
OpenAI is putting more than S$300 million into Singapore and building a 200-person applied AI team, a sign that model companies are moving closer to government and enterprise deployments.

Google’s I/O updates push Search beyond answers and links into agents, bookings, generated interfaces, and personalized task flows.

Alibaba unveiled the Zhenwu M890, a new AI accelerator aimed at agent workloads, as China’s cloud giants try to reduce their dependence on Nvidia hardware.

Andrej Karpathy is joining Anthropic’s pre-training team, a notable talent move that points to how frontier AI labs are trying to improve model development itself, not just buy more compute.

ASML expects the first products made with its High-NA EUV machines within months, a sign the next chip-manufacturing step is moving out of the lab.

Google is pairing its TPU stack with Blackstone's infrastructure capital to create a new U.S. AI cloud venture. The deal shows how AI compute is becoming a real estate, power, and financing problem as much as a chip problem.

OpenAI and major chip partners released MRC, an Ethernet-based protocol meant to keep giant AI training clusters running when networks fail.

OpenAI says two employee devices were hit by the TanStack npm supply-chain attack. The bigger lesson is that trusted build pipelines can still ship compromised packages.

YouTube is widening access to its AI likeness detection feature, giving adult users a way to find and challenge videos that copy their face.

Google’s new Googlebook category is a bigger move than another AI laptop pitch: it hints at Android becoming a serious desktop platform.

GitHub’s MCP Server can now run secret scanning from AI coding tools, moving credential checks earlier in the development workflow.

Google has added Multi-Token Prediction drafters to Gemma 4, promising up to 3x faster local inference without changing the model’s answers.