Your daily brief on the most important AI news worldwide.
🔥 Top 3
1. Nscale is raising $3.5B in pre-IPO financing
The AI compute provider is in talks to raise $3.5 billion ahead of a potential IPO, hot on the heels of a $45 billion compute deal with Anthropic. From wholesale capacity contracts to a final private round, the capital loop around AI cloud infrastructure is closing fast — private markets are already pricing compute platforms at what they expect them to be worth in public markets.
2. Thinking Machines in talks for $1B at a $40B valuation
Accel is reportedly in talks to lead a roughly $1B round for Thinking Machines at a $40B valuation — against an annualized revenue run rate just past $100 million. The premium on "top talent plus enterprise agents" keeps climbing, but the yawning gap between revenue and valuation puts a hard question on the table: how much money can agents actually make?
3. Crusoe raises $3B at a $30B valuation
The data center developer closed a $3B round shortly after reportedly securing a $13B contract with Jane Street. The "AI factory" model — bundling energy, land and compute — now has full capital-market backing: compute scarcity has shifted from a model-layer anxiety to balance-sheet expansion at the infrastructure layer.
🌐 More news
- China's first office-agent user behavior report is out: Beijing leads the country in users, and over 12% of users are overseas — office agents are past the novelty stage, and the overseas share hints at real international traction for Chinese agent products. Details
- Alibaba's agent product Qianwen Office launched the industry's first "multi-player workbench," moving conversational agents from solo use into team collaboration — the next battleground for agent products is organizations, not individuals. Details
- Startups are bringing in-context learning to robotics: longer multimodal context lets machines pick up new tasks on the spot — embodied AI is copying the LLM playbook, and "context scale" is the new scaling variable. Details
- Enterprise AI startup Wonderful is now valued at $5B after three rounds totaling $800M in two years — enterprise AI budgets are moving from "pilots" to "infrastructure," and vertical agent vendors are the new favorites. Details
- Anthropic released Fable 5.1 and cut cache-read pricing, calling it comparable to Mythos 5.1 — a frontier lab joining the price war means inference cost is now a scale game: good for the application layer, squeezing for second-tier model vendors. Details
- A lawsuit may force the administration to reveal the secret rules federal agencies use for frontier AI safety testing — if the black box opens, vendors and the public finally see what the government considers "safe," with direct consequences for the next round of regulation. Details
- Google released Gemini 3.8 Flash, its third Flash model in six weeks, focused on long-horizon reasoning and security — Pro updates look paused while Flash iterates rapidly; the mid-tier model race is now hotter than the flagship one. Details
Trend watch
- Infrastructure capital is sprinting toward the public markets. Nscale's pre-IPO round, Crusoe at $30B, and Nvidia's Hugging Face acquisition all point the same way: AI infrastructure has moved from "burning money to build" to "getting priced by capital" — IPOs are about to become the main stage of the compute story.
- Reliability is the new competitive frontier. Days after ChatGPT, Claude, Grok and Gemini went down almost simultaneously, Astra's launch locked paying users out and drew a public apology from Altman. As capabilities converge, stability and capacity planning — not benchmarks — are becoming the reason users switch products.
- The inference price war goes deep. Anthropic cutting prices with Fable 5.1, Google shipping three Flash models in six weeks, and Qwen's budget positioning are compressing model-layer margins — and pulling forward the cost turning point for the application layer.
Disclaimer: For reference only; exercise caution for investment or decisions.