The Information Machine

2026-06-05

SpaceX's IPO filing revealed it collects $2.17 billion per month from Google and Anthropic for AI compute, while both OpenAI and Anthropic publicly named recursive self-improvement as a current safety concern — the same week AI company IPO timelines hardened and community resistance to data centers produced concrete bans and moratoriums across multiple US jurisdictions.

What

Google disclosed via SEC filing that it will pay $920 million per month for 110,000 Nvidia GPUs at xAI data centers [1], placing SpaceX alongside Anthropic's $1.25 billion per month Colossus lease for a combined total of roughly $2.17 billion per month, or $26 billion annually [2] — a figure that surfaced in SpaceX's own IPO filings ahead of a planned June 12 listing. In the same week, both Anthropic and OpenAI publicly acknowledged that current AI systems show early signs of recursive self-improvement: Anthropic disclosed that Claude authored more than 80% of its production code merged in May 2026 [3][4] and called for a global slowdown in frontier AI development [5], while OpenAI's policy blueprint stated it sees 'early signs of recursive self-improvement in today's systems' and called RSI 'potentially the most consequential frontier safety issue of the coming decade' [6]. On public markets, S&P Dow Jones Indices confirmed it will not waive profitability requirements or shorten seasoning windows for newly listed companies [7], removing accelerated index entry from the IPO calculus for SpaceX, Anthropic, and OpenAI. Community and regulatory resistance to AI data centers produced concrete actions: the Stratos hyperscale project in Utah was cut by 50% after water-use opposition [8], Monterey Park, California voted to permanently ban data centers [9], Illinois is pausing data center tax incentives starting July 2026 [10], and Seattle moved toward a one-year construction moratorium [11].

Why it matters

The Google and Anthropic compute contracts establish SpaceX as the structural backbone of frontier AI infrastructure at a scale — $26 billion per year — that rivals national research budgets, creating a concentrated supply-chain dependency in a single private company whose CEO has made public statements that now directly contradict those contracts in his own firm's SEC filings. Both leading AI labs naming recursive self-improvement as a live concern while simultaneously filing for IPOs and expanding compute deals puts investors and regulators in the position of pricing safety disclosures alongside growth narratives at the same moment.

Open questions

  • With Google and Anthropic paying SpaceX roughly $2.17 billion per month combined [2], and Musk's personal denial of the Anthropic contract contradicted by SpaceX's own SEC filing [12], how do regulators and counterparties evaluate the disclosure reliability of an infrastructure provider whose CEO's public statements conflict with corporate filings?

  • Anthropic and OpenAI both named recursive self-improvement as a current rather than future safety concern [3][6][5] while preparing IPOs — do these disclosures carry legal weight as safety representations to investors, or are they read as regulatory positioning with no binding consequence?

  • S&P's refusal to fast-track newly listed companies [7] removes a key post-IPO narrative for SpaceX, Anthropic, and OpenAI — does this affect their pricing or timing strategies, given that index inclusion drives significant institutional inflows after a listing?

  • Permanent data center bans [9], construction moratoriums [11], and incentive pauses [10] are now enacted across multiple US jurisdictions — at what point do local restrictions begin to materially constrain where new AI infrastructure can be sited, and which regions emerge as permissive alternatives?

Thread movements (10)

  • spacex-ai-compute-supplier — Google's SEC filing disclosed $920 million per month for 110,000 Nvidia GPUs at xAI data centers from October 2026 through June 2029 [1], confirming SpaceX collects roughly $2.17 billion per month from Google and Anthropic combined [2] — a figure that became public in IPO-related filings ahead of SpaceX's planned June 12 listing.
  • rsi-governance-moment — Both Anthropic and OpenAI publicly named recursive self-improvement as a current safety concern in the same week: Anthropic disclosed Claude authored over 80% of production code merged in May 2026 [3][4] and called for a global AI slowdown [5], while OpenAI's policy blueprint called RSI 'potentially the most consequential frontier safety issue of the coming decade' [6], generating debate over whether the statements reflect genuine alarm or regulatory positioning.
  • datacenter-water-opposition — Community and regulatory actions against AI data centers advanced concurrently: the Stratos project in Utah was cut 50% over water concerns [8], Monterey Park, California enacted a permanent ban with 86% voter support [9], Illinois announced an incentive pause starting July 2026 [10], and Seattle moved toward a one-year construction moratorium [11].
  • ai-ipo-public-markets — S&P Dow Jones Indices confirmed it will not waive profitability requirements or shorten the seasoning window for newly listed companies [7], eliminating accelerated index entry as a post-IPO strategy for SpaceX, Anthropic, and OpenAI as all three converge on 2026 listings.
  • agentic-internet-takeover — The bot-to-human traffic split for global HTML requests is now quantified at 57.4% bots vs. 42.6% humans [114], with a new economic framing: bot-majority traffic reduces CPM and conversion-rate revenue for publishers and advertisers even where human filtering works, because those pricing models are built on human audience assumptions.
  • anthropic-rapid-ascent — SpaceX's publicly posted S-1 discloses Anthropic pays SpaceX $1.25 billion per month for compute [12], settling the SpaceX-vs-xAI attribution dispute through SpaceX's own SEC filing while leaving Musk's personal June 2 denial unretracted against his company's disclosure.
  • ai-persistent-memory-race — Broad amplification of OpenAI's Dreaming V3 continued [146], with the actual rollout scope — Plus/Pro-only vs. all users — remaining contested between OpenAI's initial framing and subsequent reporting; a third-party AgentMemory tool for Claude Code [147] surfaced as a parallel developer response to the cross-session context gap.
  • microsoft-build-2026 — Social amplification of Microsoft Build 2026 continued [183][184], with MAI-Thinking-1's SWE Bench Pro score and a Linux-at-Build angle joining the existing record without resolving the contested 'clean commercial data' training claim.
  • openai-pac-false-flag — Social media amplification of the Build American AI false flag story continued [190][191][192] with no new primary reporting, as the story settles after the initial investigative burst and organizational responses.
  • claude-opus-48-release — Minor amplification continued [193] with no substantive new claims on Opus 4.8 capabilities or the Anthropic S-1 filing beyond what prior coverage established.

Notable items (4)

  • Quoting Andreas Kling
    Simon Willison
    The Ladybird browser closed public pull requests because AI-generated code has broken the traditional assumption that a substantial patch implies substantial good-faith effort [194] — a concrete open-source governance response: contribution volume can no longer serve as a proxy for contributor accountability when any prompt can generate a large patch.
  • Google just made Gemma 4 much easier to run on phones and laptops by releasing QAT (Quantization-Aware Training) checkpo…
    Rohan Paul Twitter
    Google released quantization-aware training checkpoints for Gemma 4 that reduce the smallest model from 11.4 GB to 1.1 GB (0.84 GB text-only) [195], making a capable open model deployable on consumer phones and laptops while preserving more quality than standard post-training quantization.
  • Anthropic’s new chemistry report has a genuinely wild result.
    Rohan Paul Twitter
    Anthropic's chemistry evaluation report found Claude Opus 4.7 competitive with dedicated NMR software and capable of inferring molecular structure from spectra — working the problem in reverse [196] — a domain-specific result that, if independently verified, extends AI's reach into scientific instrumentation workflows.
  • Better self-improving agents need better solvers, not bigger update-writing models.
    Rohan Paul Twitter
    A contrarian design argument holds that self-improving AI agents gain more from a stronger solver model than from a stronger update-writing evolver model [197], directly challenging the common practice of assigning the most capable model to the evolver role in self-improvement pipelines.