Good Enough Wins: China's Cheap, Open Models Move In on American AI
Chinese models like Kimi K3 and DeepSeek V4 are only a few benchmark points behind the American frontier and up to 100 times cheaper to run. They have already taken the commodity layer of the market, and the surprise winners are the US software firms that were supposed to be AI's victims.
Read full story →A Model Built to Defend, Not Attack
Microsoft's first in-house cyber model scores zero on writing exploits, on purpose. It is built to find and patch flaws, handle 90 percent of the work, and hand the hardest cases to a cheaper OpenAI model. The 95.95 percent benchmark headline, though, deserves a closer look.
Read full story →In the AI Buildout, Power Is the New Chip
AMD just made the biggest infrastructure commitment in its history without buying a chip: 530 megawatts of capacity from a former bitcoin miner. Power, land and even financing have become the AI battleground, as AMD works around Nvidia's software moat and Nvidia turns itself into a lender.
Read full story →Two Tracks for Governing AI Take Shape in the Same Week
China launched a 29-nation AI bloc in Shanghai while OpenAI and Anthropic quietly helped Washington draft a closed federal review framework. Two blueprints for governing AI, with almost nothing in common, and a wave of open-weight models neither of them can reach.
Read full story →