May 1, 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 Musk v. Altman trial is underway in Oakland before Judge Yvonne Gonzalez Rogers, with a 9-person advisory jury and a ruling expected by mid-May. Musk testified he was "a fool" for donating ~$38M, is seeking up to $134B in "wrongful gains" plus the removal of Altman and Brockman — and on cross-examination admitted xAI "partly" used distillation on OpenAI models to train Grok. Of his 26 original claims, only two survived: breach of charitable trust and unjust enrichment.

Why it matters: Even narrowed to a charity-law dispute, a Musk win could force a structural unwinding of OpenAI's October 2025 recapitalization, blow up Microsoft's ~$135B (~27%) stake, and create unprecedented case law on whether private donors can claw back nonprofit transformations. "AI bubble could pop within days if Musk wins" is already a 721-upvote thread on r/BetterOffline.

Q1 2026 revenue $109.9B (+22% YoY, fastest since 2022), net income $62.58B (+81%), EPS $5.11. Google Cloud crossed $20B for the first time at +63% YoY, with operating income tripling to $6.6B. Cloud backlog nearly doubled sequentially to $462B; Sundar said cloud "would have been higher if we were not capacity constrained." Capex guide raised to $180-190B; stock rallied ~7%.

Why it matters: This is the print that transformed hyperscaler AI capex from "speculative spend" to "contracted revenue" in Wall Street's eyes. The $462B backlog covers ~$230B of next-two-year cloud revenue. Meta got punished -6% the same day for similar capex math; the difference is Google has demand it can't even fill.

Microsoft raised 2026 capex to ~$190B (vs prior $152-154B consensus), with CFO Amy Hood disclosing ~$25B of that increase is purely higher memory and AI-chip prices. Combined hyperscaler 2026 capex (MSFT + GOOGL + AMZN + META) tracks ~$725B, +77% YoY. MSFT expects to remain capacity-constrained even at the higher spend. Azure +40%, AI services run-rate $37B (+~123% YoY).

Why it matters: The bottleneck has officially moved from capital to memory — HBM scarcity, not GPU availability, now sets hyperscaler capex. Bernstein's Mark Moerdler captured the discomfort: "There's a disconnect that makes investors nervous between how fast they're seeing CapEx growing and how fast they're seeing revenue growing."

AWS revenue hit $37.6B in Q1 2026, +28% YoY — the fastest pace in 15 quarters on a $151B base. Custom silicon (Graviton + Trainium + Nitro) crossed a $20B annualized run rate, growing triple-digits YoY and ~40% QoQ. Jassy said it would be a $50B/yr standalone revenue line. AWS AI revenue is at $15B run rate (~260x where AWS itself was at the same stage of cloud build-out). Q1 capex $44.2B; full-year 2026 ~$200B. The catch: TTM FCF collapsed ~95% to ~$1.2B.

Why it matters: AWS is the only hyperscaler anchoring both leading frontier labs (Anthropic, OpenAI) into its silicon — Anthropic alone has signed up to 5 GW of Trainium and a $100B+/10-yr commitment. Trainium2 is reportedly ~30% better price/perf than comparable GPUs. Jassy is openly hinting at selling racks to third parties.

Anthropic is reportedly fielding unsolicited offers of $40-50B at an $850-900B valuation, surpassing OpenAI's $852B mark and making it the world's most valuable AI startup. That's more than double February's $380B Series G post-money — and a 15x markup from March 2025's $61.5B Series E. ARR went from ~$9B at end-2025 to $30B+ in March, reportedly approaching $40B now (Claude Code and Cowork carrying the number). Could be the last private raise before an IPO as soon as October 2026.

Why it matters: TechFundingNews put it bluntly: "No company in American technology history has grown at that rate." Whether you read this as the cleanest "AI bubble" data point in the set or the most aggressive enterprise revenue ramp ever, you can't ignore it. Secondary-market trades have already implied a ~$1T valuation.

The Blend

Connecting the dots across sources

Big Tech's $725B capex isn't speculative anymore — it's pre-sold, and it still isn't enough

  • Across the news today Alphabet posted a $462B cloud backlog and Sundar Pichai said cloud revenue "would have been higher if we were not capacity constrained" — meaning much of the capex is already attached to contracted demand.
  • Microsoft's CFO disclosed in earnings that ~$25B of its 2026 capex hike is purely higher memory and AI-chip prices, not more units, telling us HBM scarcity (not GPUs) is now setting the spend.
  • On X the most-shared chart of the day was the $725B aggregate hyperscaler capex (+77% YoY), and four hyperscalers all said they remain capacity-constrained through 2026 even at the higher spend.
  • In the blogs, Pragmatic Engineer documented 15 tech companies blowing through their agent-token budgets in 2-3 months, suggesting the demand pulling on this capex is structural, not seasonal.

