Mar 29, 2026

Agentic Brew Daily

Your daily shot of what's brewing in AI

Fresh Batch

Bold Shots

Today's biggest AI stories, no chaser

The $700 Billion Concrete Pour: AI Data Centers Are Reshaping America's Grid

Tech giants are deploying $630-700B in AI data center investment in 2026 alone — a 62% jump from last year. Amazon leads at $200B, Google at $175-185B, and Meta at $115-135B. Meta's building the 2,250-acre Hyperion data center in Louisiana for up to $27B and financing 10 gas-fired power plants to deliver 7.5GW — a 30% increase to Louisiana's entire grid capacity. Sanders and AOC have introduced legislation to pause new construction.

Why it matters: These aren't tech investments anymore — they're energy infrastructure decisions that will shape power grids for decades. When a single company can increase a state's grid capacity by 30%, we've entered a fundamentally different era.

Judge Rita Lin ruled the Pentagon's 'supply chain risk' designation of Anthropic was illegal First Amendment retaliation. The dispute started when Anthropic refused unrestricted Claude access for autonomous weapons and mass surveillance. Pentagon CTO Emil Michael says the ban 'still stands' despite the court order. 37 engineers from Google DeepMind and OpenAI filed a joint amicus brief in support.

Why it matters: Potentially landmark precedent that companies can't be punished for ethical positions on their technology. The Pentagon's open defiance of a federal court order is setting up a constitutional showdown.

OpenAI killed Sora after burning $15M/day in inference costs against $2.1M in total lifetime revenue. Disney exited a $1B investment and three-year licensing deal covering 200+ characters — no money ever changed hands. User engagement had collapsed 66% from peak. OpenAI retains Sora as internal research for world models and robotics.

Why it matters: If the best-funded AI company in history can't make consumer video generation work economically, it raises fundamental questions about the gap between AI capability and profitable delivery.

Claude Code's revenue hit $500M+ annual run-rate after 10x usage growth in three months. It authors 4% of public GitHub commits, projected 20%+ by year-end. Anthropic shipped auto mode letting it autonomously decide safe actions. 90% of Claude Code's own codebase was written by Claude Code.

Why it matters: 4% of all GitHub commits from a single tool is infrastructure-level penetration. This is a structural transformation of how software gets written.

The Blend

Connecting the dots across sources

AI Agents Crossed From Demo to Production Infrastructure This Week

  • Claude Code 4% of GitHub commits + Cursor 1M agent commits in two weeks + Coinbase 50% AI code
  • GitHub: superpowers repo at 2,293 stars/day; Product Hunt: 4 of top 6 products are agent tools
  • LangChain published agent evaluation readiness checklist — you don't write production checklists for toys
  • SakanaAI AI-Scientist-v2 (507 stars/day) automating scientific discovery

The AI Economics Are Sending Contradictory Signals

  • Sora burned $15M/day for $2.1M total revenue while the industry deploys $630-700B in data center capex
  • YouTube bearish (Ed Zitron 89K views, Wendover 1.37M views) vs X bullish (SoftBank $40B loan, funding excitement)
  • Latent Space reports H100 prices going UP, not down, contradicting 'compute getting cheaper' narrative
  • Efficiency research papers trending (Attention Residuals matches 1.25x compute baselines) suggesting field seeking alternatives to brute scaling

Slow Drip

Blog reads worth savoring

Analysis · EveryThe Agent That Saved My Brain

Every's growth lead built a personal command center inside Claude Code and the workflow is genuinely clever.

Analysis · Towards AII Burned Thousands of Hours on the Cursor. Here Is Why 'Auto-Accept' Is a Trap.

A cautionary tale about speed-running AI completions and the hidden maintenance debt.

Tutorial · EveryBuild Your Own Bloomberg Terminal With AI

Step-by-step from ChatGPT earnings previews to a custom investment dashboard.

