〔Day Digest〕On-Device Shift Drops Anthropic Server Utilization 26% From January Peak, Stress-Testing $990B 2027 Hyperscaler Capex — Goldman Lifts BTM Data Center Power to 67GW as 10Y Nears 5.17%

0. Weekly Arc

The AI infrastructure thesis is being stress-tested from both ends. On demand, Jefferies argues small language models are redirecting enterprise AI from centralized cloud to local hardware; Swerve Research Technologies' queue-wait "congestion index" puts Anthropic server utilization 26% below the January-February 2026 peak ^[1]. On supply, Goldman's Carbonomics team lifted its behind-the-meter data center power forecast from 40GW to 67GW, citing grid bottlenecks that have already sold out GE Vernova's 2030 gas-turbine capacity and more than half of 2031 ^[2]. Capital costs have risen with them: the 10-year UST sits near 5.17%, roughly 1 percentage point higher year-to-date, the marginal price tag of the $4.1 trillion AI-related debt pile JPMorgan expects by 2030 ^[3]. Counter-currents — JPMorgan's "constructive" setup call and Citi's 35% per-task cost decline for open-source models — argue the trade, not the thesis, has reset ^[4][5].

1. Demand-Side Test: Small Models Redirect Compute

• **[NEW] Jefferies (via Wall Street CN):** small-model performance gains and falling cost are pulling enterprise AI from centralized cloud to on-device deployment, threatening the demand side of the hyperscale data center buildout ^[1]. Anchor metric: Swerve Research Technologies' queue-wait "congestion index" shows Anthropic server utilization down 26% from its January-February 2026 peak ^[1]. Caveat: the index is a wait-time proxy, not direct revenue ^[1].
• **[NEW] Hyperscaler capex context:** Street consensus has Meta, Alphabet, Amazon and Microsoft at roughly $990 billion combined 2027 capex, with the report flagging stranded-asset risk if compute demand does not keep pace ^[1].
• **[NEW] Consumer-AI offset:** Meta's Muse consumer product, launched September 8, drove Meta's stock up 21% over subsequent weeks — illustrating the consumer-AI revenue line that the small-model, on-device thesis does not address ^[1].

2. Supply-Side Squeeze: Power Becomes the Binding Constraint

• **[NEW] Goldman Sachs Carbonomics (50-page report, September 27):** raised global behind-the-meter data center power forecast from 40GW to 67GW; gas turbines, reciprocating engines and fuel cells are expected to supply 25% of global and 28% of US data center power by 2030, up from "nearly zero" in 2025 ^[2].
• **[NEW] Gas turbine orderbook signal:** GE Vernova backlog plus reserved capacity stands at 116GW; 2030 capacity fully sold out, 2031 over 50% sold — pushing operators toward fuel cells, reciprocating engines and industrial boilers ^[2].
• **[NEW] Long-dated framing:** small modular reactors (SMRs) are flagged as the only scalable permanent solution post-2030, but construction lead times rule out 2030 delivery ^[2]. Source quality: single sell-side report; quote as Goldman view, not market consensus.

3. Cost of Capital: AI Debt Funds Into a 5.17% 10-Year

• **[ESCALATED] Funding backdrop:** the 10-year UST yield sits near 5.17%, up roughly 1 percentage point year-to-date — the marginal price tag for the next leg of AI bond issuance ^[3].
• **[NEW] JPMorgan framing:** by 2030, AI-related debt issuance by data center and adjacent companies could reach $4.1 trillion, a level already at historical highs; some investors are already flagging that risk exposures inside the existing AI debt stack are being obscured ^[3].
• **[NEW] Equity read-through:** CoreWeave, levered to debt-funded AI buildout, was up about 8% on the week; Oracle, also debt-reliant for its AI expansion, had a tougher week ^[3]. Single-source weekly performance prints; treat as directional.

4. Counter-Currents: Cheaper Inference and a Trade Recovery Call

• **[NEW] JPMorgan strategy (Mislav Matejka et al.):** the recent pullback in global AI trades has reset positioning and improved valuations, supportive of re-engagement, especially in semiconductors; "fundamentals remain constructive" even as tech is unlikely to repeat past outperformance ^[4].
• **[NEW] Citi research:** open-source per-task cost fell 35% week-on-week to $0.80; the discount versus closed models widened from 40% to 60%, while the proprietary model intelligence lead rose from 9 to 12 points ^[5]. Interpretation: closed models are extending both their intelligence lead and relative cost premium even as open-source absolute costs collapse — a separate demand-side dynamic from the on-device shift in ^[1].
• **[NEW] Anthropic chief economist Peter McCrory (Harvard Kennedy School dialogue with Jason Furman, September 24):** attributed the gap between model-capability gains and stagnant total factor productivity to a J-curve diffusion lag; cited US unemployment at 4.1% as the "missing displacement" data point ^[6].

5. Regional Tape and Source Quality

• **Regional tape — China:** Guangdong's reform committee set the "AI + robotics" cluster as a high-tech, high-growth priority cluster, with applications in electronics, new energy vehicles and high-end equipment ^[7]. Aishida deployed 100+ humanoid robots across 100 stores for a September 28-30 commercial trial, with full-shift rosters and public performance scorecards ^[8]. GigaDevice raised its 2026 related-party transaction quota with Changxin (CXMT) by 2.143 billion yuan to 7.854 billion yuan, citing higher DRAM foundry prices; the 2027 January-April quota is set at 4.411 billion yuan, pending shareholder approval ^[9]. miHoYo founder Liu Wei targeted entry into China's top-tier domestic LLM cohort within 2-3 years ^[10]. MetaX established MetaX Integrated Circuit (Chongqing), a wholly owned subsidiary spanning AI foundational and applied software ^[11].
• **Safety and policy tape:** Nvidia released a system designed to prevent AI agents from going out of control, claiming it could have blocked the Hugging Face intrusion ^[12][13]. Anthropic will not appear before the Australian Senate inquiry, per a Guardian relay ^[14]. Zhipu disclosed a compensation plan following the data-upload controversy surrounding its ZCode coding tool ^[15]. Samsung Electronics announced an expanded "Blue Pass" AI subscription; cumulative non-mobile sales are near KRW 1 trillion (~$734 million) ^[16].
• **Source quality control:** the 26% utilization drop rests on a single research note using wait-time as a proxy, not direct revenue ^[1]. The Goldman BTM revision and GE Vernova backlog are from one sell-side report; quote as the Goldman view, not market consensus ^[2]. Weekly equity performance prints for CoreWeave and Oracle come from a single source and should be treated as directional ^[3].