# China’s AI Buildout, Copilot’s Reset, and Fable 5’s Return

*By AI High Signal Digest • July 5, 2026*

China’s AI infrastructure economics, Microsoft’s Copilot consolidation, and Fable 5’s return led the day. The brief also covers Sakana’s ICML research, new multimodal product launches, Moonshot’s research-first strategy, and Alibaba’s reported Claude Code ban.

## Top Stories

*Why it matters: the biggest signals today were about infrastructure, distribution, and models proving themselves in live use.*

- **China’s AI compute push is scaling fast.** One market analysis said the national computing power network could attract **Rmb7tn** of investment in 2026, with roughly **Rmb2tn ($300bn)** of data-center spending over five years. A typical GW-scale campus is modeled as **50%+ inference**, but domestic chips still trail on performance: Nvidia is still seen holding **55%** overall share, and Huawei 910B/910C servers were said to produce only **1/6 to 1/3** of an H800’s daily token output. [^1]
- **Microsoft is merging consumer and enterprise Copilot into one app.** The August-targeted overhaul reportedly adds AI coding tools, paid AutoPilot agents, and add-ons like Copilot Cowork after cutting features customers were not using. Copilot had **20M paying users** by April, up from **15M** in January, but still trails ChatGPT’s **50M+** paid subscribers. [^2]
- **Fable 5 is back in public testing.** Arena said the model has returned to Battle Mode and Agent Mode and had previously ranked **#1** in Agent Arena; separate posts pointed to strong 3D-generation demos across 60+ hard tasks and one case where it chose propensity score matching in a retention analysis without being asked. [^3][^4][^5]

## Research & Innovation

*Why it matters: the strongest technical work focused on memory, efficiency, and better training data rather than just larger scale.*

- **Sakana AI brought a broad ICML slate.** Its 11 papers span multi-agent coordination, sparse LLMs, test-time scaling, long-term memory, and agent benchmarks. Highlights included **FwPKM** at about **75%** 5-needle NIAH accuracy at 128K context, **Doc-to-LoRA** for internalizing documents into model weights, and **TwELL** sparse kernels with **20%+** speedups on billion-parameter models. [^6][^7][^8][^9]
- **EfficientRollout targets RL’s biggest time sink.** The paper says rollout generation consumes nearly **70%** of LLM RL training time; its quantized self-drafter, roofline-based switch, and adaptive draft length produced up to **19.6%** faster rollout generation and **12.7%** faster training steps. [^10]
- **DolphinMath aims to make high-quality math data abundant.** QuixiAI released a generator for unlimited math problems from elementary to postgraduate level, with mechanically correct step-by-step solutions for pretraining, SFT, and RL. [^11][^12]

## Products & Launches

*Why it matters: new launches are turning multimodal performance and retrieval infrastructure into usable tools.*

- **Google launched Nano Banana 2 Lite and Gemini Omni Flash.** Nano Banana 2 Lite was priced at **$0.034 per 1K images** with four-second generation, while Gemini Omni Flash was priced at **$0.10 per second** for developer video generation and conversational editing. [^13]
- **LlamaIndex released Index v2 for agentic retrieval.** It exposes retrieve, read, grep, and find APIs so agents can navigate evolving knowledge bases; the legal-kb reference app adds project-scoped knowledge bases, visual citations, version control, and data export. [^14]
- **Dreamina Seedance 2.5 is coming to CapCut.** CapCut said it supports seamless generation and editing, up to **50 multimodal references**, and **30-second** scenes across web, desktop, and mobile. [^15]

## Industry Moves

*Why it matters: labs are differentiating through go-to-market choices and infrastructure strategy as much as model quality.*

- **Moonshot AI is staying research-first.** Its enterprise chief said Kimi will rely on partners for last-mile deployment instead of building a heavy services team; the company is reportedly raising **$2B** at about a **$30B** valuation, expanding via AWS, and says its KV-cache hit rate is above **90%**. [^16]
- **OpenAI’s reported Jalapeño chip points to the silicon race.** Posts said OpenAI unveiled its first custom AI chip, while a claimed **nine-month** design-to-tape-out timeline drew skepticism; a separate comment argued that at OpenAI’s scale, owning silicon is now necessary. [^17]

## Policy & Regulation

*Why it matters: model-access restrictions are starting to shape internal software policy at major companies.*

- **Alibaba is reportedly banning Claude Code at work starting July 10.** The company classified Anthropic’s coding agent as high-risk software after reports it contained checks for China-linked users; Anthropic already bars Chinese companies and foreign entities they own from using its models, and Alibaba is directing staff to its own Qoder tool. [^18]

## Quick Takes

*Why it matters: smaller updates still show where tooling, evals, and training methods are moving.*

- A team said it distilled **2.3M** Claude Fable 5 reasoning traces into **Qwen3-4B** and open-sourced the weights. [^19]
- Newer Claude models were reported to fail Pi’s edit tool more often than older versions, especially on tasks close to but not exactly on training distribution. [^20][^21]
- A practical MCP paper said tool-selection accuracy falls below **90%** after **10-30 tools**, while MCP itself adds little latency. [^22]
- **OctoTools** was accepted as an **ACL 2026 Oral**, positioning its training-free tool-use framework for wider attention. [^23]

---

### Sources

[^1]: [𝕏 post by @pequityresearch](https://x.com/pequityresearch/status/2073593220034908216)
[^2]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2073356431659827379)
[^3]: [𝕏 post by @arena](https://x.com/arena/status/2072423538641031372)
[^4]: [𝕏 post by @arena](https://x.com/arena/status/2073484893334691855)
[^5]: [𝕏 post by @_catwu](https://x.com/_catwu/status/2073439890482794966)
[^6]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2073393327622516843)
[^7]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2073394995428753820)
[^8]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2073394035646177742)
[^9]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2073393424091484165)
[^10]: [𝕏 post by @VukRosic99](https://x.com/VukRosic99/status/2073587668869337226)
[^11]: [𝕏 post by @QuixiAI](https://x.com/QuixiAI/status/2073630735420789093)
[^12]: [𝕏 post by @QuixiAI](https://x.com/QuixiAI/status/2073633208705437829)
[^13]: [𝕏 post by @dl_weekly](https://x.com/dl_weekly/status/2073451933310783972)
[^14]: [𝕏 post by @llama_index](https://x.com/llama_index/status/2072372216994374077)
[^15]: [𝕏 post by @capcutapp](https://x.com/capcutapp/status/2073261464065122562)
[^16]: [𝕏 post by @TechBuzzChina](https://x.com/TechBuzzChina/status/2073270599024042290)
[^17]: [𝕏 post by @thursdai_pod](https://x.com/thursdai_pod/status/2073618690613100639)
[^18]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2073458381906460967)
[^19]: [𝕏 post by @waterloo_intern](https://x.com/waterloo_intern/status/2073171123542573231)
[^20]: [𝕏 post by @mitsuhiko](https://x.com/mitsuhiko/status/2073488508816151038)
[^21]: [𝕏 post by @dbreunig](https://x.com/dbreunig/status/2073492460047876220)
[^22]: [𝕏 post by @TheTuringPost](https://x.com/TheTuringPost/status/2073528308411814054)
[^23]: [𝕏 post by @lupantech](https://x.com/lupantech/status/2073576304159670300)