Yesterday's biggest story: Pope Leo XIV published "Magnifica Humanitas," a 43,000-word encyclical entirely focused on AI -- the first papal document of its kind. He calls for "disarming AI" and demands "robust legal frameworks, independent oversight, and a political system that does not abdicate its responsibility."
The headline detail: Anthropic co-founder Christopher Olah sat alongside the Pope at the Vatican press conference. Olah said decisions about AI "should not be left to people in the industry." This comes while Anthropic is fighting the Trump administration over military AI use -- the Vatican explicitly rebuked autonomous weapons.
Micron just hit $1 trillion market cap, up 75% in May alone (best month since 1987). Marvell is co-leading the AI rally to start the week. NVDA +23.5% since April, TSM +19.7%. Trump singled out Micron at a rally: "Micron, boy Micron's great, they're investing hundreds of billions."
Meanwhile, software stocks are getting destroyed -- Intuit down 63%, iShares Tech-Software ETF down 37%. The market is literally repricing: AI infrastructure up, software incumbents down. The rotation from software to semis is the biggest theme on Wall Street right now.
AgentSeal scanned 1,808 MCP servers and found 66% had security findings -- code execution risks most common, toxic data-flow patterns close behind. Wallarm tracked MCP vulnerabilities jumping 270% in Q1 2026, with AI-API vulnerabilities up 57% and agentic AI flaws up 67%.
Separately, a PraisonAI CVE (CVE-2026-44338) shipped with authentication disabled by default -- exploitation attempts began within 3 hours and 44 minutes of disclosure. The agent security surface is expanding faster than anyone is patching it.
DeepSeek made its 75% promotional discount permanent on May 24. V4-Pro input tokens now cost $0.0036/million vs. Claude Opus 4.7 at $5-25/million and GPT-5.5 at $5-30/million. A workload costing $348 on DeepSeek would run $2,500 on Claude or $3,000 on GPT.
The price cut is attributed to increased supply of Huawei Ascend 950 chips. Anthropic has accused DeepSeek of distillation attacks. This is the AI price war going nuclear -- and it has massive implications for agent economics.
Unitree Robotics heads to its STAR Market IPO hearing on June 1, seeking to raise 4.2 billion yuan ($619M) at a ~42 billion yuan valuation. The problem: Q1 2026 profit dropped 52% YoY to 40.3 million yuan despite 68% revenue growth. Rising R&D costs, sales expenses, and reliance on institutional procurement over commercial demand are squeezing margins.
Context: Unitree shipped 5,500 humanoid robots in all of 2025. EngineAI is targeting 10,000 units from Shenzhen. UBTECH delivered 500+ Walker S2 industrial units and targets 10,000 in 2026. The humanoid volume story is real -- but so is the margin compression.
Anthropic acquired Stainless for ~$300M on May 18. Stainless built the SDK infrastructure for OpenAI, Google, Cloudflare, Cerebras -- and Anthropic's own SDKs. One 4-year-old startup was the on-ramp for every major model provider. Now Anthropic owns the pipes. @aakashgupta's viral X thread mapped the full customer list.
The 66% vulnerability stat is getting discussed across dev communities. GitHub's response (building MCP security scanning) and the Wallarm 270% jump in MCP vulnerabilities are generating a "told you so" moment for security-conscious builders.
The image of an atheist AI researcher sitting beside the Pope to present an encyclical on AI is striking. Olah's quote -- "these decisions should not be left to people in the industry" -- lands differently from someone who co-founded one of the industry's most powerful companies.
Intuit down 63%, Adobe down, Salesforce down. Meanwhile Micron hits $1T, NVDA at record revenue. The market is making a clear bet: AI infrastructure eats software. Zacks is calling it the biggest rotation since cloud killed on-prem.
Thread concept: Unitree's IPO filing reveals the fundamental tension in humanoid robotics right now. Revenue is surging (68% YoY) but profit is halving. R&D costs are exploding, commercial demand hasn't caught up to institutional procurement, and margins are compressing even as units ship.
The insight: humanoid robots are following the EV playbook -- volume growth with margin compression -- except the TAM is even less proven. Unitree shipped 5,500 in 2025. EngineAI and UBTECH are both targeting 10,000 in 2026. We're watching a hardware scale race before the software stack is mature enough to justify the hardware cost. The winners will be the ones who solve autonomy (software), not just locomotion (hardware).
Connect to your thesis: this is why you killed the Embodied Agents hardware concept -- the value is in the software/protocol layer, not the metal.
Thread concept: Three data points from this week paint a picture. (1) 66% of MCP servers have security vulnerabilities. (2) MCP vulnerabilities jumped 270% in Q1 2026. (3) PraisonAI shipped with auth disabled by default and was exploited in under 4 hours.
The take: we're at the "move fast and break things" phase of AI agents. Every agent framework is racing to add tools and integrations. Nobody is racing to add trust, verification, and behavioral monitoring. The next wave of agent infra companies won't be about what agents CAN do -- they'll be about what agents SHOULD do. GitHub sees it (MCP security scanning). AgentSeal sees it. The market will follow.
Thread concept: Tie together three seemingly unrelated events from the past 72 hours into a single narrative about where AI is headed. (1) Pope Leo XIV publishes history's first papal encyclical on AI, with Anthropic's co-founder at his side, calling to "disarm" AI. (2) DeepSeek makes its 75% price cut permanent, making frontier-class inference 100x cheaper than Western alternatives. (3) Micron hits $1 trillion on pure AI infrastructure demand while software stocks crater.
The synthesis: the AI industry is simultaneously becoming cheaper (DeepSeek), more powerful (semis boom), and more dangerous (Vatican intervention). These three forces are pulling in different directions, and the companies that navigate all three -- building powerful, affordable, AND trustworthy AI -- will define the next decade. That's the agent infra thesis in a sentence.
Anthropic launched Claude Code Channels, letting developers control AI agents from Telegram and Discord. Plugin-based architecture, full local access, sender allowlists for security. Available to Pro/Max users. This bridges messaging platforms with Claude Code sessions for async agent workflows.
The World Humanoid Robot Games kicked off in Fuding, Fujian Province. Instead of lab demos, robot teams were deployed directly in tea plantations, working alongside tea masters. Unitree G1 robots are being tele-operated for sorting, picking, and pressing. Meanwhile China launched OpenHarmony, the country's first dedicated robot OS.
NVDA: $81.6B Q1 revenue (+85% YoY), $80B buyback, +23.5% since April
MU: $1T market cap, +75% in May, best month since 1987
TSM: +19.7% since April, 3nm demand strong
INTU: -63% from 52-week high, software selloff poster child
DeepSeek V4 Pro: $0.0036/M input tokens (permanent) vs Claude $5-25/M
MCP: 66% of servers vulnerable, vulnerabilities +270% QoQ
Unitree: Q1 profit -52% YoY, revenue +68%, IPO hearing June 1
Anthropic/Stainless: ~$300M acquisition, owns SDK infra used by OpenAI/Google/Cloudflare