Last Week in AI #326 - Qualcomm AI Chips, MiniMax M2, Kimi K2 Thinking

# Qualcomm Announces AI Chips to Compete with AMD and Nvidia — Stock Soars 11%

[![image](https://blog.aitoearn.ai/content/images/2025/11/img_001-6.jpeg)](https://substackcdn.com/image/fetch/$s_!tkew!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2773e195-7d24-4120-a6dd-031adf1542e8_800x418.jpeg)

## Overview  
Qualcomm has unveiled two new **data center AI accelerators**:

- **AI200** — shipping in 2026  
- **AI250** — shipping in 2027  

This marks a **major pivot** away from Qualcomm’s mobile/wireless roots.  
Both will be available in **fully liquid‑cooled server racks** — competing with **Nvidia** and **AMD** rack‑scale solutions clustering up to **72 accelerators per system** for advanced AI model training/deployment.

The chips leverage **Qualcomm’s Hexagon NPUs**, originally from smartphones, scaled for **high‑performance data center workloads**.  
Following the announcement, **Qualcomm shares jumped 11%**, reflecting investor confidence in its AI strategy.

---

### Market Impact  
Competition in the AI data center segment will intensify.  
Liquid‑cooled, multi‑accelerator racks are emerging as the **standard** for both **AI training** and **inference**.

---

### Related Links  
- [MiniMax Releases MiniMax M2: A Mini Open Model Built for Max Coding and Agentic Workflows](https://www.marktechpost.com/2025/10/28/minimax-open-sources-minimax-m2-a-mini-model-built-for-max-coding-and-agentic-workflows-at-8-claude-sonnet-price-and-2x-faster/)  
- [Alibaba-backed Moonshot releases second AI update in four months as China’s AI race heats up](https://www.cnbc.com/2025/11/06/alibaba-backed-moonshot-releases-new-ai-model-kimi-k2-thinking.html)

---

## AiToEarn in the AI Content Ecosystem  

Open-source platforms like **[AiToEarn](https://aitoearn.ai/)** help creators:

- **Generate** AI‑driven content  
- **Publish simultaneously** on multiple platforms: Douyin, Kwai, WeChat, Bilibili, Rednote (Xiaohongshu), Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X (Twitter)  
- **Monetize** efficiently with analytics and model ranking  

**Docs & Code:** [AiToEarn开源地址](https://github.com/yikart/AiToEarn)  

![image](https://blog.aitoearn.ai/content/images/2025/11/img_002-45.png)

---

## MiniMax‑M2 Overview

[MiniMax‑M2](https://substackcdn.com/image/fetch/$s_!QvFc!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea71b548-6356-4929-951b-10715de6490d_2560x1060.png)

### Key Specs  

- **Open MIT‑licensed Mixture of Experts model** for coding and agentic workflows  
- **Model weights** on Hugging Face  
- **229B total parameters** but **~10B active per token** => low memory, reduced tail latency  
- **Optimized for tools**: shell, browser, retrieval systems, code editors  
- **Interleaved thinking** via `...` reasoning blocks — must remain intact for optimal performance.

---

### Creator Applications  
Integrate **MiniMax‑M2** into AI content automation pipelines via platforms like **AiToEarn** to combine intelligent planning/execution with global, multi-platform distribution.

Resources:  
- [AiToEarn官网](https://aitoearn.ai/)  
- [AiToEarn博客](https://blog.aitoearn.ai)  
- [开源地址](https://github.com/yikart/AiToEarn)

---

## Kimi K2 Thinking Model Release

![image](https://blog.aitoearn.ai/content/images/2025/11/img_003-3.jpeg)

### Specs  

- **Mixture-of-Experts architecture**  
- **1 trillion parameters** / **32B active per inference**  
- **256k context window**  
- **INT4 quantization** for efficiency  
- **Training cost:** ~$4.6M  
- **Agent capabilities**: 200–300 autonomous tool calls.

---

Extends DeepSeek’s architecture, increasing expert count and vocabulary size while reducing inference cost. Incorporates **FlashAttention** and **MuonClip** innovations.

