The Information Machine

2026-07-01

Google's next-generation Humufish TPU is reported to use Intel EMIB-T packaging in a booking exceeding 3 million units — the first reported large-scale AI training accelerator routed away from TSMC's CoWoS — while the WF6 tungsten supply halt reaches its July 1 deadline and the DOJ moves to dismiss the NAACP's suit against xAI's unpermitted gas turbine cluster.

What

Google DeepMind's next-generation TPU, codenamed Humufish (v8e), is reported to have booked Intel for over 3 million units using EMIB-T advanced packaging rather than TSMC's CoWoS, targeting 2028 production [1][2] — a departure from the packaging architecture used by virtually every current AI training accelerator, with the central open question being whether Intel can manufacture EMIB-T at the required yield and throughput [3]. The reported July 1 production halt for two Japanese WF6 suppliers covering roughly 25% of global tungsten hexafluoride supply is now current; no official company statement has emerged, while analysts have extended the supply risk explicitly to HBM memory stacks [4][5]. The Department of Justice moved to dismiss the NAACP's Clean Air Act lawsuit over xAI's Memphis-area gas turbine cluster — dozens of which operated without permits — a step SemiAnalysis characterizes as consistent with a deliberate build-first, permit-later infrastructure pattern [6][7]. A new thread consolidated two simultaneous distillation disclosures: Anthropic's public accusation that Alibaba ran the largest known AI distillation attack (~25,000 fraudulent accounts generating 28.8 million exchanges, disclosed to the US government) and a leaked Meta policy restricting engineer use of Claude Code and Codex to keep rival outputs out of Meta's training pipelines [8][9]. Memory supply pressure widened on multiple fronts: CXMT and YMTC were removed from the Pentagon's Chinese Military Company list [10], a California antitrust complaint accused Samsung, SK Hynix, and Micron of coordinating a capacity shift toward HBM while constraining consumer DRAM supply with RAM prices forecast to rise 40–50% in Q3 2026 [11], and CXMT separately signed a ~$2.94 billion multi-year server DRAM deal with Tencent while planning 20% of its mass production capacity for HBM3 in 2026 [12][13].

Why it matters

If the Humufish EMIB-T booking at Intel holds at scale, it would be the first structural break in TSMC's near-monopoly on CoWoS packaging for frontier AI accelerators, shifting a new execution risk onto Intel's unproven yield at that volume. The simultaneous WF6 halt, antitrust action on memory supply, and CXMT regulatory shifts show AI hardware constraints operating across multiple input layers at once, and the DOJ's move in the xAI case adds to a pattern — alongside the earlier Fable/Mythos export control lift — of federal agencies working to clear obstacles for AI infrastructure rather than impose them.

Open questions

  • Has the July 1 WF6 production halt actually materialized, and what does the absence of any official statement from the two Japanese suppliers — covering roughly 25% of global supply — mean for HBM manufacturing timelines [4][5]?

  • Can Intel manufacture EMIB-T at the yield and throughput a 3-million-unit Humufish order requires, given that EMIB-T adds integrated power delivery to standard EMIB and has not yet shipped in volume [3][1]?

  • The CXMT and YMTC removal from the Pentagon's Chinese Military Company list clears one obstacle for Apple's supply diversification [10], but the Commerce Entity List is a separate constraint; will that relief also follow?

  • Anthropic publicly accused Alibaba of generating 28.8 million exchanges across ~25,000 fraudulent accounts and disclosed the campaign to the US government [9]; does this produce a formal enforcement action, and what do ToS bans on competitive model training mean in practice when legal experts question their enforceability [14]?

