Agentic Brew Daily
Your daily shot of what's brewing in AI
Fresh Batch
Bold Shots
Today's biggest AI stories, no chaser
On April 28, AWS and OpenAI announced GPT-5.4 (with 5.5 weeks away), Codex, and Bedrock Managed Agents going live on Amazon Bedrock — exactly one day after Microsoft and OpenAI restructured the Azure exclusivity deal into a non-exclusive IP license through 2032. Amazon also tossed in a $50B investment ($15B upfront, $35B contingent on IPO/AGI) and OpenAI committed tens of billions to AWS Trainium, including a 2-gigawatt Trainium 3/4 build. Codex is already at 4M weekly users and AWS claims 30-40% cost savings vs. Nvidia.
Why it matters: The 'one model, one cloud' era of frontier AI is officially over. Microsoft trades exclusivity for a 27% stake worth ~$135B, and AWS converts Trainium from a price-sensitive alternative into a credible primary platform overnight.
Anthropic is in advanced talks for a ~$50B primary at >$900B post-money (decision at the May board meeting), while on April 28 its secondary shares on Forge Global traded at an implied $1T — eclipsing OpenAI's ~$880B implied. Annualized revenue went $9B to $30B in a single quarter, Claude Code alone is at $2.5B+ ARR, and 8 of Fortune 10 are paying customers. Amazon committed up to $25B and Google up to $40B in the last ten days alone.
Why it matters: Anthropic just leapfrogged OpenAI on implied valuation while being structurally tethered to two hyperscalers who are simultaneously its largest shareholders, customers, and only compute pipelines. The ~$65B in pledged equity is effectively reserved gigawatts.
Google inked a classified agreement on April 28 letting the Pentagon use Gemini for 'any lawful government purpose,' including on classified networks, with Google obligated to adjust safety filters at government request. The deal closed the day after 600+ Google employees (including 20+ directors and DeepMind researchers) signed an open letter to Sundar opposing it — and right after the Pentagon branded Anthropic a 'supply-chain risk' for refusing to drop guardrails on mass surveillance and autonomous weapons.
Why it matters: 'Supply-chain risk' is now procurement-speak for 'won't drop your published guardrails.' Compare to 2018's Project Maven (4,000 signers killed the contract); this 2026 letter arrived a day late.
Jury selection in Musk's civil trial against Altman, Brockman, OpenAI, and Microsoft started April 27 in front of Judge Yvonne Gonzalez Rogers. Of 26 original claims, only two survived — unjust enrichment and breach of charitable trust — but Musk is asking for $130-150B and testified that Altman and Brockman 'looted' the charity. The nine-person jury is advisory; the judge issues a binding ruling by mid-May.
Why it matters: If any version of the charitable-trust theory sticks, every future AI nonprofit-to-PBC conversion gains a fresh litigation surface — and Microsoft is exposed on aiding-and-abetting at exactly the moment its Copilot business depends on the partnership.
Alphabet posted Q1 2026 revenue of $109.9B (+22% YoY), Google Cloud at $20.03B (+63% YoY), backlog nearly doubling QoQ to $462B, and capex guidance up to $180-190B (with 2027 going 'significantly' higher). Microsoft hit $82.9B (+18%), Azure +40%, AI ARR at $37B (more than doubled YoY). Combined hyperscaler 2026 AI capex got re-rated from ~$670B to ~$725B in about 80 seconds of earnings.
Why it matters: Cloud margins are doing the analytical heavy lifting (Google Cloud op margin 32.9%, op income tripled to $6.6B). The binding constraint isn't demand — it's data centers, silicon, and energy. Half of US 2026 data centers have been delayed or canceled.
The Blend
Connecting the dots across sources
Compute scarcity is now the organizing principle of the industry
- Across the news today, five separate stories describe the same physical bottleneck: OpenAI's $138B AWS commitment plus a 2GW Trainium order, Anthropic's roughly $65B in pledged hyperscaler equity reframed as reserved gigawatts, Alphabet capex jumping to $180-190B with Pichai admitting Google Cloud was capacity-constrained, and Meta's CFO conceding the company continued to underestimate compute needs.
- On X, the data-center capacity wall topic notes that out of 12 GW of US AI data center capacity announced for 2026, only about 5 GW is actually under construction — a near-50% delay or cancellation rate.
- In the research, DeepSeek-V4 cuts per-token inference FLOPs by 73% and KV-cache by 90% at one million tokens, which signals that efficiency is the active research frontier specifically because compute is the binding constraint.
Agentic coding went from hype to infrastructure crisis in the same week it became a $2.5B business
- Across the news today, Anthropic's Claude Code is at $2.5B+ ARR and Cursor just shipped a TypeScript SDK to launch, steer, and compose custom agents — both on April 29.
- On X, the AI Agents Mature topic centers on GitHub buckling under 17 million agent-generated pull requests in March alone, with one of GitHub's earliest power users moving Ghostty off the platform after 18 years because it is no longer a place for serious work.
