May 14, 2026

Agentic Brew Daily

Your daily shot of what's brewing in AI

Fresh Batch

Bold Shots

Today's biggest AI stories, no chaser

Altman testified for the first time on Tuesday, spending roughly four hours in federal court in Oakland defending OpenAI's nonprofit-to-for-profit restructuring against Musk's claim that he and Microsoft "stole a charity." He told the jury Musk's opening demand at OpenAI's founding was 90% of the equity, and that in a 2017 conversation Musk floated handing the company to his children if he died. Musk's lawyer Steven Molo went hard at credibility, leaning on prior critical testimony from Mira Murati, Ilya Sutskever, and Helen Toner. Closing arguments Thursday; Judge Yvonne Gonzalez Rogers could rule next week.

Why it matters: A bad ruling could unwind October 2025's reorganization into OpenAI Group PBC, force Altman and Brockman out, and torpedo the planned $1T IPO — with Microsoft on the hook as a co-defendant. Only two of Musk's original 26 claims survived to trial, but both (breach of charitable trust and unjust enrichment) are the dangerous ones. Vivian Dong put it bluntly: it would be "unprecedented" for a court to order the structural changes Musk wants.

At The Android Show: I/O Edition on May 12, Google reframed Android as "moving from an operating system into an intelligence system." The Gemini Intelligence package includes Auto Browse in Chrome, Gemini-powered Autofill, a Gboard dictation tool called Rambler, and "Create My Widget" for natural-language homescreen and Wear OS Tile generation. Then they dropped Googlebook — a premium Android-based laptop line with a Gemini-powered "Magic Pointer" cursor shipping fall 2026 from Acer, ASUS, Dell, HP, and Lenovo. Aluminium OS (built on Android 17) is officially eating ChromeOS.

Why it matters: This is Google's pre-emptive strike before Apple's anticipated AI reboot, and they compressed software, hardware, and a Chrome-level agent into one news cycle. The collateral damage is going to be brutal: OS-level Rambler dictation could flatten Wispr Flow, Typeless, and Willow overnight. As TechCrunch put it, "when a platform player enters a market at the OS level, stand-alone apps need a compelling reason" — better accuracy, deeper features, or stronger privacy.

The 17-CEO roster that went out Monday for Trump's Beijing trip very conspicuously did not include the most important chip CEO on the planet. Then media noticed, Trump personally phoned Huang on Tuesday, and Huang flew to Anchorage to board Air Force One during a refueling stop. The confirmed delegation now includes Musk, Cook, Fink, Solomon, Schwarzman, and Mehrotra for the May 14-15 summit. NVIDIA rose ~2.4% premarket; Chinese AI-linked stocks rallied on bets the visit unlocks more H200 supply.

Why it matters: Semiconductor export policy is now sitting at the center of US-China trade. The wild contrarian wrinkle: even after Trump approved H200 sales in December 2025 (with a 25% revenue cut to Treasury), zero H200s have actually entered China because Beijing customs is blocking them. NVIDIA reports fiscal Q1 on May 20 — one week after the summit — which makes summit deliverables a direct earnings catalyst. China is a ~$50B opportunity NVIDIA's currently locked out of.

On May 13, NVIDIA became the first company ever to reach a $5.5T market cap, with shares climbing roughly 3% to an intraday high of $227.16. They added ~$500B in fresh market value in less than seven months since first crossing $5T on October 29. The company is now worth more than the annual GDP of every nation except the U.S. and China — eclipsing Germany ($5.01T), Japan ($4.34T), the UK ($4.00T), and India ($3.92T). The surge happened to coincide with Huang joining Trump's Beijing trip.

Why it matters: The slope is the story. NVIDIA hit $1T in May 2023, $4T in July 2025, $5T in October 2025, $5.5T this week — a ~1,500% gain in five years. Hyperscaler 2026 capex of $635-670B is the demand engine, but every top-four customer is now building its own accelerator (TPU, Trainium, MAIA, MTIA). Dan Ives's "new oil or gold" line writes itself; the real question is how long the moat holds.

