The Information Machine

2026-06-25

IBM's 0.7nm nanostack chip, OpenAI and Broadcom's 'Jalapeño' inference silicon, and China's LineShine taking the TOP500's top rank all land on the same day as OpenAI data showing its Codex agent generates roughly 99.8% of the company's own token output.

What

OpenAI's internal data shows its Codex coding agent accounts for roughly 99.8% of the company's own token output [1], and Goldman Sachs projects agent token usage industry-wide will grow 24-fold by 2030 [2] — the first concrete scale measurements of an agent economy that multiple firms are building around. On hardware, IBM announced a 0.7nm nanostack chip using vertical transistor stacking to fit roughly 100 billion transistors on a fingernail-sized chip and claiming ~50% performance gain over 2nm, with the '0.7nm' label a performance-node metaphor rather than a physical dimension [3]; simultaneously, OpenAI and Broadcom revealed 'Jalapeño,' a custom LLM inference chip built in nine months, with SemiAnalysis flagging the timeline requires 'making no mistakes' [4]. China's LineShine supercomputer was confirmed at 2.198 exaflops by CNN, Reuters, and The Guardian — the first domestically developed Chinese system to top the TOP500 since 2017 — though HPC experts note that ranking may not capture AI workload performance [5] [6]. The Trump administration's documented 10% equity stake in Intel from August 2025 [7] now provides concrete precedent for its public signals about taking financial stakes in AI companies, and Meta's engineer reassignment to data labeling is now reported at 30–50% of its engineering workforce [8], larger than prior estimates and unreconciled with them.

Why it matters

IBM's vertical stacking and OpenAI's custom inference silicon both signal that the next structural gains in AI compute come from architecture shifts rather than continued physical shrinkage — a transition with supply-chain and manufacturing implications separate from the familiar GPU scaling story. The Codex token-output figure suggests that within at least one major lab, agentic AI has already displaced single-turn interaction as the dominant use pattern, which changes how infrastructure capacity, pricing, and cost pressure models should be constructed across the industry.

Open questions

  • Codex accounts for ~99.8% of OpenAI's own token output [1] and Goldman Sachs projects 24x industry growth by 2030 [2]; does this concentration in coding agents reflect a general cross-industry pattern or an OpenAI-specific effect, and how does it interact with the cost pressure that has already led enterprises like Uber and Microsoft to pull back on expensive agent deployments?

  • IBM's 0.7nm designation is a performance-node metaphor relying on vertical stacking, not a physical dimension — chips below 1nm are not manufacturable with current methods [3]; at what point does the vertical stacking approach face its own density limits, and does volume production require TSMC or Samsung in ways that introduce supply dependencies not present in IBM's prior node claims?

  • LineShine at 2.198 exaflops is fully domestically developed [5], but Reuters and HPC experts say TOP500 rankings may not reflect AI workload capability [6]; does China's supercomputing leadership on this metric translate to frontier AI training advantages, or does it remain primarily a scientific and geopolitical signal?

  • The US government's 10% equity stake in Intel [7] was framed as industrial policy and criticized as reverse privatization; if similar stakes are pursued in AI companies, what governance conditions would attach — and does the Intel model suggest the terms would be commercially constraining in ways that affect labs' investment and partnership decisions?

Thread movements (22)

