The Information Machine

2026-06-22

SpaceX's AI compute business reaches $2.32B/month with a third customer deal, OpenAI says compute scarcity is blocking products it has already built, and the Financial Times argues Anthropic's safety rhetoric contributed to triggering its own export control suspension.

What

SpaceX signed Reflection AI as a third AI compute customer, paying $150 million per month for Nvidia GB300 chips at a newly disclosed facility called Colossus 2, potentially totaling $6.3 billion through 2029, bringing SpaceX's combined monthly AI compute revenue to roughly $2.32 billion [1]; OpenAI president Greg Brockman stated that global compute supply is insufficient for current AI demand and that OpenAI cannot launch products it has already built for lack of infrastructure [2]. The Financial Times published a quantitative analysis arguing that Anthropic's safety-focused public communications — at 8x the volume of OpenAI's — may have contributed to triggering the export control directive that suspended Fable 5 and Mythos 5 [3], while Nathan Lambert separately identified GLM-5.2 as the first open-weight Chinese model competitive with Opus 4.5 in coding agent workflows, adding competitive pressure on Anthropic's core revenue base during the outage [4]. Sakana AI (Tokyo) launched Fugu Ultra on June 22, a multi-agent orchestration system routing decomposed tasks across GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro through a single endpoint using a 7B RL-trained routing model, with self-reported benchmark claims of parity with Fable 5 and Mythos [5] but no independent verification [6], and a live coding test found it approximately 17x more expensive than comparable models [7]. Hedonically-adjusted semiconductor and computer import prices reached levels with no precedent in available historical data, rising 3.6% in May 2026 alone and 14.4% year-over-year, with increases reflecting genuine supply shortages rather than product upgrades — and the figures are pre-tariff, meaning actual buyer costs are higher still [8][9].

Why it matters

SpaceX collecting $2.32B/month from three AI compute contracts while OpenAI's president confirms compute scarcity is already blocking product launches, and semiconductor import prices are at historic highs, shows the infrastructure gap is actively constraining what can be built and shipped today, not merely inflating future costs. The FT's argument that Anthropic's own rhetoric contributed to triggering its export control suspension, combined with a credible open-weight Chinese alternative appearing during the outage, shifts the political and competitive texture of that story from straightforward government overreach toward something more contested.

Open questions

  • SpaceX's Reflection AI contract includes a mutual cancellation clause [1] that the anchor Anthropic and Google compute contracts apparently lack — does this structural difference signal different confidence levels on both sides, and what is SpaceX's exposure if demand conditions shift at Colossus 2?

  • Sakana Fugu Ultra claims benchmark parity with Fable 5 and Mythos [5] but all data is self-reported with no independent third-party verification [6], and a live test found it 17x more expensive than comparable models [7] — under what conditions does multi-model orchestration outperform a single frontier model on a cost-adjusted basis?

  • The FT analysis argues Anthropic's safety rhetoric at 8x OpenAI's volume contributed to the export control directive [3], while Anthropic's posture frames the ban as disproportionate government overreach — which interpretation is operationally accurate, and does the framing affect the path to reversal?

  • UK PM Starmer sought an exemption from the Anthropic export restrictions, the Trump administration refused, and the UK government then officially denied having sought one [10] — what does the episode reveal about whether the US is using AI access as a bilateral lever, and does the UK denial leave a formal diplomatic channel open or closed?

Thread movements (21)