The agent economy is the workload that's actually breaking capacity

  • In the news, Stripe shipped Link as a wallet for autonomous agents and Cloudflare announced agents can now create accounts, buy domains, and deploy without a human in the loop — agents went from API consumers to customers.
  • On GitHub today, four of the top five trending repos are agent infrastructure: warpdotdev/warp, mattpocock/skills, obra/superpowers (174K stars), and TauricResearch/TradingAgents.
  • On X, Vercel's note that removing 80% of tools from their agents made them 3x faster with 37% fewer tokens turned "context and harness engineering" into the engineering meme of the week.
  • In the research, HuggingFace's GLM-5V-Turbo paper argues vision-first foundation models — not adapter retrofits — are what unlock reliable agent behavior, putting an academic floor under the same trend.

Reasoning benchmarks keep rising while users say models feel dumber

  • In the news GPT-5.5 became the second model to complete the UK AISI's 32-step end-to-end corporate-network attack simulation, and Baidu's ERNIE 5.1 Preview took the #1 Chinese spot on LMArena at 1,476 — above DeepSeek V4-Pro and GPT-5.5.
  • In the research, Apple's "The Illusion of Thinking" paper (top X-trending with 6,521 votes) found reasoning models actually reduce their thinking effort as problems get harder, even with budget to spare.
  • On X, 22,000 people signed a petition asking OpenAI to bring back GPT-4o because "AI is getting dumber" — an unusually loud user-side counter-signal to the benchmark wins.

Slow Drip

Blog reads worth savoring

Analysis · The Pragmatic EngineerThe Pulse: token spend breaks budgets – what next?

Inside data from 15 tech companies on how exploding AI agent spend is breaking budgets — and what they're doing about it. The perfect chaser to today's $725B capex stories.

Analysis · Latent Space[AINews] The Inference Inflection

Latent Space steps back and maps why the "inference age" rewrites economics and product design.

Tutorial · Amazon EngineeringOrganizing Agents' Memory at Scale: Namespace Design Patterns in AgentCore Memory

A concrete AWS playbook for namespace hierarchies, retrieval patterns, and IAM-based access control so your agents' memory doesn't collapse under production load.

Tutorial · Data Science CollectiveI Built an AI Agent That Knows When to Quit. Here's How.

A LangGraph walkthrough for building support agents with stop conditions and structured human-handoff packets — the missing piece in most agent demos.

News · The Neuron AIBig Tech just spent $130B on AI in Q1; Google literally ran out of capacity

One day, four hyperscaler earnings calls, the same shocking message — even $130B isn't enough.

News · Cloudflare BlogAgents can now create Cloudflare accounts, buy domains, and deploy

Cloudflare just made AI agents first-class customers — they can sign up, pay, register domains, and deploy without a human ever touching the dashboard.

Research · Google DeepMindEnabling a new model for healthcare with AI co-clinician

DeepMind sketches a credible path to AI-augmented care from the lab that knows medical AI best.

The Grind

Research papers, decoded

Reasoning6,521 upvotes · arxiv
The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity

Apple researchers stress-tested "thinking" LLMs (DeepSeek-R1, Claude Sonnet Thinking, etc.) on controlled puzzles like Tower of Hanoi and River Crossing where difficulty can be dialed precisely, sidestepping benchmark contamination. They found three regimes: at low complexity standard models can beat reasoning models, at medium complexity reasoning models genuinely help, and at high complexity both collapse equally — and counterintuitively, reasoning models reduce their thinking effort as problems get harder, even with budget to spare. The takeaway: today's "reasoning" models are sophisticated pattern matchers with hard scaling cliffs.

Theory200 upvotes · alphaxiv
There Will Be a Scientific Theory of Deep Learning

Argues deep learning is graduating from trial-and-error engineering into a real science of "learning mechanics," and synthesizes five mathematical pillars: solvable toy models (deep linear networks), infinite-width "lazy vs. rich" regime analysis, empirical scaling laws, hyperparameter transfer theory, and universal representations across architectures. For practitioners, a mature theory promises to replace expensive hyperparameter sweeps with predictable scaling rules.