Cerebras priced its US IPO at $185 per share — above the upwardly revised $150-$160 marketed range — selling 30 million Class A shares for ~$5.55B in proceeds. That values the company at $56.4B fully diluted, making it the largest stock-market debut globally so far in 2026 per Dealogic. Trading begins on Nasdaq under "CBRS" today, May 14. The offering ran roughly 20x oversubscribed and Cerebras raised the marketed range twice in a single week.

Why it matters: The re-rate is absurd — $8.1B in September 2025, $23B in February, $56.4B at IPO. 7x in eight months. It opens the 2026 AI IPO wave and signals capital is rotating from training to inference. But the customer concentration is genuinely scary: MBZUAI was 62% of 2025 revenue and G42 was 24% (86% from two UAE-affiliated entities), and the $24.6B backlog is dominated by OpenAI's ~$20B commitment. Oh, and OpenAI lent Cerebras $1B secured by warrants for 33M+ shares. That's not a typo.

The Blend

Connecting the dots across sources

One financial flywheel, four different lenses

  • NVIDIA crossing $5.5T, Microsoft's confirmed $100B+ OpenAI spend, and Cerebras's $56.4B IPO are not separate stories — they are the same flywheel viewed from different equity stakes, with each milestone landing within the same 48 hours.
  • Reporters on X have been showing GPU memory roadmaps and capex charts that all trace back to the same handful of hyperscaler checks, with Coatue's NVIDIA memory chart drawing 200K views and Bloomberg confirming the Microsoft commitment number the same day.
  • Chamath's Substack primer this week argued the structural thesis — that AI now writes 75% of new code at Google and Claude Code daily commits crossed 134,000 — which is exactly the demand story justifying every one of those market-cap milestones.
  • The Cerebras filing showing OpenAI as a $20B backlog customer plus a $1B lender stapled to share warrants makes it explicit that these companies are increasingly each other's customers, lenders, and shareholders all at once.

Skills are now the organizing unit of the agent ecosystem

  • GitHub's daily trending list is suddenly stacked with skill repos: mattpocock/skills hit 3,372 stars in a day for what is literally one engineer's personal Claude directory, and obra/superpowers added 1,419 stars as an explicit skills framework.
  • Vercel Labs's find-skills meta-skill has crossed 1.5M installs on skills.sh, making it the single most-installed item in the entire ecosystem and effectively the front door for everything else.
  • Product Hunt's Open Vibe (289 votes) is a SaaS tutor built directly on Claude Code skills and Jotform's Claude App (233 votes) ships as a first-class Claude integration — both betting their distribution on the skill surface.
  • The supply-chain pressure is already showing up: ClawHub's Skill Vetter has logged 238K downloads as a security-first vetting layer before anyone installs anything, which only exists because the ecosystem got big enough to need one.

Anthropic is having its "have it both ways" moment

  • Anthropic shipped Claude for Small Business and 10 finance agent templates this week, signaling that they're aggressively expanding into high-trust, ship-ready agent workflows for non-technical users.
  • The YouTube CFO explainer from Krishna Rao detailed how Anthropic is sourcing roughly $100B in compute across AWS, Google, and NVIDIA — the back-end that all those new agent products actually depend on.
  • At the same time X is in open revolt, with #savesonnet45 trending and one user racking up 75K views complaining "we pay $200/month for Claude just to hit the limit in under an hour" — the user-facing cost of the same compute squeeze.
  • Anthropic responded by raising weekly Claude Code limits 50% through July 13, which TestingCatalog tied directly to Colossus 1 compute coming online, meaning the rate-limit pain and the multi-cloud compute deal are literally the same story.

Slow Drip

Blog reads worth savoring

Analysis · Chamath Palihapitiya (Substack)Deep Dive: The Agentic AI Economy

High-signal primer arguing why 75% of Google's new code being AI-generated and 134K+ daily Claude Code commits in early 2026 mark a structural shift, not hype.

Analysis · Pragmatic EngineerRevisiting "No Silver Bullets" in the age of AI

Orosz revisits Fred Brooks' 40-year-old engineering classic to test whether AI is finally the silver bullet software engineering has been chasing.

