18: Compute is the new oil

Jul 25, 2025

Hey, it’s Marc.

The White House dropped a 90-step AI roadmap focused on deregulation, international exports, and wiping “ideological bias” from U.S.-built AI models.

What’s in it: [AI Action Plan]

Export U.S. AI tech: Full-stack packages (chips to software) for allies
Fast-track data centers: Easier permits + workforce boost
Kill red tape: Remove federal rules that slow down AI development
Bias crackdown: Government contracts only for “ideology-free” models

The takeaway: Trump wants to lock in U.S. leadership, flood allies with American AI, and take a hard stance against what the administration calls “woke” systems.

Top Agent signals this week:

Amazon acquires AI wearable startup, Bee
Qwen3-Coder launches as an open model for agentic coding tasks
GitHub Spark lets Copilot Pro+ users build full-stack apps from text prompts
Proton launches Lumo AI chatbot
Gemini 2.5 Flash-Lite is now stable—Google’s fastest, cheapest AI model yet.
Replit launches Queue for efficiency

👉 Get your brand in front of 30,000+ decision-makers — book your ad spot now.

Power Moves

xAI to raise $12B in debt for 1M GPU supercomputer

Elon Musk’s xAI is reportedly raising up to $12B in debt to fund Colossus — a Memphis-based AI supercluster targeting 1M Nvidia GPUs. That’s up from the 200K already deployed. xAI has already raised $10B (50/50 debt + equity) in July. [NEWS]

So what? It’s an industrial-grade buildout of vertically integrated AI infrastructure, where X supplies data, Tesla provides energy, SpaceX handles distribution, and Grok runs inference. Think AWS + OpenAI + Twitter in one stack.

OpenAI’s $500B compute bet hits the wall

OpenAI’s $500B AI infrastructure plan with SoftBank has stalled, with just one small data center expected this year. While SoftBank and OpenAI pledged $100B “immediately,” internal clashes and unclear execution have sidelined Stargate’s moonshot launch. [NEWS]

By the numbers: Global data center capacity is growing 23% per year through 2030

OpenAI projects 4M Nvidia-class chips (2M A100 equivalents) by year-end
Oracle’s 4.5 GW data center deal = power of 2 Hoover Dams
$30B+ annual spend for compute with Oracle alone — triple OpenAI’s projected revenue

Why it matters: OpenAI’s growing compute needs are outpacing supply chains, partner patience, and grid capacity. Enterprise AI roadmaps will bottleneck on chips, data centers, and sovereign compute policy, not models.

Vendor Spotlight: Qwen3-Coder

Qwen launches Qwen3 Coder, a coding assistant that enables you to use developer tools, solve coding problems and plan out complex software projects step by step.

Features of Qwen3 Coder:

Reads and understands entire codebases, not just single files
Remembers context from long conversations (up to 256,000 words)
Can work with pull requests and understand how code changes affect the whole project
Knows how different parts of your application connect together

Why it matters: Qwen3-Coder is an open-source model with no subscription fees for the base model. It is designed for programming tasks and performs as well as expensive commercial alternatives. Plus, it scales with the developer’s needs.

Zooming in: It plugs into popular code editors and development environments and uses the existing APIs and tools. It is compatible with the existing workflows and works through command-line interfaces that developers are familiar with.

How Big and Powerful Is It?

This AI was trained on a massive amount of code and programming knowledge:

Built using 480 billion parameters (that's the "brain size")
Trained on 7.5 trillion pieces of text, mostly code
Can handle conversations as long as a short novel
Performance matches top commercial AI models like Claude

What Can You Actually Do With It?

For Everyday Coding

Write complete functions and classes based on your description
Debug existing code by explaining what's wrong and how to fix it
Add new features to existing applications
Convert code from one programming language to another
Generate documentation and comments for your code

For Bigger Projects

Plan entire applications from scratch
Design databases and how they connect to your app
Create web applications with both frontend and backend
Build APIs that other applications can use
Set up automated testing for your code

Specific Examples

Web Development: Create complete online stores, social media apps, or business websites
Data Analysis: Build tools to analyze spreadsheets, create charts, and find patterns
Games and Simulations: Code simple games or physics simulations
Automation Scripts: Create tools that handle repetitive computer tasks
Mobile Apps: Help design and code smartphone applications

Three Ways to Get Started

1. Command Line Tool (Qwen Code)

Install with: npm i -g @qwen-code/qwen-code
Type qwen in your terminal to start coding
Works like having a coding buddy in your command line

2. Through Claude Code

Install with: npm install -g @anthropic-ai/claude-code
Use familiar Claude interface but powered by Qwen3-Coder
Good if you're already used to Claude's style

3. Direct API Access

Connect through Alibaba Cloud
Integrate into your own applications or tools
Use the same way you'd use OpenAI's API

Qwen3 coder is free to try, works with tools you already know, and can handle everything from simple scripts to complex applications. In the world of vibe coding, it is an attractive option as an open-source model.

Top AI Vendors This Week

👉 We’re tracking 1000s of AI and blockchain vendors so you find the right vendors, partners, and bet on the right tech.

More from 51:

That’s all for today.

Thanks,

Marc & Team

Discussion about this post

Ready for more?