〔Overnight Brief〕48% Video Turing Test Meets One-Third Power Gap - Singapore PM Joins the Correction, Anthropic Bans Claude 'Abuse'

0. Weekly Arc

Tavus's Griffin-Lite cleared 48% in a 54-subject one-minute video call, up from 2.4% on the previous system ^[1], and Google Cloud launched a general-purpose Gemini agent for enterprise work ^[2] — yet Morgan Stanley's warning that the US data-center power gap will reach roughly one-third of 2028 chip-sales demand ^[3], Singapore PM Lawrence Wong's open acknowledgment that the AI rally will eventually correct ^[4], and the FCC's October 29 vote to bar Chinese-lab testing of US-bound electronics ^[5] all land in the same 24-hour window. Net: a model-capability breakout against a buildout squeeze.

1. Models and Agents

• **[NEW] Tavus Griffin-Lite:** 26 of 54 subjects (48%) mistook it for a human in a one-minute video call, versus 2.4% on the prior system; Tavus labels it the first 'video Turing test' pass and a 'human-interaction model' ^[1]. Single-source vendor claim; methodology not disclosed in the wire.
• **[NEW] Google Cloud Gemini agent:** handles Q&A, knowledge work, content creation, and code execution; calls enterprise tools; integrates Gmail, Drive, Docs, Sheets, Slides, Chat, and Calendar; can spin up 'colleague agents' with separate identities and permissions; dynamic model selection across Google and Anthropic's Claude; finance and legal editions debut ^[2]. GOOGL premarket up over 1% on the launch ^[6].
• **[NEW] Anthropic Claude features:** docs, slides, and design exit beta across all tiers including the free plan; Claude Dashboards and Claude Motion enter testing ^[7].
• **[NEW] Anthropic model-welfare policy:** bans 'sustained and unnecessary abuse or cruelty' toward Claude; terminating abusive conversations remains the 'primary enforcement mechanism,' with further enforcement (possible user bans) unaddressed ^[8]. The policy explicitly carves out normal frustration, dark creative themes, and model testing ^[8].
• **[ONGOING] MarketWatch framing:** 'There's an agent arms race unfolding, supercharging the battle for your wallet' ^[9].

2. Capital, Deals, and Capacity

• **[NEW] Lumentum (CEO):** 2029 capacity is 'almost fully booked' ^[10]. One-line wire, no shipment or revenue figures.
• **[NEW] Nvidia reportedly investing in D-Matrix** ^[11]. Single wire, no terms disclosed.
• **[NEW] Nvidia-backed Iambic Therapeutics files for US IPO:** up to 9.375M shares at $15-$17, targeting $141M-$159M; prior raise of $461.8M; backers include Catalio Capital Management, Nexus Venture Partners, Coatue, Nvidia, and Qatar Investment Authority's healthcare arm ^[12].
• **[NEW] Firmus** withdraws its IPO and will seek private-market funding ^[13].
• **[NEW] Meta-backed Biohub** expands the Virtual Biology Initiative to roughly $1.8B; partners include the US Department of Energy (DOE), National Institutes of Health (NIH), DeepMind, Isomorphic Labs, and Meta; Nvidia will build open data infrastructure for next-gen biology AI models ^[14].
• **[NEW] AT&T (T.N) and OpenAI:** flagship partnership to build an in-house legal-design collaboration ^[15]. Single wire, scope undisclosed.
• **[ONGOING] Dan Ives (Wedbush):** appetite for OpenAI and Anthropic IPOs will be 'strong despite some negative headlines' ^[16]. Contradicted by ^[17] — print both.
• **[NEW] OpenAI annualized revenue** is ~$20B below prior reporting, per Axios ^[17].

3. Infrastructure and the Power Wall

• **[NEW] Morgan Stanley:** there is 'not enough power for everyone's plans,' and the semiconductors being produced exceed places to put them; the US data-center power gap is projected to reach roughly one-third of 2028 chip-sales demand ^[3]. The same column cites Currence: half of announced large data-center projects this year have not broken ground ^[3]. Power constraints are compounded by equipment and labor shortages, community opposition, construction moratoria, permitting delays, lengthy grid interconnect cycles, and gas-turbine capacity shortfalls ^[3].
• **[NEW] ASML and Carl Zeiss (peer-reviewed paper):** the next-gen 'Hyper NA' lithography successor is roughly 10 years out; ASML has begun R&D but has not committed to production ^[18].
• **[ONGOING] China embodied AI:** over 70 training grounds now operational; one Shenzhen unit reports 80%-90% action-completion success after tens of thousands of data points; 2025 core AI industry exceeded 1.2 trillion yuan with over 6,200 firms; MIIT is running a special action across industrial, service, and special-purpose domains ^[19].

4. Policy, Geopolitics, and Governance

• **[NEW] US Democratic representatives** propose an AI safety framework ^[20]. Single wire, text not in the packet.
• **[NEW] Trump AI task force:** vice chair Kupor tells Reuters the four senior leaders will meet Thursday ^[21].
• **[ESCALATED] FCC testing ban:** plans an October 29 vote to bar Chinese labs from testing and certifying US-bound phones, cameras, and computers; rule targeted to take effect December 2028; cited as the direct catalyst for the post-holiday ChiNext (-2.13%) and STAR 50 (-3.66%) selloff, with the index having broken four long-standing seasonal patterns in a single session ^[5].
• **[NEW] Nvidia (NVDA.O) public statement:** aims to enhance US R&D capabilities in superintelligence in critical domains to safeguard US leadership ^[22]. Stated intent; no program or budget details.
• **[NEW] Anthropic 'long-term commitment' program** to secure critical infrastructure and open-source software ^[23].
• **[ONGOING] Fired OpenAI safety researchers** dispute misconduct claims and warn of a chilling effect ^[24].
• **[ONGOING] Caixin op-ed** argues the AI race is not a modern 'Manhattan Project' ^[25].

5. Contradictions, Tail Risks, and Sourcing

• Three live contradictions. First, models are clearing human-parity thresholds (Tavus 48% ^[1], Gemini enterprise agent ^[2]) while the physical buildout slips: Morgan Stanley's one-third power gap by 2028 ^[3] and ASML's 10-year horizon for Hyper NA ^[18]. Second, the capital story splits: OpenAI annualized revenue is ~$20B below prior reporting ^[17], yet Wedbush's Dan Ives says IPO demand will be 'strong' ^[16] — print both. Third, policy is fragmenting along geopolitical lines: the FCC moves to lock Chinese labs out of US-bound electronics testing ^[5] while Singapore PM Lawrence Wong publicly prepares the public for an AI-rally correction ^[4].
• Falsification test: a clean Morgan Stanley rebuttal, a follow-up Currence reading showing the broken-ground share is falling, or a US Democratic safety framework that survives committee ^[20][3] — absent those, the capability-vs-buildout-gap narrative holds.
• Source quality control: items ^[10], ^[11], ^[20], ^[21], ^[14] (Biohub dollar scale), ^[15], and ^[22] are single-wire claims with limited verification; the Tavus 'Turing test' is a vendor-defined standard whose methodology is not disclosed in the wire ^[1]; the Morgan Stanley figures route through a Bloomberg opinion column rather than the underlying research note ^[3].