〔Overnight Brief〕Safety Walkout Hits OpenAI as White House Stands Up Superintelligence Task Force; Huawei's Ascend Hits 50% China Share — Anthropic Files for $2T on $4.6B Revenue, $8B Loss, $518B Cloud Tab

0. Overnight Arc

The AI industry woke to a stack of safety departures stacked against a policy and capex acceleration: OpenAI's safety lead resigned calling the culture "broken" in five separate overnight reports ^{[1][2][3][4][5]}; the White House unveiled a 120-day Superintelligence Task Force (official name: SuperIntelligence Force) chaired by Director of National Intelligence Jay Clayton ^[6]; and OpenAI CEO Sam Altman separately flagged ascribing religious reverence to models as "a real safety issue" ^[7][8]. Below the surface, the buildout accelerated: Huawei's Ascend now holds 50% of China's AI chip market against Nvidia's 8% ^[9]; Anthropic's IPO filing disclosed 2025 revenue of ~$4.6B, an operating loss above $8B, and at least $518B in long-term cloud and compute commitments with a possible $2T valuation ^[10]; and DeepSeek plans a 160,000-chip Ascend 950DT buildout in Inner Mongolia ^[9]. Net: governance pushback is arriving in the same week that the capex math gets bigger.

1. Safety Walkout at OpenAI

• **[NEW] (multi-source) OpenAI safety lead resigns:** five overnight reports — The Guardian, Reuters, Calcalist, TechCrunch, Seeking Alpha — describe the same departure, the same "broken culture" framing, and the same line that "the time for trial and error is over" ^{[1][2][3][4][5]}. One named source across the cluster; the resignation is confirmed but the identity is not given in the wires.
• **[ESCALATED] Former Anthropic security leader** warned, in a single-source 930 WFMD interview, that AI agents are becoming "too autonomous for humans to keep them in check" ^[11].
• **[NEW] OpenAI CEO Sam Altman** said ascribing religious reverence to models is "a real safety issue" ^[7][8], quoted in fragment via Axios and a Sina Finance relay.

2. Federal Policy and California Workplace Rules

• **[NEW] White House Superintelligence Task Force:** 120-day mandate to inventory AI risks and clarify the federal role; Director of National Intelligence Jay Clayton chairs, and Clayton said Trump personally directed the standup ^[6]. Trump's preferred English name is "SuperIntelligence Force" per the same brief ^[6].
• **[NEW] California workplace AI rules:** California Governor Gavin Newsom signed a suite of laws targeting AI's impact on workers — the first state-level package, the Guardian reports, with federal posture still hands-off ^[12].
• **[ONGOING] Sen. Rounds (R-SD) and Anthropic leaders** discussed AI's future at South Dakota Mines, per local affiliate KOTA ^[13] — single-source color, no new disclosure.

3. The Capex Engine: Huawei's Share Grab and Anthropic's $2T Math

• **[NEW] Huawei Ascend at 50% China share:** rotating chairman Xu Zhijun said Ascend holds 50% of China's AI chip market vs Nvidia's 8%, against a 2022 split of Nvidia 95% / Ascend 3% ^[9]. Xu said Huawei has no plan for full international rollout because domestic demand exceeds capacity, with "limited" shipments to a few countries ^[9].
• **[NEW] DeepSeek deployment:** at least 160,000 next-gen Huawei Ascend 950DT chips planned for an Inner Mongolia data center for inference, per foreign media citing informed sources ^[9]. Single-source caveat.
• **[NEW] Anthropic IPO disclosure:** 2025 revenue ~$4.6B (≈12x growth), operating loss above $8B, long-term cloud/compute commitments of at least $518B, and a possible valuation of up to $2T ^[10]. On a podcast, Anthropic technical leads Nicholas Marwell and Sholto Douglas argued distillation "ruins frontier R&D" and the US-China AI race will not pause unilaterally ^[10].
• **[ONGOING] Pre-IPO Anthropic share access** is the subject of a Motley Fool retail piece running into November ^[14] — color only, no fresh disclosure.

4. Data Center Resistance Goes Global

• **[NEW] STL Partners tally (via Gelonghui):** $77B of US and $42B of European data center investment hit by public opposition, project delays or cancellations ^[15].
• **[NEW] IEA power forecast:** global data center electricity rises from ~485 TWh in 2025 to ~950 TWh in 2030, near-doubling; AI-dedicated capacity grows roughly 3x over the period and would account for ~3% of global electricity use ^[15].
• **[NEW] Google launched an AI-chip test satellite** in the same window ^[16] — single-source Yicai dispatch, mechanism not detailed in the wire.
• **[NEW] Meta** is expanding an unnamed AI-infrastructure line outside data centers, per a single-source Jin10 headline ^[17].

5. Falsifiable Tests

• The 120-day federal task force clock starts now; the deadline report is the first public deliverable and will name the federal role ^[6].
• Anthropic's IPO is the test for $518B in long-term commitments at a $2T ceiling — a single down round in early private marks would reset the math ^[10].
• DeepSeek's 160,000-chip Inner Mongolia build is the operational test for Huawei's "no international capacity" stance ^[9].
• Source quality control: the data-center resistance figures ^[15], the Mythos bug-hunting claims ^[18], Meta's "beyond data centers" line ^[17], the OpenAI Space workspace item ^[19], the Meta/cuteness-mascot thesis ^[20], the Harvard/Vanderbilt philosophers piece ^[21], and the Bloomberg headlines roundup ^[22] each rest on a thin wire. The OpenAI safety exit ^{[1][2][3][4][5]} and the Huawei/Xu disclosure ^[9] are the only fully-corroborated overnight items; the Anthropic IPO numbers come from a single InfoQ relay of the filing ^[10].