  • spacex-ai-compute-supplier — SpaceX signed Reflection AI as a third AI compute customer at $150M/month for Nvidia GB300 chips at a newly named Colossus 2 facility, potentially totaling $6.3B through 2029 with a mutual cancellation clause [1]; Greg Brockman stated that compute scarcity is already preventing OpenAI from launching products it has already built [2].
  • fable-mythos-export-control — The Financial Times published a quantitative analysis arguing Anthropic's safety communications — at 8x OpenAI's volume — may have contributed to triggering the export control directive [3]; Nathan Lambert separately identified GLM-5.2 as the first open-weight Chinese model competitive with Opus 4.5 in coding workflows, adding competitive pressure on Anthropic's Claude Code revenue during the outage [4].
  • sakana-fugu-ultra — Sakana AI launched Fugu and Fugu Ultra on June 22 — a multi-agent orchestration system routing tasks across GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro via a 7B RL-trained router, with self-reported benchmark claims of parity with Fable 5 and Mythos [5] but no independent verification [6] and approximately 17x the cost of comparable models in live testing [7].
  • ai-chip-price-inflation — A new thread formed around AI-driven semiconductor price inflation: hedonically-adjusted import prices hit levels with no historical precedent, rising 3.6% in May 2026 and 14.4% year-over-year, with increases reflecting genuine supply shortages rather than product mix changes — and all figures are pre-tariff, meaning actual buyer costs are higher [8][9].
  • g7-ai-frontier-summit — The UK export control carve-out episode is now fully documented: UK PM Starmer sought an exemption from Anthropic export restrictions, the Trump administration refused, and the UK government officially denied having sought one [10] — all three stages now on record, with the UK denial directly contradicting press accounts.
  • europe-ai-sovereignty-deficit — Claude Mythos 5 has been publicly named as the first frontier model subject to US export controls, with the full text of Commerce Secretary Lutnick's letter to Anthropic now public and commentary noting litigation is possible [89][90]; a new signal — reports that US regulations grant allies unlimited AI access [91] — leaves Europe's actual exposure to the controls unresolved.
  • ai-agent-identity-infrastructure — A new thread formed around products giving AI agents their own persistent real-world identities: Delos launched Workers, positioning agents as full digital employees with their own email, phone, Slack, Teams, computer, and company-specific memory [93]; Atomic Mail (API-first agent inboxes) and Cloudflare ephemeral Workers deployments extend the same architectural direction [94][95].
  • ai-power-concentration-risk — A new thread formed around two independent concentration warnings: Microsoft CEO Satya Nadella argued AI power is concentrating differently from prior technology waves, focused on control of compute, capital, data centers, and user access [98]; Jack Clark's Import AI newsletter synthesized research showing AI systems reliably out-persuade expert humans in text interactions and framed this as a societal power-concentration concern [99].
  • rsi-governance-moment — The export control directive is now precisely dated and scoped: issued June 12, targeting Fable 5 and Mythos 5 specifically, with Anthropic disabling both globally on June 13 and SK Telecom's inclusion in Project Glasswing identified as the trigger [100][101]; a secondary reaction emerged arguing the suspension proved centralized AI can be shut off overnight, strengthening arguments for decentralized alternatives.
  • chinese-ai-competitive-rise — New items extended the Chinese AI competitive narrative [4][102][103], building on GLM-5.2's reception among practitioners and ASML's formal public denial — calling press reports of EUV machine shipments to China 'inaccurate and damaging to our reputation' — which remains the sharpest open factual dispute in the thread.
  • nvidia-isc-ai-science — A new thread formed around NVIDIA's ISC 2026 opening-day announcements: four coordinated blog posts covered production science results from JUPITER (Europe's Grace Hopper exascale system at Jülich) [104], a third LANL supercomputer (Veritas) with Vera CPUs and Rubin GPUs targeting 2027 [105], expanded NAIRR contributions spanning 700+ US research projects [106], and new scientific computing software with large benchmark claims across astrophysics, materials science, and particle physics [107].
  • spacex-cursor-acquisition — Three new angles complicated the deal's competitive framing: multiple reports confirm Anthropic separately struck a compute partnership with SpaceX for Colossus 1 access; xAI launched Grok Build as a live coding-agent product targeting Claude Code and Codex directly; and an analyst introduced data showing Cursor's coding-spend share had already been slipping from approximately 41% before the acquisition.
  • ai-macro-economic-disruption-signals — WSJ and Yahoo Finance coverage confirmed Accenture's approximately 50% year-to-date stock decline [110][111], with a signal that the drop reflects forward AI disruption concerns rather than quarterly earnings alone — consistent with the thread's established framing that markets are repricing AI's impact on incumbent IT services.
  • senior-researchers-agi-skepticism — A named rebuttal to LeCun's LLM-skepticism position appeared, directly arguing that LLM scaling could plausibly reach AGI and drawing Hacker News community debate — partially filling the counterargument side of the thread that had lacked a named proponent.
  • ai-datacenter-buildout-geography — Engineering News-Record reported FERC is weighing federal oversight of AI datacenter grid connections [119], a potential expansion of FERC's posture beyond the tariff-rewrite order covering six grid operators already in the thread.
  • meta-ai-workforce-disruption — A new report added that Meta plans to address its morale crisis — which CTO Andrew Bosworth characterized as near a 20-year low — partly by restoring employee snack perks, a response that sits in tension with the severity of the damage Bosworth himself described; other new items were amplification of the April layoff announcement with no new factual claims.
  • us-ai-policy-regulation — A single new item [120] extended coverage of the US AI policy landscape without adding perspectives beyond the established picture of federal enforcement actions and the Bloomberg characterization of Anthropic's safety principles as making it a White House target.
  • google-ai-talent-exodus — Social amplification of the Shazeer and Jumper departures continued across multiple languages and regions through June 21-22 [121][122][123], with a Fast Company article confirming the Shazeer move; the story has moved fully from breaking news to settled fact with no new substantive claims.
  • us-government-ai-ownership — New items were social media amplification with no claims or angles beyond the established picture of Sanders' stock-tax fund proposal and the Trump administration's equity-stake framing [124].
  • ai-coding-agents-robot-training — Coverage of the Project Fetch and ENPIRE results continued to cool — new items consisted entirely of social media posts and a YouTube video with no substantive new claims.
  • openai-chatgpt-superapp-pivot — No new substantive developments; all new items were social media amplification of the SpaceX/Cursor acquisition and OpenAI/Anthropic IPO race framing already established in prior coverage.

Notable items (1)

  • LLM-Driven Feature Discovery
    Alignment Forum
    Josh Engels (Alignment Forum) published a method he calls a 'black box SAE' — using an LLM autorater to generate and cluster behavioral features from model transcripts without internal activation access — applied to 100k Gemini chat transcripts and surfacing clusters including the model tracking its token budget, detecting roleplay vs. reality, and entering infinite reasoning loops [125]; the approach enables interpretability-style behavioral analysis of any model using only outputs, with no model internals required.