👾 Why AI Will Bankrupt The Power Grid (Leaked Report)

The US alone needs 50 GW more power for AI leadership. That's 30+ nuclear reactors. Powerful open-source AI at low cost changes everything and how to vibe code.

Author

Tobias Wupperfeld
August 05, 2025

Welcome to The Agent Roundup

The numbers are wild, but no longer theoretical. When a single AI training run can consume as much power as an entire city, we're facing the largest infrastructure challenge in modern history.

The U.S. provides an example of what's coming globally. Anthropic's Build AI in America report describes the reality: by 2028, America alone needs at least 50 GW of additional electrical capacity just to maintain AI leadership. That's nearly double New York State's peak demand, equivalent to 30+ nuclear reactors, or more than half the entire U.S. nuclear fleet.

This splits into two critical categories:

Frontier Training: The workloads to train new foundation models. By 2028, training a single state-of-the-art model could demand a 5 GW datacenter.

AI Inference: The distributed network serving billions of real-time responses across the globe.

Here's the challenge: existing datacenter infrastructure works fine for inference (close to users, leveraging urban fiber) but breaks down for frontier training's power demands.

Goldman Sachs' Powering the AI Era report came out:

  • $1T in private hyperscaler investment by 2027

  • 160% growth in datacenter power consumption this decade

  • $5T total infrastructure requirement by 2030

Capital is flowing through entirely new channels: private credit vehicles, long-term power purchase agreements, and vertically integrated infrastructure.

Three brutal realities emerge:

  • Datacenters can be built quickly but bureaucracy used to be slow

  • GPUs lose value every 18 months, but infrastructure loans expect decades

  • What happens to non-AI sectors when trillions flood into compute infrastructure?

Bottom line: Nations worldwide face the same fundamental challenge. The U.S. example shows the scale, but every major economy will grapple with similar infrastructure demands as AI becomes essential.

This week’s topics:

  • Open-source power at low cost

  • Anthropic research on auditing agents

  • How to vibe code in production responsibly

  • More market updates, trending AI tools, and more

AI Agent News Roundup

💥 Breakthroughs

Open-Source Power at Low Cost with GLM-4.5

z.ai glm 4.5 model family on benchmarks

Source: Z.ai

Z.ai introduces GLM-4.5 and GLM-4.5-Air frontier models, marking a significant effort toward unifying reasoning, coding, and agentic capabilities into a single model.

This addresses a common limitation where previous LLMs often excelled in one specific domain but not across the board.

Achieves strong overall performance across 12 diverse benchmarks, securing a top-3 ranking for GLM-4.5 among leading models from various major AI labs.

High Performance at Low Cost with Step3

step3 model cost comparison

Source: StepFun

StepFun's Step3 model is presented as a cutting-edge multimodal reasoning model built on a Mixture-of-Experts (MoE) architecture, with a primary design goal of minimizing decoding costs while maintaining top-tier vision-language reasoning performance.

It achieves cost-effective multimodal intelligence and high decoding efficiency.

The model processes a vast amount of pre-training data, including 4 trillion image-text mixed tokens, emphasizing its multimodal capabilities.

Automated Agents Audit Models for Alignment Issues

anthropic text with magnifying glasses

Agent Roundup via Reve

Anthropic introduces LLM-based auditing agents.

Main Innovations:

Development of three specialized agents: a tool-using investigator agent for open-ended investigations, an evaluation agent for building behavioral evaluations, and an agent for eliciting concerning behaviors.

The research findings indicate that interpretability tools and semantic search are critical for success in complex auditing.

📈 More Market Updates

🇺🇸 Tesla-Samsung: $16.5B manufacturing deal for Tesla's next-gen AI6 chips signals major semiconductor partnership expansion. Musk emphasizes the strategic importance as Tesla deepens its AI hardware capabilities.

🇺🇸 DOGE: New "DOGE AI Deregulation Decision Tool" targets elimination of 100,000 federal regulations from 200,000 total by Trump's first anniversary. Automated AI analysis will drive most removal decisions with minimal staff oversight.

🇺🇸 Meta: Q2 revenue climbed 22% to $47.5B with net income reaching $18.3B as advertising strength continues.