Tutorial · The Product CompassClaude Code for PMs: The Beginner's Guide

Zero-code-required walkthrough for installing Claude Code, wiring up MCP servers, and running frontier models for free, with a ready-to-use template.

Tutorial · Towards AIClaude Code Hooks: The Developer's Secret Weapon for AI-Controlled Automation

Goes beyond Claude.md prompting to show how hooks deterministically enforce guardrails, formatting, and tests at every lifecycle step.

News · Alibaba Cloud EngineeringQwen Conference 2026: A First Look at the Exhibition Highlights!

First-party preview of what Alibaba is showcasing at Qwen Conference 2026 — useful for tracking China's open-model push.

News · Simon Willison's Weblogllm 0.32a2

Release notes flag that reasoning-capable OpenAI models now route through /v1/responses, unlocking interleaved reasoning across tool calls for GPT-5-class models.

Research · Arxiviq SubstackFast Byte Latent Transformer

FAIR/Stanford team (Zettlemoyer, Potts, et al.) propose a faster byte-level transformer that could finally make tokenizer-free LLMs practical.

The Grind

Research papers, decoded

Language Modeling183 upvotes · alphaxiv
Continuous Latent Diffusion Language Model

Cola DLM rethinks text generation by moving away from left-to-right token prediction. It compresses text into a continuous semantic latent space, runs a block-causal diffusion transformer over that space, then decodes back to words — treating language modeling like image diffusion at the meaning level. Hits strong MMLU and RACE numbers with only 8-16 denoising steps versus hundreds for discrete diffusion, and offers a clean structural bridge toward unified text-image models on a shared semantic prior.

Video / Vision14 upvotes · huggingface
Relit-LiVE: Relight Video by Jointly Learning Environment Video

Relights video realistically without requiring known camera poses — the constraint that has blocked most prior methods from working on everyday footage. Fuses raw RGB frames with decomposed scene intrinsics and jointly predicts per-frame environment maps, cutting light-direction angular error from 50+° to ~20° while maintaining temporal stability on hard materials like glass and water. For video editing, AR object insertion, material editing, and delighting pipelines, this removes a major preprocessing step and works on in-the-wild video.

On Tap

What's trending in the builder community

mattpocock/skills

Matt Pocock open-sourced his personal Claude skills directory and it took off — 3,372 stars today, 78K total. Pure proof that the skills wave is now a personality-driven distribution channel.

CloakHQ/CloakBrowser

Stealth Chromium that passes every bot detection test, drop-in Playwright replacement with source-level fingerprint patches. 1,829 stars today. If you've lost a scraper agent to Cloudflare, you get it.

tinyhumansai/openhuman

Private personal-AI stack written in Rust — 1,595 stars today. The local and private AI movement continues to compound.

obra/superpowers

Agentic skills framework and software development methodology that works — 1,419 stars today, 189K total. Tied directly to the broader skills ecosystem boom.

rohitg00/agentmemory

Persistent memory for AI coding agents, benchmark-backed — 1,335 stars today. Addresses the one thing every coding agent currently doesn't have.

github/spec-kit

GitHub's own toolkit for spec-driven development — 1,299 stars today, 97K total. Still climbing fast.

Kelviq

Payments, tax, and billing built specifically for SaaS and AI companies — 3.5% + 40¢. End-to-end monetization including subscriptions, usage-based billing, and license keys.

Open Vibe

Turns Claude Code into a SaaS tutor that teaches concepts while you build. Targeted at builders who get stuck shipping.

Jotform Claude App

Build, edit, and analyze forms directly in Claude. Conversational form-building as a first-class Claude app integration.

Hyperswitch Prism

Apache-2.0 stateless library by Juspay to plug-and-switch payment processors with a few lines of code.

display.dev

One command publishes agent-generated HTML behind company SSO with in-line comments. Built for shipping agent artifacts inside an org.