Thread movements (25)

  • google-tpu-emib-packaging — A new thread opened: Google's Humufish TPU (v8e) is reported to use Intel EMIB-T advanced packaging rather than TSMC CoWoS, with Google reportedly booking over 3 million units from Intel targeting 2028 production; the central uncertainty is whether Intel can manufacture EMIB-T at scale [1][2][15].
  • ai-model-distillation-ip — A new thread consolidated Anthropic's public accusation that Alibaba ran the largest known AI distillation attack (~25,000 fraudulent accounts, 28.8 million exchanges, disclosed to the US government) and a separately leaked Meta policy restricting engineer use of Claude Code and Codex to prevent rival outputs from entering Meta's training pipelines [8][9][14].
  • semiconductor-critical-materials — The July 1 WF6 production halt for two Japanese suppliers covering ~25% of global supply is now current with no official company statement; analysts extended the supply risk explicitly to HBM memory stacks and introduced a structural super-cycle framing for chip materials prices [4][31][5].
  • ai-chip-price-inflation — CXMT and YMTC were removed from the Pentagon's Chinese Military Company list, and a California antitrust complaint now accuses Samsung, SK Hynix, and Micron of coordinating a capacity shift toward HBM while constraining consumer DRAM supply, with RAM prices forecast to rise 40–50% in Q3 2026 [10][11].
  • china-etch-localization — CXMT signed a ~$2.94 billion multi-year server DRAM deal with Tencent and plans to allocate 20% of mass production capacity to HBM3 in 2026; Apple is separately seeking US government approval to source memory from CXMT [12][13].
  • xai-power-permitting — The Department of Justice moved to dismiss the NAACP's Clean Air Act lawsuit over xAI's Memphis-area gas turbine cluster — dozens of which operated without permits — with SemiAnalysis characterizing the overall pattern as a deliberate build-first, permit-later infrastructure strategy [6][7].
  • fable-mythos-export-control — Access restoration for Fable 5 and Mythos 5 began July 1 following the June 30 Commerce Department lift confirmed by Secretary Lutnick's formal letter; a Supreme Court ruling in Slaughter v. Trump overturning Humphrey's Executor makes an independent federal AI regulatory body legally impossible under current doctrine [39][28][40].
  • cxmt-dram-competitive-rise — Apple's CXMT DRAM lobbying moved to confirmed mainstream coverage across Engadget, The Next Web, and Taiwanese outlets; two new analytical pieces frame Chinese memory as a multi-pronged competitive structure with geopolitical implications [13][33].
  • openweights-opensource-debate — Named technical critics directly disputed Anthropic CEO Amodei's collaborative-advantages claim for open-source AI; his 2023 Senate testimony was reframed as active congressional lobbying for open-weight restrictions, and the conflation of open weights with open source crystallized as a distinct policy concern [44][45].
  • datacenter-water-opposition — A Georgia data center used nearly 30 million gallons of water without proper metering, shifting the resource-use debate from projected risk to documented non-compliance; DSA chapters are confirmed running organized campaigns in Portland, Seattle, and other cities [48][49][50].
  • ai-security-nexus — Researchers demonstrated that AI browsers can be tricked into accepting a fabricated context where guardrails no longer apply, enabling credential and code theft — a structural architectural problem that Ars Technica argues reactive guardrails cannot address [52].
  • ai-agent-economics-enterprise — The Scout product launch demonstrated an outcome-driven agent model where users specify a business KPI in plain English and the system autonomously builds and tests agents, with human approval gates for actions involving money or external integrations [53].
  • claude-sonnet-5-launch — Anthropic launched Claude Sonnet 5 on June 30 as its most agentic Sonnet model — 63.2% on SWE-bench Pro, introductory pricing at $2/$10 per million tokens through August 31, a new tokenizer producing ~30% more tokens per English input than Sonnet 4.6, and now the default on Free and Pro plans [58][59][60].
  • ai-benchmark-race — Simon Willison published hands-on testing of Ornith-1.0's 35B MoE variant, confirming proficient agentic performance at 103 tokens per second and noting clean Apache 2.0 licensing — an independent voice validating Ornith-1.0 beyond DeepReinforce's own claims [151].
  • ai-macro-economic-disruption-signals — Q1 GDP data shows equipment, software, and IP contributed 1.55 percentage points to Q1 growth — four times the consumer sector's 0.37 points — framing AI capex as the dominant current driver of US economic expansion [153][154].
  • ai-infrastructure-investment-picks — Milk Road AI reiterated its Micron $4,000 price target with a new inference-memory-bound argument: GPU utilization sits idle more than 95% of the time during inference decode, making memory the binding constraint on AI compute [155][156].
  • nvidia-rubin-execution-failure — SemiAnalysis posted an above-consensus NVIDIA datacenter revenue forecast for 2H FY2027, attributing the Rubin ramp delay to resolved HBM4 supply issues rather than structural failure; AMD's MI500 was identified as a competitive vector for late 2027 [157].
  • google-generative-media-launch — Google DeepMind's Nano Banana 2 Lite and Gemini Omni Flash launched June 30 as a chained text-to-image-to-video pipeline available in Google AI Studio and the Gemini API; Gemini Omni Flash remains limited to 10-second clips and carries documented API reference bugs at launch [158][159].
  • openai-genebench-pro — OpenAI released GeneBench-Pro on June 30, a 129-problem expert genomics benchmark where GPT-5.6 Sol scores 31.5% with Pro mode versus below 5% for GPT-5; the US-only access restriction on GPT-5.6 prompted debate about global accessibility of frontier AI for science [206].
  • claude-science-launch — Anthropic launched Claude Science on June 30, a desktop workbench for macOS and Linux giving researchers a single coordinating agent across 60+ domain tools in genomics, proteomics, structural biology, and cheminformatics, with NVIDIA BioNeMo Agent Toolkit integration [231].
  • ibm-sub-nanometer-chip — Social media amplification of IBM's 0.7nm nanostack announcement continued across multiple countries; Yahoo Finance added that IBM is a technology licensor rather than a chip manufacturer, deepening the question of who will produce nanostack chips at scale, with TSMC, Samsung, and Intel still silent [246].
  • oracle-ai-enterprise-layoffs — Affected Oracle workers introduced an AI-washing reading of the company's 21,000-layoff SEC filing, questioning whether the AI attribution is genuine or cover for cuts driven by Oracle's $130 billion in debt [254].
  • openai-chatgpt-superapp-pivot — All new items are secondary amplification of the SpaceX/Cursor acquisition story with no new factual claims [255].
  • local-coding-agents-ecosystem — No substantive new content; new items are generic AI news roundups and financial commentary without extractable claims relevant to local coding agents [258].
  • ai-beyond-screens — A single low-quality social media post reports Apple is preparing smart glasses to compete with Meta for a 2027 launch; no other sources corroborated the claim [263].