🇺🇸 Microsoft: Revenue hit $76.4B with next year's capital spending exceeding $100B for AI datacenter expansion. The 14% budget increase positions Microsoft aggressively in the infrastructure arms race against Alphabet ($85B) and Amazon ($100B).

🇺🇸 Amazon: Q2 operating income reached $19.2B on $167.7B sales, crediting AI deployment across retail, cloud, and logistics. AWS revenue jumped 17.5% to $30.9B while North America retail climbed 11% to $100.1B.

🇺🇸 Apple: Fiscal revenue of $94.04B beats estimates, marking strongest growth since 2021 with 10% year-over-year increase. iPhone sales rose 13% while services hit June-quarter records.

🇺🇸 Figma: $1.2B IPO saw shares rocket 250% on first day, achieving $65B+ valuation and marking biggest US debut pop in three decades.

🇺🇸 Reddit: Q2 advertising revenue hit $465M, comprising 93% of total revenue as AI-powered ad tools drive growth.

🇳🇴 OpenAI: Stargate Norway becomes first European AI datacenter with 100,000 Nvidia GPUs, targeting full renewable energy by late 2026.

Together with Syllably.io

Create Faceless Videos in Minutes

Generate scripts, captions, and faceless short-form videos using AI—all from one dashboard.

Thousands of creators are using Syllaby.io to grow across TikTok, YouTube, and Instagram.

How to Vibe Code in Production

Key insights of the Code w/ Claude event from Anthropic on the best practices of vibe coding.

What Is Vibe Coding?

It's defined as "where you fully give into the vibes, embrace exponentials, and forget that the code even exists". The key is forgetting the code itself, while still understanding the product.

Why Should You Care?

The length of tasks AI can perform is currently doubling every seven months.

length of tasks Ai can do by model

Source: metr.org

While engineers can still review hour-long tasks today, soon AI will generate days or even weeks of work. It will become impossible to keep up by reviewing every line of code manually.

To leverage this exponential growth, engineers must learn to responsibly trust AI-generated code, similar to how developers eventually trusted compilers without reading the assembly output. Otherwise, engineers risk becoming a bottleneck and missing out on the capabilities of future models.

How to Vibe Code in Production Responsibly?

Be the AI's Product Manager (PM): Instead of asking "what AI can do for you," ask "what you can do for AI".

  • Treat the AI like a new junior engineer: provide comprehensive guidance, context, requirements, specifications, and constraints.

  • This often involves spending 15-20 minutes collecting information and building a plan with AI, exploring the codebase, and identifying files and patterns.

  • You need to be able to ask the right questions

Focus on Leaf Nodes: Apply vibe coding primarily to "leaf nodes" in your codebase – parts of the system that nothing else depends on.

  • It's acceptable for these parts to have some technical debt because they are less likely to change or have other features built upon them.

  • The core architecture (branches and trunks) still requires deep human understanding and protection to ensure extensibility and flexibility.

Think About Verifiability: Design your systems and changes so you can verify correctness and stability without needing to read every line of code.

  • Strategies include writing acceptance tests, using the product to ensure expected behavior, spot-checking key facts, and designing systems with easily human-verifiable inputs and outputs

Security Considerations: For production systems, the engineer needs to know enough about the context to understand what is dangerous or safe.

Test-Driven Development (TDD): TDD is very useful in vibe coding.

  • Encourage the AI to write minimalist, general, end-to-end tests (e.g., happy path, error cases)

Tool Spotlight

Microsoft integrates AI assistant directly into Edge browser, enabling cross-tab search functionality, automated task execution, and proactive action recommendations for enhanced browsing productivity.

🎬 Wan2.2

Alibaba’s open-source video generation model delivers cinematic-quality output with advanced motion dynamics, supporting both text-to-video and image-to-video synthesis capabilities.

Browser-based AI agent automates complex spreadsheet workflows, including financial model construction and SEC data integration, featuring chatbot interface and seamless Excel export functionality.

Black Forest Labs and Krea collaborate on open-weight image generation model designed to eliminate artificial aesthetics, delivering enhanced photorealism and superior visual quality.

More Resources

Blog: In-depth articles on AI workflows and practical strategies for growth
AI Tool Collection: Discover and compare validated AI solutions
Consultancy: Explore AI potential or make your team AI-fit
Agency: Production-ready AI implementation services

See you next time!

Tobias