AI industry shakeups
AI chatter this week centers on a stealth Chinese 1‑trillion‑parameter model, Anthropic outspending OpenAI in enterprise deployments, Mistral launching Forge for custom models, plus NVIDIA GTC positioning around agentic AI — a fast pivot toward agentic/autonomous layers in enterprise ( ). At the same time, self‑evolving models like Minimax M2 are being discussed for recursive research workflows that could outpace compute improvements — a trend creators and startup founders should watch for tooling changes (x.com).
An anonymous model called “Hunter Alpha” appeared on OpenRouter on March 11, advertised with roughly 1 trillion parameters and a context window of up to one million tokens, and Reuters’ testing reported the model described itself as “a Chinese AI model” with a May 2025 data cutoff. (tech.yahoo.com) Xiaomi’s MiMo team later stated Hunter Alpha was an early internal test build of MiMo‑V2‑Pro and that the system is intended as a “brain” for agent-style multi‑step tools, a disclosure traced by TechRepublic to Xiaomi’s announcement and Reuters reporting. (techrepublic.com) Ramp customer‑spend data cited by Axios shows Anthropic capturing over 73% of spending among companies buying AI tools for the first time, a metric Axios reported on March 18 as evidence of Anthropic’s enterprise momentum. (axios.com) Menlo Ventures’ Mid‑Year LLM Market Update found enterprise LLM API spending rose to $8.4 billion and reported Anthropic at roughly 32% enterprise share versus OpenAI’s ~25%, a market‑share flip Menlo published in its 2025 market update. (menlovc.com) At NVIDIA GTC (March 16–19, 2026) vendors framed a push toward agentic and “AI factory” stacks—NVIDIA’s event page and multiple recaps highlighted agentic systems as a central theme while Jensen Huang’s keynote outlined new hardware and software layers for autonomous AI. ( ) Mistral announced Forge at GTC as an enterprise platform for training custom frontier models on proprietary data with pre‑training, reinforcement‑learning alignment, synthetic‑data generation and lifecycle tooling, positioning the company to offer full model ownership to customers rather than just RAG/fine‑tuning wrappers. ( ) MiniMax says its new M2.7 release is a “self‑evolving” agentic model that participated in parts of its own training and can automate an estimated 30–50% of reinforcement‑learning research workflows, a claim covered by VentureBeat and described in MiniMax’s product notes; the broader research community is concurrently publishing surveys and papers cataloging “self‑evolving” or recursive agent approaches. ( )