Dark Factory: How OpenClaw Ships Faster Than You Can Read the Diff — Vincent Koc

AI Engineer talk. Vincent Koc argues static benchmarks fail for adaptive agents and introduces "intent engineering" — agents curating their own eval suites from production traces. Required watch for anyone running agents in prod.

How Claude went from $9 billion to $30 billion in one year | CFO explains

Invest Like The Best. Anthropic CFO Krishna Rao on procuring gigawatts of compute across AWS/Google/NVIDIA, sourcing $100B in compute, scaling laws, and how Anthropic's own finance team uses Claude.

The $1M+ Solo AI Agent Business (Full Course)

Greg Isenberg with Nick from Orgo. Tactical playbook for unlimited-agent offers at $5K-$10K/mo, vertical selection (avoid healthcare/finance), full stack (Hermes, Orgo, Composio, Obsidian), and model picks.

LLMjacking: How hackers steal your AI API keys and stick you with the bill

IBM Technology. A startup's API bill ballooned from $180 to $82,000 in two days after key theft. Covers attack surface, defenses, and AI-augmented red teaming.

Token经济学:AI时代的新货币战争

硅谷101. Deep dive into token pricing economics, "token-maxxing," Chinese-model cost arbitrage, and the rise of token-arbitrage business models like OpenRouter and Metronome.

NVIDIA crosses $5.5T and Microsoft confirms $100B+ OpenAI spend

Coatue's NVIDIA GPU memory roadmap (15K likes, 200K views) and Bloomberg's confirmation of Microsoft's $100B+ OpenAI commitment landed within the same news cycle. NVIDIA is now larger than Germany's GDP.

the gpu crunch is real

Julien Chaumond's 5-word distillation of Trump landing in Beijing with Jensen Huang for AI chip diplomacy. 2.5K likes.

First $250K bug bounty paid to an AI

"An AI just got paid $250,000 for finding a vulnerability that every top human auditor missed. Largest bounty ever paid to an AI." 7.1K likes. AISI says frontier-model cyber-task length is doubling every few months.

find-skills

Vercel Labs's meta-skill that discovers and installs skills from the open ecosystem. By far the most installed skill — the front door for everything else.

frontend-design

Anthropic's skill for "distinctive, production-grade frontend interfaces that reject generic AI aesthetics." Strong opinions, refreshingly so.

Self-Improving Agent

ClawHub skill that captures learnings, errors, and corrections to enable continuous agent improvement across runs.

Roast Calendar

Upcoming events & gatherings

2026 SaaStr AI Demo Day & After PartyThu, May 14 at 1:30 PM PT | San Francisco, CA
Sushidata Launch PartyThu, May 14 at 12:00 PM PT | San Francisco, CA
Agent Builders Breakfast - Founders & Builders in SOMAThu, May 14 at 8:00 AM PT | San Francisco, CA
SF Dinner: Coding Agents for Product & Engineering TeamsWed, May 13 at 7:00 PM PT | San Francisco, CA
Operators Supperclub: Peninsula AI LeadersWed, May 13 at 6:30 PM PT | San Mateo, CA
Building the Business Behind Open Source AI @ AWS SFThu, May 14 at 1:30 PM PT | San Francisco, CA
Vibecoding: World Building w/ LovableWed, May 13 at 6:30 PM PT | San Francisco, CA

Last Sip

Parting thoughts & a teaser for tomorrow

If you only remember one thing from today, make it this: every single one of the five big stories is downstream of the same flywheel. NVIDIA's $5.5T, Cerebras's $56.4B IPO, Microsoft's $100B commitment, Altman testifying in Oakland, Jensen on Air Force One — these are all different angles on a capital-and-compute concentration that is genuinely without historical precedent. The interesting part of tomorrow's news isn't whether the flywheel keeps spinning (it will, at least for a quarter). It's whether the legal system in Oakland, the customs office in Beijing, or NVIDIA's own top-four customers building their own chips end up being the first thing to chip a tooth on it. Closing arguments in Musk v. Altman are tomorrow. NVIDIA earnings hit May 20. Watch those two clocks.

We'll see you tomorrow. Stay caffeinated.