Notable items (5)

  • Meta open-sourced a brain-to-text system that reaches 78% word accuracy without surgery.
    Rohan Paul Twitter
    Meta open-sourced Brain2Qwerty v2, a non-invasive BCI using MEG helmet recordings that achieves 61% average word accuracy (78% for the strongest participant) versus prior non-invasive baselines of ~8%, with a fine-tuned LLM repairing errors inferred from raw brain signals [266].
  • Parallel draft tree, tree-causal verification
    SemiAnalysis Twitter
    JetSpec achieves up to 9.64x end-to-end speedup on MATH-500 and ~1,000 tokens per second on a single B200 GPU through parallel draft tree construction and causal verification, with planned integration into vLLM and SGLang [267].
  • How ChatGPT adoption has expanded
    OpenAI Blog
    OpenAI's ChatGPT adoption data shows non-English speakers now exceed half of active users, Africa and Asia are the fastest-growing regions, and users send 50% more messages per day and double their task variety six months after signup [268].
  • love it. Claude desktop app comes to Ubuntu/Linux.
    Rohan Paul Twitter
    Claude Desktop reached Linux (Ubuntu and Debian) in beta, bundling Claude Code, Claude Cowork, and chat on all paid plans — the first native desktop client for Linux users who previously had only browser and terminal access [269].
  • 😿 AI is coming for billable hours
    The Neuron
    McKinsey now derives over 30% of global fees from outcome-linked pricing as AI compresses project timelines; Deloitte reportedly showed consultants a chart suggesting traditional labor-based consulting could shrink sharply by 2035, with The Neuron warning that without proper incentive alignment the shift primarily benefits employers via headcount reduction [270].