Multimodal Agents71 upvotes · huggingface
GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents

Built so vision is a first-class citizen in the model's reasoning loop instead of a vision encoder bolted onto an LLM after the fact. Combines a CogViT encoder, Multi-Token Prediction with a learnable image placeholder, and reinforcement learning across 30+ task categories. Reports an 8x jump on multimodal search and beats competitors on UI-to-code without hurting pure text coding. The argument: native multimodality, not adapter retrofits, is what unlocks reliable agent behavior.

On Tap

What's trending in the builder community

warpdotdev/warp

Warp is an agentic development environment, born out of the terminal. 8,262 stars today, 48,327 total (Rust). The agentic terminal is having a moment.

mattpocock/skills

"Skills for Real Engineers. Straight from my .claude directory." 6,175 stars today, 48,343 total (Shell). Riding the explosive Claude Skills wave.

TauricResearch/TradingAgents

Multi-agent LLM financial trading framework. 2,203 stars today, 56,950 total (Python).

obra/superpowers

An agentic skills framework and software-development methodology that actually works. 1,623 stars today, 174,286 total (Shell).

iamgio/quarkdown

Markdown with superpowers — papers, presentations, websites, books, knowledge bases. 799 stars today, 12,867 total (Kotlin).

Plurai

Vibe-train evals and guardrails tailored to your use case. Plugs straight into the harness-engineering moment.

Open Wearables

Open infrastructure for wearable-powered health products.

KarmaBox

Run your own Claude Code in your pocket. Yes, really.

UXPin Forge

Generate UI from your design system, not around it.

Gro v2

Spot signals, trigger outreach — turn posts into pipeline.

How GPT-5, Claude, and Gemini are actually trained and served – Reiner Pope

Dwarkesh Patel goes deep on training and serving infra for frontier models. The infra deep-dive of the week.

The AI Frontier and How to Spot Billion-Dollar Companies Before Everyone Else — Elad Gil

Tim Ferriss interviews Elad Gil on how to spot generational AI companies early.

The Only 20 Ways to Make Money with AI in 2026

Dan Martell's pragmatic 2026 AI monetization tour.

Making $$ with AI Agents

Greg Isenberg with Airtable's Howie Liu on HyperAgent.

find-skills

Vercel Labs' meta-skill for discovering skills.

vercel-react-best-practices

Opinionated React best practices, packaged as a skill.

frontend-design

Anthropic's recipe for distinctive, production-grade frontend interfaces that reject generic AI aesthetics.

Skill Vetter (ClawHub)

Security-first skill vetting from ClawHub.

Roast Calendar

Upcoming events & gatherings

AI & Scientific Researcher Dinner: Autopoiesis Sciences x RhoThu, Apr 30 at 6:30 PM PT | San Francisco, CA
Weighted Moves: AI & Chess MixerThu, Apr 30 at 7:00 PM PT | San Francisco, CA
GTM Engineering DinnerThu, Apr 30 at 6:30 PM PT | San Francisco, CA
AI DevDay After Hours (TensorWave)Thu, Apr 30 at 7:00 PM PT | San Francisco, CA
AIC & ISN Guest Speaker: Lead Data Engineer at Crusoe & NetflixThu, Apr 30 at 6:30 PM PT | Santa Clara, CA
Shopify & NetSuite User Group: The Real-Time Commerce StackThu, Apr 30 at 6:30 PM PT | San Francisco, CA
Founder Photo Day @ ReductoThu, Apr 30 at 6:30 PM PT | San Francisco, CA

Last Sip

Parting thoughts & a teaser for tomorrow

If today had a thesis, it's this: the AI economy is now real enough to print a $462B backlog and weird enough to have $25B of capex inflation come from memory chips alone — while a courtroom in Oakland decides whether Microsoft's $135B stake in OpenAI even gets to exist in its current form. We're watching the agent economy quietly assemble its plumbing (payment rails at Stripe, account creation at Cloudflare, memory namespaces at AWS), and the GitHub trending list is basically all agent skills now. Tomorrow we'll be watching the Musk v. Altman testimony continue, the Anthropic board's response to those $900B offers, and whether anyone outside Big Tech can afford an HBM module before Christmas. Stay caffeinated.