Tutorial · LangChain BlogAgent Evaluation Readiness Checklist

If you're shipping agents to production, this is the quality bar you should be hitting.

Industry Signal · Cursor BlogNVIDIA commits 3x more code across 30,000 developers with Cursor

When NVIDIA goes all-in on AI-assisted coding across 30K engineers, that's a signal.

Industry Signal · Latent Space[AINews] H100 prices are melting UP

Counterintuitive GPU pricing trends that challenge the 'compute getting cheaper' narrative.

The Grind

Research papers, decoded

Neuroscience / BCI3,613 upvotes · arxiv
Stable, chronic in-vivo recordings from a fully wireless subdural-contained 65,536-electrode brain-computer interface device

A fully wireless BCI with 65,536 electrodes — roughly 60x more than current commercial devices — that records stable signals without penetrating brain tissue.

Machine Learning / Efficiency527 upvotes · alphaxiv
Attention Residuals

Content-dependent attention fixing PreNorm dilution. +7.5 on GPQA-Diamond, +3.6 math, +3.1 code. Matches baselines requiring 1.25x more compute.

Computer Vision / Tools9 upvotes · huggingface
VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models

Converts complex raster figures into clean, editable SVG vector graphics using VLMs. Practical for scientific figures and design assets.

On Tap

What's trending in the builder community

obra/superpowers

Agentic skills framework flying at 2,293 stars/day.

datalab-to/chandra

OCR for complex tables and forms at 679 stars/day.

virattt/dexter

Autonomous financial research agent at 583 stars/day.

Agentation

Visual feedback for AI agents. 407 votes.

Claude Code auto-fix

Auto-resolves CI failures. 345 votes.

Anthropic Claude Mythos Leak

CMS misconfiguration revealed 'Claude Mythos' under new 'Capybara' tier. 4.5M views.

Boris Cherny: 100% AI-Written Code

No manual edits since November, shipping 10-30 PRs daily. Garry Tan repost hit 293K views.

The AI boom is a lie: Fake data centres and unused GPUs

Ed Zitron's contrarian take on data center utilization. 89K views.

Anthropic's Generational Run, OpenAI Panics, AI Moats, Meta Loses Major Lawsuits

All-In Podcast covers Anthropic's multi-front dominance. 146K views.

Roast Calendar

Upcoming events & gatherings

AgentKit Hackathon Demo DayToday, March 29 @ 9:00 AM PT | San Francisco
Anthropic Equity Structure vs OpenAI & xAIToday, March 29 @ 3:00 PM PT | Sunnyvale
Design Engineer x Wonder - Creative HoursTonight, March 29 @ 6:00 PM PT | San Francisco
HackwithBay 2.0Sunday, March 30 @ 9:00 AM PT | San Francisco
AI Operator Intensive: 5-Day WorkshopSunday, March 30 @ 9:00 AM PT | San Francisco
Breakout (Bittensor)Sunday, March 30 @ 9:00 AM PT | San Francisco
STAK Ventures Speaker SeriesSunday, March 30 @ 12:00 PM PT | Oakland

Last Sip

Parting thoughts & a teaser for tomorrow

Here's what's sticking with me today: the safety-focused AI company got blacklisted by the Pentagon for being too safe, and then leaked its most dangerous model through a basic CMS error. You genuinely cannot write this stuff.

But zooming out, the real story this week is the tension between ambition and economics. $700B flowing into data centers while Sora proved you can burn $15M/day and still fail commercially. Claude Code printing $500M ARR while researchers publish papers on efficiency because compute isn't getting cheaper. The industry is simultaneously the biggest investment story in history and a cautionary tale about the gap between capability and sustainability.

Tomorrow we'll be tracking the Pentagon's response to the injunction (will they actually defy a federal court?), any fallout from the Mythos leak, and whether the Sanders-AOC moratorium bill gains traction over the weekend. Plus, HackwithBay 2.0 kicks off — always fun to see what people build under pressure.

Stay caffeinated. See you tomorrow.