---

## Udio–UMG Licensing Deal

[![image](https://blog.aitoearn.ai/content/images/2025/11/img_004-5.jpeg)](https://substackcdn.com/image/fetch/$s_!_BE1!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F15f1195e-aed7-468c-9227-f44a19aa00d1_2560x979.jpeg)

- **Settlement** ends claims of mass training on copyrighted catalogs  
- **Licensed platform launch** in 2026 with walled‑garden distribution  
- Current generator (incl. “BBL Drizzy”) remains but with restricted distribution  
- Temporary 48‑hour download window after user backlash.

AiToEarn parallels: open-source tools for monetizing creative output within controlled multi-platform ecosystems.

---

## Microsoft’s $15.2B UAE AI Infrastructure Investment

[![image](https://blog.aitoearn.ai/content/images/2025/11/img_005-4.jpeg)](https://substackcdn.com/image/fetch/$s_!Gz1T!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb10aa0db-4e57-4324-92df-8a947e2ce6f6_960x640.jpeg)

### Deal Highlights  

- **60,000 Nvidia GPUs** incl. A100/H100/H200/GB300  
- Equity stake in **G42**  
- $4.6B data center CAPEX (2025) + $7.9B (2026-29)  
- UAE as test case for **US AI diplomacy** with strict export controls.

---

## Other Industry Highlights

### Google TPU v7 Launch  
High FP8 performance, 3D torus architecture, optical switch fabric.

### Moonshot Kimi K2 Thinking v2  
Low training cost & high autonomous tool use.

---

## Multi‑Agent AI Developments

- **GitHub AI Agent HQ** — dashboard to run/manage multiple coding agents  
- **Cursor 2.0 Composer Model** — proprietary coding AI + parallel agents  
- **Microsoft MAI‑Image‑1** — photorealistic/artistic AI image generation  
- **Google AI Mode** — autonomous booking capability.

---

### Canva Proprietary Design Model  
Generative multi-layer design files, spreadsheet integration, analytics tools.

---

### Sora Android Launch  
~470,000 day-one installs, removing invite requirement.

---

### Instacart White-Label AI Shopping Bot  
Tailored recommendations for enterprise retail clients.

---

## AI Business Landscape

- **Nvidia** valuation tops $5T  
- **Coca‑Cola AI ad** widely criticized  
- **Amazon Project Rainier** for Anthropic Claude  
- **Lambda–Microsoft multi‑billion AI infra deal**  
- **Apple–Google Gemini Siri upgrade talks** (~$1B/year)  
- **Google–Reliance partnership** for Gemini Pro access  
- **Baidu’s Apollo Go robotaxi milestone**  
- **Waymo expansion** to 3 more cities  
- **Pony AI** $863M Hong Kong IPO  
- **Shopify AI-driven e-commerce growth**  
- **People Inc.** AI licensing with Microsoft.

---

## AI Research Highlights

- **Scaling Latent Reasoning via Looped LMs** — 2–3× parameter efficiency  
- **Future Summaries Pretraining** — +5% long-range reasoning gains  
- **Supervised Reinforcement Learning** — dense rewards for stepwise reasoning  
- **End-to-End LM Decoding** — dynamic sampling parameters  
- **Continuous Autoregressive LMs** — next-vector prediction  
- **FP16 Training–Inference Match** — stability + performance improvements  
- **Kimi Linear Attention** — efficient + expressive  
- **Remote Labor Index** — AI automates ~2.5% of remote work tasks  
- **arXiv policy changes** — block low-effort AI-generated submissions.

---

## Policy & Analysis

- **OpenAI copyright case** — potential billion-dollar liabilities  
- **Stability AI win vs. Getty Images** — minimal trademark infringement  
- **AI artist Xania Monet** — Billboard debut sparks ethics debate  
- **Jerome Powell** warns of near-zero net job creation due to AI.

---

## Productivity Trends & Profitability  

Media Lab study: AI gains remain **sector-specific** and **niche**.  
Constraints: legacy systems, governance, compliance, re-skilling needs.  
Recommendation: domain-specific AI, end-to-end processes, iterative culture.

---

**AiToEarn** exemplifies unified content generation, publishing, and monetization — bridging technical capability with economic opportunity.