the current level of subsidy cannot last, so customers must ultimately see return on investment. Nadella attended the state dinner during the summit, and Bosa said any direct US-China communication line emerging from the talks would be a positive step.

Microsoft CEO Satya Nadella said the company's new cloud-based Autopilot feature represents the next generation of automation in the enterprise, with high ROI expected for businesses that adopt it. In an interview with Deirdre Bosa, Nadella described Autopilot as an agent with its own identity that completes jobs and hands off work, while the inner agent loop lives inside applications such as Excel. The new Copilot update has three modes: chat, code, and Autopilot, and Microsoft is positioning Copilot as the operating layer for enterprise AI, with adoption at about 30 million customers so far. Nadella said the bet is that companies should not have to track the fast-moving model race, and that Microsoft will serve the right model for each task, using more capable models for hard tasks and cheaper models for easier ones.