AI News in December 2025: Devstral 2, Agent Standards and Data-Center Power

On December 9, 2025, I expect GPT-5.2 to arrive soon and look at prediction markets for its release date. I see a sharp fall in the market’s implied chance of a December 9 release and an increase for December 11. OpenAI subsequently announced GPT-5.2 on December 11.

I suggest that insiders may have moved the market. A price change, even a sudden one, does not establish who traded, whether they possessed private information or whether any law or rule was broken. The chart is evidence of changing prices, not proof of insider activity.

Mistral announced two coding models, Devstral 2 at 123 billion parameters and Devstral Small 2 at 24 billion, alongside the Mistral Vibe command-line assistant. I welcome available weights and tools that developers can run or integrate. Availability and pricing described in this December 2025 release may change.

The licensing distinction matters. Devstral Small 2 uses Apache 2.0, while Devstral 2 uses a modified MIT license, not standard MIT. The published license restricts use by companies or employers whose consolidated monthly revenue exceeded $20 million in the preceding month, so the larger model is not unrestricted for all users.

Mistral reports 72.2% for Devstral 2 and 68.0% for Devstral Small 2 on SWE-bench Verified. Its displayed third-party human comparison also shows Devstral 2 winning 21.4% against Claude Sonnet 4.5, tying 25.5% and losing 53.1%. These are reported results for specific evaluations and setups, not universal rankings.

OpenAI CEO Sam Altman appeared on The Tonight Show with Jimmy Fallon and answered broad questions about ChatGPT and AI. I take the need for basic explanations on a mass-audience show as a reminder that many people are still encountering these tools for the first time.

The Linux Foundation announced the Agentic AI Foundation, with Anthropic contributing Model Context Protocol (MCP), OpenAI contributing AGENTS.md and Block contributing its goose agent framework. MCP connects AI applications to tools and data; AGENTS.md provides project-specific instructions for coding agents. The foundation aims to support open, collaborative governance.

The announcement reports more than 10,000 public MCP servers and adoption across several AI products. I am pleased to see shared agent infrastructure move into a foundation. These counts and adoption descriptions are from the 2025 announcement rather than a measure of current usage or quality.

OpenAI also announced certificate courses for AI skills. I describe the idea as learning through ChatGPT and earning a credential; course coverage and any assessment requirements depend on the specific offering.

Boom announced Superpower, a 42-megawatt natural-gas turbine intended for data centers and based on its supersonic-engine technology. The company says launch customer Crusoe ordered 29 turbines totaling approximately 1.21 gigawatts and describes an order book exceeding $1.25 billion. Those are announced orders and specifications, not evidence that the capacity is already operating.

Boom CEO Blake Scholl shows a historical chart comparing electricity generation in China and the United States. I argue that power availability matters for AI infrastructure and like the proposal to use modular arrays of turbines. A natural-gas installation also depends on fuel supply, permitting, grid or site integration and its emissions; a turbine cannot simply be placed beside a data center without that supporting infrastructure.

Scholl describes discussing the concept with Sam Altman, a long-time Boom investor, before securing a launch order. He compares modular turbine arrays to blade servers, where individual units can be serviced or replaced without treating the entire installation as one machine.

I turn to investor Gavin Baker’s discussion of data centers in space. The attraction is more continuous solar exposure in a suitable orbit and optical links between satellites. Google’s Project Suncatcher explores such a constellation as an early research proposal, not an already operating orbital data center.

Baker describes space cooling as free and suggests radiators on the shaded side. I acknowledge that thermodynamics makes cooling more complicated. In vacuum, heat must be rejected primarily through radiation, requiring appropriately designed radiators and thermal systems; sunlight and battery needs also depend on the orbit and mission.

Light travels faster in vacuum than through optical fiber, but that fact alone does not establish that a satellite network would outperform every terrestrial data center. Link bandwidth, distances, pointing, ground communications, reliability and radiation all matter. I am excited by the concept while recognizing that viable deployment and sustainable power remain engineering questions.