  • ai-agent-economics-enterprise — A new thread opens with the first consolidated synthesis of agent-economy scale evidence: OpenAI internal data showing Codex generates ~99.8% of the company's token output [1], Goldman Sachs projecting 24x agent token growth by 2030 [2], and Eric Schmidt publicly framing this as 'the agentic period in AI' in which every business will eventually build competing agents [9].
  • ibm-sub-nanometer-chip — IBM announced the 0.7nm nanostack chip on June 25, stacking transistors vertically to fit ~100 billion on a fingernail-sized chip and claiming ~50% performance gain over 2nm, with the '0.7nm' label explicitly a performance-node metaphor rather than a physical dimension — the technology targets AI data center workloads [3].
  • asic-gpu-market-dynamics — OpenAI and Broadcom announced 'Jalapeño,' OpenAI's first custom LLM inference chip, developed in nine months with a performance-per-watt claim against current accelerators — the first major AI model operator to enter custom silicon for inference — with SemiAnalysis noting the nine-month timeline 'requires making no mistakes' [4].
  • nvidia-isc-ai-science — China's LineShine supercomputer at 2.198 exaflops was confirmed as TOP500 #1 by CNN, Reuters, The Guardian, and South China Morning Post — fully domestically developed, first Chinese system to lead since 2017 — with Reuters and HPC experts qualifying that the TOP500 ranking may not reflect the best systems for AI workloads [5] [6].
  • ai-infrastructure-investment-picks — Micron's Q3 2026 earnings added concrete corporate backing to the de-commoditization thesis: CEO Sanjay Mehrotra publicly called memory 'a strategic asset,' and Micron disclosed Strategic Customer Agreements that Milk Road AI argues structurally end the company's 47-year commodity pricing model [82] [83].
  • us-government-ai-ownership — The Trump administration's 10% equity stake in Intel in August 2025 — framed as industrial policy and criticized by experts as reverse privatization — is now publicly documented [7], giving concrete precedent to the administration's signals about taking financial stakes in AI companies.
  • meta-ai-workforce-disruption — The scale of Meta's engineer reassignment to data labeling is now reported at 30–50% of its engineering workforce [8] — substantially larger than prior estimates and unreconciled with them — with The Next Web reporting a specific revolt at Meta's Applied AI unit and employees publicly using the term 'gulag.'
  • ai-beyond-screens — Figure AI signed a named commercial agreement with Catalyst Brands for warehouse humanoid deployment and documented Helix-02 handling 8-hour factory shifts [96] — the first named enterprise customer in the thread, moving the story from robot-count milestones to commercial contracts.
  • oracle-ai-enterprise-layoffs — Senator Elizabeth Warren publicly warned that AI-driven job losses could become a major disruption for American workers [101], adding a political dimension to Oracle's SEC disclosure explicitly attributing 21,000 layoffs to AI deployment, while Fast Company introduced a 'SaaSpocalypse' framing situating Oracle within a sector-wide displacement of SaaS roles.
  • fable-mythos-export-control — Legion LegalTech's federal lawsuit challenging the directive that suspended Fable 5 and Mythos 5 for foreign nationals — filed June 24 and confirmed by The Next Web, Law360, and others — is now established as the first formal judicial challenge to the order [102] [103].
  • sakana-fugu-ultra — Sakana's Fugu Ultra social coverage continues to bifurcate between geopolitical 'export controls hack' framing [104] and the 'not actually a model' counterframing, which gained independent amplification through new accounts [105] — while the arXiv technical report and a Requesty.ai reverse engineering of the orchestration architecture are now publicly available.
  • ai-ipo-public-markets — SpaceX stock has settled to ~$156/share with a reported $2.1T diluted market cap [108], and an unverified social media claim puts Anthropic's IPO target at October 2026 with $47B in annualized revenue [109] — higher than prior figures and without a sourced basis.
  • rsi-governance-moment — Social media commentary on the Fable 5/Mythos 5 shutdown is growing, with observers calling it underreported and among the most consequential AI regulatory developments of 2026 [113], but no new sourced facts or policy developments emerged this pass.
  • claude-tags-slack-launch — Bloomberg added mainstream financial press coverage [116], and a new strand of behavioral skepticism appeared: critics arguing that organizational delegation problems are human rather than technical, meaning a Slack integration does not resolve them [117].
  • telecom-ai-agent-platforms — Nokia announced upgraded agentic AI capabilities for autonomous networks and expanded its AWS partnership to run those systems on cloud infrastructure at DTW26 [119], adding a second major vendor platform alongside NVIDIA's and introducing a cloud-delivery model not previously documented in the thread.
  • nvidia-enterprise-ai-ecosystem — SemiAnalysis reported a firmware bug in NVIDIA's GB300 NVL72 rack hardware requiring a full system reboot every 66.5 days [120], introducing a concrete reliability concern into a platform narrative that had been uniformly positive.
  • ai-chip-price-inflation — Micron's Q3 2026 earnings drew market commentary citing an 84.9% gross margin at a memory chipmaker [124] — anomalous for a historically commodity sector — and SK Hynix is reportedly pursuing a $29B Nasdaq listing as price hikes spread from memory to foundry and OSAT segments.
  • ai-macro-economic-disruption-signals — Big Tech shed $2.7 trillion in market cap in June 2026 alone, and AI labs are projected to spend $725 billion on capex in 2026 — up 77% from $410 billion in 2025 [125] — adding a concrete near-term investment-as-inflation mechanism to the question of why AI's disinflationary productivity thesis has not materialized.
  • ai-security-nexus — The official NSA press release confirming the Five Eyes advisory [126] elevates the June 23 warning that AI models capable of severe attacks on governments and businesses could arrive within months from a press statement to a formal government record.
  • us-ai-policy-regulation — Multiple outlets confirmed the Trump administration is pressing Meta to join the voluntary AI pre-release review system [127], with no new detail on Meta's response or the administration's leverage beyond amplification of the established story.
  • ai-agent-identity-infrastructure — Press amplification of Claude Tag continued across Fortune, Quartz, The New Stack, and ITPro, with a new critical voice arguing that Claude Tag's enterprise-only, Slack-specific scope limits its reach despite heavy promotion [129].
  • spacex-cursor-acquisition — A social media post raised a fringe question about whether Anysphere might pursue an independent IPO rather than close the SpaceX acquisition [131], with no substantive basis cited; Quartz confirmed co-founders' net worths doubled, consistent with prior reporting.

Notable items (6)