Dec 23, 2025 · 24-Hour AI Briefing: ERNIE tops LMArena in China, cloud spend hits $102.6B, and DingTalk’s Agent OS wants to become the enterprise “new substrate”
The past 24 hours weren’t about one splashy launch—they were about three compounding battles: models fighting for user preference and “conversation quality,” cloud platforms shifting from raw scale to full-stack AI ecosystems, and enterprise software moving from point tools to agent-native orchestration. The common thread: rankings and announcements are only the opening move—durable advantage comes from repeatable, scalable productivity.
1. Baidu’s ERNIE-5.0-Preview-1203 scores 1451 on LMArena, ranking No.1 in China
Baidu’s new ERNIE-5.0-Preview-1203 posted a 1451 score on LMArena, taking the top China ranking. The model is reported to perform especially well in creative writing and complex task-oriented conversations.
Commentary:
LMArena is fundamentally a human-preference, head-to-head win-rate system. A high score is a meaningful signal that “chat experience” and subjective quality have improved—but it does not automatically translate to absolute leadership in reproducible benchmarks like reasoning, math, or coding.
ERNIE-5.0-Preview-1203 is positioned as a native multimodal foundation model released in Nov 2025, reportedly at 2.4T parameters with unified modeling that can process and generate text, images, audio, and video.
China is no longer a spectator in the global model race—it is increasingly shaping the pace and the rules. The score is momentum; the real competition is whether creative and complex-task strength can be turned into reliably deliverable, scalable productivity. Expect competition among ERNIE, ChatGPT, Qwen, Gemini, and Grok to intensify.
2. Omdia: Global cloud infrastructure spend reached $102.6B in Q3 2025 (+25% YoY); AWS leads with 32% share
An Omdia report estimates global cloud infrastructure services spending hit $102.6B in Q3 2025, up 25% year over year. AWS led with a 32% market share and reported 20% revenue growth.
Commentary:
Cloud is moving beyond “who has the most compute” into a full-stack contest: AI platform capabilities, model ecosystems, and industry solutions. AWS, Azure, and Google Cloud together account for roughly two-thirds of the market, reinforcing an entrenched top-tier.
Even with Azure and Google Cloud growing quickly, AWS still controls the center of gravity due to its scale. But AWS growing slower than the overall market is the signal to watch: it must translate AI platform depth into scalable, billable revenue faster.
Against Microsoft and Google’s aggressive AI-native architecture investments, the real question is not whether AWS has AI—it’s whether AWS can operationalize AI into compounding growth while defending share.
3. DingTalk launches 20+ AI products plus enterprise AI hardware, and unveils Agent OS for “agent-native work”
Alibaba’s DingTalk announced more than 20 AI products, including enterprise AI hardware, and introduced an agent-oriented “work intelligence operating system” called Agent OS.
Commentary:
Agent OS is not a traditional operating system. It’s better understood as a foundational runtime and coordination layer designed for AI agents to operate and collaborate.
By pushing a “product matrix + operating system” story, DingTalk is signaling a shift from selling point features to defining how agents will orchestrate people, workflows, data, permissions, and devices inside the enterprise.
The hardware angle is strategically coherent: embed AI entry points into high-frequency contexts—meeting rooms, reception desks, storefronts, desks—to solve the “last-mile adoption” problem. But the risks are equally real: deployment and maintenance complexity, cost models, and security/compliance boundaries will determine whether this can scale.
Could DingTalk become one of the keys to the next generation of enterprise operating systems?
For more context, here are the most important AI events from the past 72 hours: