The Information Machine

2026-06-29

A tungsten supply squeeze threatens every advanced chip as China's export controls drive WF6 prices up 90% and Mitsubishi Materials warns investors, while NVIDIA cancelled its flagship Rubin Ultra GPU and the US government's Anthropic ban expanded to cover military contractors.

What

Two hardware supply crises emerged simultaneously. China controls roughly 80% of global tungsten mining and refining and has cut exports approximately 50% year-over-year, driving prices of tungsten hexafluoride — the gas used in chemical vapor deposition to form tungsten contacts in every advanced chip — up roughly 90%, with Mitsubishi Materials making alarming disclosures to investors about availability [1][2]. Separately, CXMT became the world's fourth-largest DRAM manufacturer, and Apple is actively lobbying the White House to source memory from CXMT despite its Pentagon military company listing because the shortage is that severe [3][4]. On the compute side, NVIDIA cancelled the original 4-die Rubin Ultra GPU three months after announcing it at GTC 2026, with the replacement delivering roughly half the intended performance; SemiAnalysis separately reported that Anthropic already runs substantial Claude Code inference on AWS Trainium and trains Claude on Google TPUs [5][6]. The US government's action against Anthropic was confirmed as broader than previously reported: CNN confirmed the ban extends to military contractors as well as federal agencies, with OpenAI securing a Pentagon deal within hours [7][8].

Why it matters

The tungsten-WF6 squeeze operates at the materials layer rather than the equipment or fab layer — it affects every advanced chip regardless of which fab or vendor produced it, making it a more broadly structural constraint than prior semiconductor shortage stories. NVIDIA's Rubin Ultra cancellation and Anthropic's documented migration to Trainium and TPUs arriving together suggest hardware alternatives to NVIDIA are becoming substantive enough to affect production decisions at leading AI labs, not just cost projections.

Open questions

  • With WF6 prices up roughly 90% [1] and Mitsubishi Materials warning investors about tungsten availability [2], how quickly can non-Chinese tungsten mining and refining capacity scale — and which chip manufacturers face the greatest exposure if the shortage deepens?

  • Apple is lobbying to source DRAM from CXMT, a Pentagon-listed Chinese military company, because the memory shortage is structurally severe [4]; does the White House grant an exception, and what does that signal about the coherence of its China technology policy?

  • NVIDIA cancelled Rubin Ultra and Anthropic has moved substantial Claude inference to AWS Trainium [6][5]; is this a leading-lab anomaly or evidence that CUDA dominance over AI training and inference is narrowing in practice?

  • The US government ban on Anthropic now covers military contractors [7] and OpenAI moved into the Pentagon gap within hours [8]; does the administration have a stable policy rationale for concentrating US government AI access in OpenAI alone, or is this an ad hoc sequence of decisions?

Thread movements (17)

  • semiconductor-critical-materials — A new thread: China's export controls have cut tungsten exports roughly 50% year-over-year and driven WF6 prices up approximately 90%, with Mitsubishi Materials making alarming investor disclosures about availability — WF6 being the critical gas for tungsten contacts in every advanced chip [1][2][9].
  • cxmt-dram-competitive-rise — A new thread: CXMT reached approximately 7.67% of global DRAM revenue, Apple is lobbying the White House to source DRAM from CXMT despite its Pentagon military company listing, and CXMT secured a multi-year server DRAM deal with Tencent worth approximately $2.94 billion [3][4][16].
  • nvidia-rubin-execution-failure — A new thread: NVIDIA cancelled the original 4-die Rubin Ultra GPU three months after announcing it at GTC 2026, with the replacement delivering roughly half the original performance, while SemiAnalysis reported Anthropic runs substantial Claude Code inference on AWS Trainium and trains on Google TPUs — framing both as evidence of CUDA moat erosion [5][6][27].
  • fable-mythos-export-control — CNN confirmed the US government's Anthropic ban extends to military contractors as well as federal agencies, OpenAI secured a Pentagon deal within hours, and Zvi Mowshowitz called WSJ reporting that China's GLM-5.2 matched Mythos 'obvious nonsense' [7][8][30].
  • us-ai-policy-regulation — Meta is confirmed still holding out on the government AI review framework while all other major US labs are engaged, and critics labeled the government-gated access model 'AI feudalism' [35][33].
  • ai-macro-economic-disruption-signals — The BIS warned that debt-financed AI infrastructure — hyperscaler bond issuance over $100B, private credit funds at 15% AI exposure — could seed a major financial shock, and Chinese hedge funds Wealspring and Banxia separately concluded AI stock valuations have entered super-bubble territory [39][40].
  • openweights-opensource-debate — Dario Amodei's 2023 Senate testimony warning open-source AI was heading down 'a very dangerous path' was surfaced alongside his current 'red herring' framing, and SemiAnalysis drew a parallel to Microsoft's 2001 congressional warnings about Linux [49][50].
  • rsi-governance-moment — New items documented the US government's pattern of shifting from safety-focused model review to direct frontier AI access allocation across both the Anthropic and OpenAI interventions [52][53].
  • nvidia-enterprise-ai-ecosystem — Claude models on NVIDIA GB300 NVL72 reached general availability on Azure through the Microsoft/NVIDIA/Anthropic partnership, and Palantir deployed NVIDIA Nemotron open models in air-gapped environments for US government agencies with full data sovereignty retained by the agencies [54][55].
  • ai-infrastructure-investment-picks — Milk Road AI's 'Save This' series expanded to Marvell and Meta, while Apple's reported exploration of CXMT as a memory source triggered a Micron selloff — the first named-customer signal that Chinese supply alternatives are being actively evaluated [80][81][82].
  • ai-chip-price-inflation — Apple is seeking US government permission to source DRAM from CXMT, a Pentagon-listed Chinese supplier, adding a geopolitical supply-chain dimension and revealing the depth of Apple's supply concentration risk; Micron's Q4 guidance of $49-51B extends its financial momentum [15][85].
  • ai-beyond-screens — Meta sold 7 million AI glasses in 2025 — triple the combined 2023-2024 total — in a market that grew 322%, with production capacity discussions targeting 20 million units annually by end 2026 [86][87][88].
  • ai-ipo-public-markets — Social media commentary noted Oracle stock falling on OpenAI IPO delay concerns, and one outlet cited a December 2026 internal OpenAI IPO target in tension with NYT/CNBC reporting about a 2027 lean [89][90].
  • local-coding-agents-ecosystem — Anthropic's June 12, 2026 export-control access suspension converted the API access-risk argument from theoretical to a documented incident, with the enterprise response characterized as building hybrid stacks that retain frontier APIs while deploying open-weight models as continuity insurance [94].
  • spacex-cursor-acquisition — New items amplified the previously reported story of Elon Musk confirming Grok 4.5 was trained on Cursor data with no new factual additions [95][96].
  • china-etch-localization — A report corroborated the ongoing trajectory of China's domestic etch equipment localization at CXMT, consistent with the established picture of Naura displacing imports at China's leading DRAM fab [97].
  • ai-agent-economics-enterprise — New items were social media amplification of the OpenAI agents report with no original claims [98].

Notable items (4)