Microsoft Pulls Back from Oracle AI Compute as OpenAI Valuation Hits $1.25 Trillion

Key Takeaways

• Microsoft reportedly phasing out Oracle AI compute, reassessing cloud partnership
• OpenAI raises $11 billion at a $1.25 trillion valuation
• Sources say CSRC has asked for a slower pace of AI and chip IPOs

Compute Procurement: Microsoft Scales Back Oracle Resources

Microsoft is gradually winding down its rental of Oracle's artificial intelligence computing resources, according to Bloomberg, which cited people familiar with the matter. The move is part of a reassessment of the two companies' multibillion-dollar cloud partnership. The report did not disclose the scale or timeline of the phase-out, nor whether Microsoft will shift the workloads to its own infrastructure or to other suppliers. (Source: kaopunews / Bloomberg)

The tip itself carries limited information, but the direction is worth noting: against the narrative of chronic AI compute shortages, how leading cloud providers weigh third-party compute leases is becoming an observable variable. If more similar adjustments follow, it would suggest compute expansion is shifting from "take as much as you can get" to a reallocation based on cost and utilization. If it proves isolated, it should not be overinterpreted.

Capital Side: OpenAI at $1.25 Trillion, IPO Pace Faces New Variables

OpenAI has raised another $11 billion in a new funding round at a $1.25 trillion valuation, making it one of the most valuable companies in the world, according to Bloomberg as cited by The Verge. The company previously raised $122 billion at an $852 billion valuation in March this year. The report said OpenAI is still seeking more investment, with the total potentially rising to $20 billion. Its initial target was to raise up to $30 billion at a $1.4 trillion valuation, later adjusted due to the tech stock selloff and midterm election results. (Source: The Verge / Bloomberg)

At the same time, sources told Lianhe Zaobao this week that the China Securities Regulatory Commission has issued window guidance to investment banks, asking them to slow the IPO process for AI and chip companies. The two reports point to different markets, but share a common thread: the financing pace of AI assets is being shaped by both the valuation environment and regulatory stance. (Source: kaopunews / Lianhe Zaobao)

Echoing this, CNBC's Cramer segment put next week's focus on earnings from CoreWeave and Nebius as well as key inflation data, arguing that AI data center stocks face a new test. (Source: CNBC Technology)

Research Side: Multi-Agent Collaboration and Reviewer Reliability Draw Heavy Attention

Several new papers on arXiv cs.AI that day centered on multi-agent systems. "Who Should Know What?" proposes an "empirical information design" framework, testing multi-resource trading, bilateral bargaining, and double auctions across ten frontier models. It found that the effectiveness of finer information disclosure depends heavily on the model: in a trading scenario, disclosing precise valuations to GPT-6 Sol cut its realized gains from 98% of the maximum to 91%, while Claude Sonnet 5 rose from 72% to 86%. (Source: arXiv cs.AI)

"PromptOps" defines the performance degradation of LLMs as reviewers as "reviewer drift" and proposes a closed-loop framework to determine when to optimize, by what method, with which feedback records, and how to update prompts. It verifies that iterative optimization can restore performance after controlled distribution shifts. (Source: arXiv cs.AI) Another paper, "The Two-Ledger Model," addresses cost accounting for enterprise multi-agent systems, noting that billing data stays at the settlement level and cannot answer "how much AI cost did this order incur." It proposes combining a settlement ledger with an attribution ledger. (Source: arXiv cs.AI)

Engineering in Practice: From Chemical Design to Lego Animation

"ChemE-as-Code" offers a verifiable agent engineering case: an open-source environment lets LLM agents call tools for catalysis, process simulation, equipment layout, HAZOP, and lifecycle assessment. Computation runs outside the LLM, every value records its source, and automated checks report inconsistencies by file and line number. In an ammonia case, the agent developed a written task into a preliminary basic engineering package, with API costs for the design session at $224. The authors also note that fault-injection tests show the checks cover only what the rules reach, and that the design has not been reviewed by an independent engineer. (Source: arXiv cs.AI)

The LDraw Mecha project on Hacker News demonstrates another type of application: an agentic LLM reads annotated Lego model source code, infers submodel structure, identifies mechanical mechanisms, plans animations, and executes them in Blender, outputting .blend files with skeletons and animations. (Source: Hacker News)