〔Overnight Brief〕AMD Tops $1.07T on a World Labs Bid as Mistral Ships 1T-Param le Chonk and OpenAI Spreads Compute Across Atlassian, Ironclad and Lean-Proven Math - Capacity Sprint vs Memory Wall and Partisan AI Fight

0. Weekly Arc

The AI buildout has split into three simultaneous races - compute (AMD's $1.07T cap, TSMC's US/Southeast Asia capex, TOWA's equipment plant, Google's nuclear deal with Constellation), capability (Mistral's 1T-param open release, OpenAI's Lean-formalized math) and distribution (Atlassian, Ironclad, Walmart, Google Docs, Sierra-Meta agent protocol) - yet the memory wall and a partisan Washington fight over Big Tech's role now set the ceiling ^[1][2][3]. Net: a capability blitz against a structural supply cap and a regulatory fork.

1. Compute, Capital and the Memory Wall

• **[NEW] AMD** rallied 4%+ to $658.52, market cap $1.07T for the first time; CEO Lisa Su expects chip demand to stay "very high" and plans a 2027 supply ramp ^[4]. Separately, AMD is pursuing a high-price bid for World Labs ^[5]. Source quality: the World Labs item is a single Caixin headline; the cap figure is a same-session print ^[4][5].
• **[NEW] TSMC and other Taiwan chip firms** are lifting AI capex in the US and Southeast Asia, per Nikkei ^[6].
• **[NEW] TOWA (Japan)** plans a chip manufacturing equipment factory to harden Japan's supply chain ^[7].
• **[NEW] Google** signed a nuclear power deal with Constellation sized to roughly one new reactor's worth of output ^[3].
• **[ONGOING] Memory shortage**: Morgan Stanley maps three workarounds - down-spec (Nvidia's Rubin offering multiple LPDDR5 sizes), inference workload split, and CXL pooling - and names Astera Labs (ALAB), Marvell (MRVL) and Cerebras (CBRS) as beneficiaries, keeping Micron (MU) and SanDisk (SNDK) at Overweight on the view that AI absorbs all available supply ^[2].

2. Capability, Distribution and Agents

• **[NEW] Mistral** released Mistral Large 4 / "le Chonk" - 1T parameters, open-source, trained on 4,000 Nvidia Grace Blackwell GPUs in Mistral's own European data centers over two months; preview is live for developers, security leads and governments, with general release later this month ^[8]. The Europe-sovereignty pitch is partially derivative on Nvidia silicon, and Mistral closed a €3B / $3.4B Series D at a €21B valuation in September ^[8].
• **[NEW] OpenAI** is publishing frontier-model math research on GitHub with Lean-formalized proofs verifiable by computer, averaging ~3 hours of ChatGPT Pro deep-reasoning compute per result and advised by the Princeton IAS Mathematics and AI Independent Advisory Group ^[9][10].
• **[NEW] OpenAI-Atlassian**: GPT-6 Astra and the GPT-5.6 series power Atlassian's Rovo agents and full platform; OpenAI says the framework will extend to more software partners ^[11][12].
• **[NEW] OpenAI-Ironclad**: joint research into AI agents handling complex contract workflows ^[13].
• **[NEW] Sierra + Meta Platforms + Walmart** (and others) are jointly developing an agent protocol ^[14].
• **[NEW] Anthropic**: Claude now interoperates with Google Docs, Sheets and Slides ^[15].
• **[ESCALATED] Anthropic** expanded its network certification to three permission tiers and is running deep per-agency reviews with the US government for the specialized access tier ^[16].
• **[NEW] Subscription value**: a study reported by The Register finds Anthropic's Claude subscription delivers more value than OpenAI's ^[17].

3. Regulation, Personal Agents and the Washington Fork

• **[ESCALATED] Personal AI race** is compressing: Meta shipped Muse on Sep 8 and Muse Charm (hardware) on Sep 23; OpenAI shipped "Dots" on Sep 29 for always-on personal agents; Instinct raised $1B Series C at a $10B valuation on Sep 28; Chinese players (Stepfun, Doubao and others) are following ^[18].
• **[NEW] OpenAI** will watermark ChatGPT outputs by default - but only in the EU - capping the geographic reach of the change ^[19][20].
• **[ESCALATED] Washington partisan split**: Sens. Elizabeth Warren and Richard Blumenthal (D) demanded disclosure on whether Big Tech lobbying reshaped the Trump administration's May-June AI framework; Sen. Bernie Moreno (R) wrote Anthropic CEO Dario Amodei warning against "alarmist" AI rhetoric and flagging the contradiction between Anthropic's pre-IPO posture and its extinction warnings ^[1]. Source quality: the bipartisan letter cluster is single-pass sourced from the same Wall Street CN piece ^[1].
• **[NEW] OpenAI's Human Rights Lead** said what the military could do with AI "keeps me up at night" - a quote that almost certainly feeds the Warren/Blumenthal letter ^[21].
• **[ONGOING] Regulatory debate**: a Guardian letter column argues independent oversight and regulation alone cannot prevent unsafe frontier releases, citing OpenAI's late-September model withdrawal on safety grounds ^[22].

4. China, Funding and Tail Risks

• **[NEW] China data center westward shift**: BNEF forecasts operational IT capacity rising from 20.6 GW (2025) to 45.5 GW by 2030, a 17% CAGR; power demand up 151% to 329 TWh (2.4% of national electricity), with Ulanqab (Inner Mongolia) emerging as a hub hosting Huawei, Alibaba, Apple, Kuaishou and ByteDance ^[23].
• **[ESCALATED] Huawei**: rotating chairman Xu Zhijun (Eric Xu) said Ascend's China market share has surpassed Nvidia's in measurable data, with supply - not share - now the binding constraint; the company unveiled the Peerium architecture and UnifiedBus (UB) with Haisi chief scientist Liao Heng ^[24].
• **[NEW] Funding chase**: OpenAI, DeepSeek and Moonshot are pursuing billions in new rounds, per Bloomberg's "Pulse with Francine Lacqua" ^[25].
• **Falsifiable tests**: (1) does Mistral's 1T-param open release dent frontier-API pricing, or is the sovereign-derivative frame a marketing one? ^[8]; (2) does the Warren/Blumenthal-Moreno split harden into a disclosure demand that reaches the White House AI framework? ^[1]; (3) does the memory wall force a real capex pause, or do CXL pooling and inference-split workarounds hold? ^[2]; (4) does AMD's $1.07T cap survive the World Labs integration and the 2027 supply ramp? ^[4][5].