The Information Machine

2026-06-18

Noam Shazeer is leaving Google for OpenAI less than two years after Google paid $2.7 billion to bring him back, as the DOJ designates xAI's Mississippi data center as national security infrastructure and Jefferies puts 2026 AI data center power demand at 24.1 GW on a path toward a 44.4 GW projected peak.

What

Reuters reports that Noam Shazeer — Google's Gemini co-lead and one of the central model builders of the Transformer era — is joining OpenAI, less than two years after Google paid $2.7 billion in 2024 to reacquire him alongside Character.AI talent [1]. The DOJ has moved to protect xAI's Mississippi data center as national security infrastructure, extending federal AI policy from review frameworks into direct asset protection for a specific company's facility [2]. Jefferies (via Milk Road AI) quantifies the AI power buildout: data center capacity ran at 6.9 GW in 2025, surged to 24.1 GW in 2026, and is projected to peak at 44.4 GW, with power availability and physical construction identified as the primary bottlenecks rather than capital [3]; xAI's infrastructure is priced at approximately $50 billion per gigawatt [4]. Anthropic's Fable 5 and Mythos 5 models remain under US export controls and globally disabled; 'dozens of security experts' are reportedly telling Washington the ban is backfiring [5], though no reinstatement timeline has emerged. Meta CTO Andrew Bosworth said employee morale is near its lowest point in Meta's 20-year history, comparing the internal mood to the Cambridge Analytica era following roughly 10% staff cuts and a significant shift of remaining employees toward AI roles [6].

Why it matters

Shazeer's departure from Google — after a $2.7 billion reacquisition — illustrates how expensive and unstable frontier AI talent retention has become, and his move directly strengthens OpenAI's foundational model research at a moment when Google and OpenAI are in direct competition on coding agents and frontier models. The DOJ's national security designation of xAI's Mississippi facility adds a new instrument to the federal AI toolkit: direct protective action for individual company assets, which sits in tension with state-level consumer enforcement targeting the same companies. The Jefferies GW data makes the physical-limit thesis on AI infrastructure quantitative rather than rhetorical.

Open questions

  • Reuters reports Shazeer is leaving Google for OpenAI [1] after Google paid $2.7 billion to bring him back in 2024 — was the reacquisition always understood as a fixed-term arrangement, and what does this mean for Gemini's leadership continuity?

  • Dozens of security experts are reportedly telling Washington the Fable/Mythos export ban is backfiring [5] — does organized expert opposition change the Commerce Department's calculus on reinstatement, given Lutnick's stated intent to keep the models offline?

  • The DOJ has designated xAI's Mississippi data center as national security infrastructure [2] — does this protection extend to other AI company facilities, or is it specific to xAI given its SpaceX and defense adjacency?

  • Jefferies projects AI data center capacity peaking at 44.4 GW [3] with xAI's infrastructure priced at $50 billion per gigawatt [4] — at what point does power availability become more constraining than capital, and which geographic markets hit that limit first?

Thread movements (12)

  • us-ai-policy-regulation — The DOJ has designated xAI's Mississippi data center as national security infrastructure [2], adding direct asset protection to the federal AI toolkit alongside existing review frameworks and creating a new asymmetry with state-level consumer enforcement targeting the same companies.
  • nadella-token-capital-ai-economics — Jefferies (via Milk Road AI) adds quantitative grounding to the AI infrastructure thesis: data center capacity at 6.9 GW in 2025 surged to 24.1 GW in 2026 and is projected to peak at 44.4 GW, with non-financial bottlenecks as the primary limiting factor [3]; xAI's $50 billion per gigawatt infrastructure pricing enters the thread as a neocloud valuation anchor [4].
  • fable-mythos-export-control — 'Dozens of security experts' are reportedly telling Washington the export ban is backfiring [5], which may indicate organized expert opposition beyond Luta Security's previously documented letter; all other new items are news amplification with no extracted claims.
  • llm-efficiency-vs-scale — Tensordyne's inference chip is now named 'Napier,' and IEEE Spectrum describes its claimed 17x tokens-per-watt and 13x throughput over NVIDIA Blackwell as 'bold' [21][22], introducing the first explicit editorial skepticism about the figures; the claims remain based on internal simulations with no third-party validation.
  • ai-agents-software-paradigm — A June 2026 arxiv paper argues AI agents are restructuring software from pre-written logic into systems that plan behavior on demand, with Kai-Fu Lee and enterprise vendors amplifying the thesis that multi-agent systems and MCP-managed API layers represent the next step in software development [23][24][25].
  • spacex-cursor-acquisition — Additional news coverage of the $60 billion Anysphere acquisition accumulated across outlets [26][27][28] with no substantive new claims beyond what was already established about deal terms, Composer 3, and the Q3 2026 expected close [29].
  • claude-fable-5-mythos-launch — New items are tutorials, social media posts, and Spanish-language commentary about a purportedly leaked Fable 5 system prompt [43][44][45], with no named sources or original reporting; the suspension's open disputes are unchanged.
  • chinese-ai-competitive-rise — Z.ai's GLM-5.2 (753B parameters, MIT license) now leads the Artificial Analysis Intelligence Index v4.1 with a score of 51 [49], extending the mid-June Chinese open-weight release cluster alongside MiniMax M3; Simon Willison is a new voice covering the release.
  • g7-ai-frontier-summit — Additional coverage of the Évian-les-Bains summit accumulated [50] with no substantive new claims; the thread's central story — frontier AI CEOs at a G7 working lunch arguing for shared infrastructure over national fragmentation — is unchanged.
  • ai-foreign-disinfo-operations — Further social media amplification of the June 10 OpenAI disclosure now includes Chinese- and Japanese-language posts on Instagram and Facebook [71][72][73]; one English-language comment frames data-center energy-cost targeting as a permanent feature of geopolitical influence operations, a modest analytical extension with no new factual developments.
  • openai-rosalind-biomedical — A new item was added to the thread [74]; the substantive development already in the synthesis — the AI chemist paper documenting yield improvements across 10,080 high-throughput reactions with explicit safety-bounded framing [75] — is unchanged.
  • clinical-ai-performance-benchmarks — All new items are empty social media amplification of the Nature Medicine study [76][77][78] with no extracted claims or new voices; the story has run its course as a news event.

Notable items (6)