- On YouTube, Karpathy's Sequoia talk reframes the move from vibe coding to disciplined agentic engineering and calls LLMs statistical ghosts, while in the blog coverage, the Pragmatic Engineer interview with Pi creator Mario Zechner argues human judgment still anchors the agent era.
The open-weights frontier is closing on three fronts at once
- Across the news today, the Pentagon branded Anthropic a supply-chain risk for keeping its published guardrails, then handed Gemini a classified contract requiring adjustable safety filters, which leaves no single lab as a gatekeeper on classified workloads.
- In the blog coverage, AlphaGo's creator David Silver just left DeepMind with $1.1B to bet that LLMs hit their data wall, while IBM Granite 4.1 stands out as a rare deep-architecture disclosure on an enterprise-grade open model.
- In the research, AutoResearchBench shows Claude-3.5-Opus and Gemini-1.5-Pro scoring under 10% on long-horizon scientific search despite topping 80% on general web benchmarks, a strong counter-narrative to the deep research agent wave.
Slow Drip
Blog reads worth savoring
Pi creator Mario Zechner and Armin Ronacher debate where AI coding hits its limits and why human judgment still anchors the agent era.
Skips the benchmarks and hype to lay out the operational friction nobody warns you about before you commit to running your own LLMs.
Hands-on walkthrough of Gemini's Deep Research Agent for autonomously planning, searching, and synthesizing long-horizon research tasks into cited reports.
Demystifies the PyTorch-to-CUDA pipeline by building a complete six-stage ML compiler from scratch in pure Python.
Cursor opens up a TypeScript SDK to launch, steer, and compose custom agents directly on its platform.
David Silver, the architect behind AlphaGo and MuZero, just walked away from DeepMind with $1.1B in funding to bet against the data scaling curve.
IBM pulls back the curtain on the architecture and training decisions behind Granite 4.1, a rare deep look at an enterprise-grade open model.
The Grind
Research papers, decoded
Formalizes a paradox: when firms automate jobs to cut costs, each one pockets the full savings but only suffers a small share (1/N) of the resulting collapse in consumer demand. Competitive pressure drives an automation arms race past the level that would maximize collective profit. Universal Basic Income and capital taxes don't fix it — only a Pigouvian tax on automation itself does.
A new MoE family (V4-Pro at 1.6T params, V4-Flash at 284B) built around a hybrid Compressed Sparse + Heavily Compressed Attention stack. At 1M tokens, V4-Pro cuts per-token inference FLOPs by 73% and KV-cache by 90% vs. V3.2, while posting a 62.7% Chinese-writing win rate against Gemini-3.1-Pro and 67% on an internal code-agent benchmark.
A 1,000-task benchmark from Renmin University split into 600 'Deep Research' tasks and 400 'Wide Research' tasks. Claude-3.5-Opus and Gemini-1.5-Pro score under 10% here despite topping 80% on general web-browsing benchmarks. The bottleneck isn't search turns — it's long-horizon scientific reasoning.
On Tap
What's trending in the builder community
AI talent agent that learns your preferences over iMessage/WhatsApp and surfaces roles plus direct intros.
'General AI Agency' that converts any GitHub skill into a tagged team (COO/CMO/CTO) with shared persistent memory.
One API for TikTok, Instagram, YouTube, X, LinkedIn, and Facebook scraping.
Natural-language website/web-app builder, now mobile-first.
Edge/on-prem vector database claiming a 22x QPS advantage over Milvus and Qdrant at 10M vectors.
Y Combinator. Hassabis lays out unsolved AGI problems — memory integration, continual learning, reasoning limits — and gives founders strategic guidance.
Sequoia Capital. Karpathy reframes the move from 'vibe coding' to disciplined 'agentic engineering,' calling LLMs statistical 'ghosts.'
AI Engineer. A practitioner playbook for the LFM2.5 ~1B-param recipe.
OpenAI. Researchers describe using ChatGPT to crack a 42-year-old conjecture.
@RoundtableSpace's free-AI-courses thread hit 49k likes and is the highest-engagement AI topic on X today.
@rohanpaul_ai documents agent-load buckling GitHub and Hashimoto pulling Ghostty off the platform.
@TFTC21 surfaces the data-center capacity wall driving the entire compute-scarcity narrative.
Anthropic's skill for production-grade frontend interfaces that reject generic AI aesthetics.
Discover and install skills from the open agent skills ecosystem.
Captures learnings, errors, and corrections for continuous improvement; Clawhub's most-installed skill.
Roast Calendar
Upcoming events & gatherings
Last Sip
Parting thoughts & a teaser for tomorrow
Here's the thing that's sticking with me: every story today — the AWS-OpenAI move, the Anthropic $1T print, the Pentagon deal, even the $725B capex number — is downstream of compute scarcity. We're watching an industry pivot from 'who has the best model' to 'who can rack the most GPUs the fastest, and at what political cost.' Tomorrow we'll be watching for the Anthropic May board meeting decision, any movement out of Oakland on the Musk trial, and whether any of those delayed data centers actually break ground. Drink your water, ship something small, and we'll be back in your inbox tomorrow morning.