Elon's Anthropic Deal, The Next AI Monopoly?, FDA for AI Panic, Trading the AI Boom

Key points

• **Anthropic ↔ Elon deal — Anthropic leased all of Colossus 1.** 220,000+ Nvidia GPUs, 300+ MW power. XAI training moves to Colossus 2. Claude code rate limits doubled, peak caps removed for paid users, Opus API volumes raised. **Direct execution of last week's [Chamath prediction](/issues/2026-05-03#compute-bottleneck-three-floors): 'Elon and Dario should do a deal tomorrow' — it happened 5 days later.** Anthropic's compute floor lifted; not solved.
• **'Elon Web Services' as the new hyperscaler.** Brad estimates **$4-5B incremental revenue this year** for SpaceX from leasing capacity, on top of analyst mid-$20Bs. Three facilities: Colossus (H100-heavy, leased), Macrohard (1.2 GW Blackwell), Macroharder. The deal solves the SpaceX IPO bear case where investors red-team the orbital data centre line item — the terrestrial Neocloud business is now real revenue that subsidises Grok training. **SpaceX IPO will trade at 40-50x revenue** (Brad).
• **Sacks's monopoly thesis — the issue's sharpest new framing.** Anthropic Q1 ARR: $10B (Jan 1) → $30B (Mar 31) → $44B (Apr). If 10x/yr exponential holds for 18 more months, $1T by 2027 = bigger than the rest of Mag 7 combined. Sacks's Rockefeller analogy is the central rhetorical move: 'Imagine if Rockefeller had called it Safe Oil and asked for an agency to regulate kerosene safety — he'd have built the biggest monopoly in history while everyone debated wick thickness.' Direct implication: read AI-safety lobbying as regulatory-capture cover.
• **Chamath's vibe-shift framing.** 'Tech leaders get a D-minus on PR, trending to F.' Build-up of antibodies against AI is real and avoidable. Three trillion-dollar net worths controlling 'the keys' is the actual political problem. The community + Main Street reaction has turned negative; DC will follow. **9 GW of data centres meant to come online this year — almost 50% now being protested**, most will get blocked (continues last week's Chamath line; Brad adds these protests are 'highly organised activists moving across the country', same playbook that killed US nuclear 30 years ago).
• **'FDA for AI' panic, then walk-back.** NYT reported Trump considering executive order for an 'AI working group' that would review new models pre-release. Catalyst: Anthropic's Mythos cyber-capability spooking the White House. Hassett mentioned it on Fox Business, Bessent followed up with softer 'innovation + safety calculus.' Susie Wiles statement walked it back. Sacks: 'no senior official supports an FDA for AI.' Polymarket 'Trump federal review of AI by May 31': **21%**.
• **Brad disclosed 25% of his portfolio is in memory** — SK Hynix (5x earnings), Samsung (6x), Micron (7x). Says they're 'not the stuff bubbles are made of.' **This is the direct portfolio expression of the Pope/Patel memory thesis from Issue 03** ('50% of 2026 hyperscaler CapEx is going on memory'). A high-profile public allocator-side confirmation.
• **Cyber upgrade cycle remains the cleanest sectoral call.** Mythos and GPT-5.5 Cyber both at frontier; all top-4 labs hit Mythos-grade in 3-6 months; Chinese models 80% there. KYC on frontier model APIs as Brad's policy proposal (pre-release period only; relax in general release). Beneficiaries: CrowdStrike, Palo Alto Networks, plus AI-pentest startups. Same list as Issue 03.
• **Hyperscaler revenue confirms the AWS/Azure/GCP race.** AWS $150B run-rate (+28% YoY), Azure $108B (+39%), Google Cloud $80B (+63%, fastest growing). Combined annualised ~$338B. Mag 5 headcount up only ~3% over 3 years while revenue compounded — direct evidence of the operating-leverage shift the issue keeps pointing to.
• **Chamath's 500-day reckoning timer is the issue's most actionable bear case.** 'There is literally not a scintilla of evidence that AI has helped lift the operating margins of the S&P 500.' S&P op margins went 11.0% (2023) → 11.8% (Q1 2024) → 13% (now), but he says that's post-Covid fitness, not AI causally. The fork ~500 days out: (a) opex shrinks (margins via job cuts and political backlash) or (b) revenues grow (margins via more output, society wins). Brad/Sacks counter that startups and the Mag 5 are already seeing it; Chamath holds it's not yet in S&P 500 aggregates.
• **Brad's stack-rank: 'No one is better on planet Earth than Elon at converting electrons to tokens.'** The five-layer SpaceX cake (per Shawn Maguire tweet Brad cites): launch → connectivity → compute → hyperscaler → space data centres → apps/models/other. The Tesla-SpaceX merger thesis: by year-end / mid-2027, expects Elon Corp consolidation; trades at premium because 'everyone else has stopped innovating' (Chamath's frame).
• **Two political wildcards worth tracking.** (1) Spencer Pratt for LA Mayor — viral ad team, debating Karen Bass; if he wins, LA tries the same retirement-protection ballot initiative that would block CA wealth tax. (2) Ken Griffin publicly leaving NYC for Florida after Mamdani's video targeting his house; says NYC repeating Chicago's exit pattern. NYC 'becoming a flyover city' (Brad). Both feed the tech-capital geographic-redistribution thesis from Issue 03's California episode.
• **Software TAM math** — Sacks: total spend on software developers ≈ $1T/yr; coding tokens 10-100x that creative capacity; software TAM 'easily doubles to $2T+'. This is the bull case for the Anthropic monopoly thesis — the addressable market exists, the question is concentration.

Notable quotes

Themes

• Anthropic compute floor lifted via Elon's Colossus deal but not solved
• Sacks's 'biggest monopoly in human history' framing for Anthropic
• Chamath's 500-day reckoning timer on AI margin causality
• Memory thesis validated by allocator portfolio disclosure
• FDA-for-AI political panic walk-back and the regulatory-capture frame