The Big Bets & Power Shifts

The race for AI supremacy is being fought with checkbooks and compute contracts. In the biggest move, Google plans to invest up to $40 billion in Anthropic in a mix of cash and compute credits. This monumental deal, coming just after Anthropic’s limited release of its cybersecurity-focused “Mythos” model, underscores the frantic scramble to secure both talent and the vast computational capacity needed to train frontier models. Source

Meanwhile, the chip landscape is getting scrambled. Meta has signed a deal for millions of Amazon’s homegrown AI CPUs (not GPUs), earmarked for powering AI agentic workloads. This signals a potential new front in the silicon wars, where specialized CPUs for agent coordination could become as sought-after as GPUs for model training. Source

The Model Frontier & Creator Tools

On the model front, China’s DeepSeek has previewed its new V4 model, claiming it “closes the gap” with leading frontier models from US rivals like OpenAI and Anthropic on key reasoning benchmarks. The open-source model shows particular strength in coding, intensifying global competition. Source

For creators, ComfyUI has hit a $500 million valuation after raising $30 million. The investment surge highlights growing demand for granular control over generative AI processes, as professionals move beyond simple prompts to orchestrate complex image, video, and audio generation workflows. Source

Infrastructure & The Human Element

As AI agents proliferate in enterprises, a new problem emerges: coordination. A commentary argues that to avoid “automation waste,” companies need interaction infrastructure—a governed framework for agents to communicate context and coordinate tasks across different systems. Source

The demand for running models locally is also causing hardware ripple effects. Apple’s Mac mini is sold out and spawning marked-up listings on eBay, driven by enthusiasts and developers favoring its compact form for running local AI models, highlighting the consumer-grade hardware crunch. Source

In Other News

• Research: MIT scientists have built the world’s largest open collection of Olympiad-level math problems (30,000+), creating a tougher benchmark for AI reasoning and a resource for students. Source
• Policy & Ethics: A report details how Project Maven’s AI targeting systems have taught the military to rely on AI, accelerating operations like recent strikes on Iran. Source Separately, the World Press Photo contest has reaffirmed that AI-generated images are not eligible, defining “a photo” as a moment captured by light, not generated by algorithms. Source
• Talent & Gossip: The talent war continues as Meta and Thinking Machines Lab engage in a two-way poaching street. Source And the Musk vs. Altman lawsuit is set to begin, turning their feud into a legal spectacle. Source

---

Editorial Take: Today’s news paints a picture of an industry entering a hyper-consolidation phase, where victory seems to belong to those who can lock down the most compute, data, and capital. The $40B Anthropic deal isn’t just an investment; it’s a fortress being built. Yet, amidst these titanic clashes, counter-narratives persist—from open-source models closing the gap to creators demanding more control. The central tension is clear: will AI’s future be shaped by a few resource-hoarding giants, or can more open, distributed, and specialized approaches carve out lasting space?