〔Day Digest〕AI Off-Balance-Sheet Pledges Hit ~$3T, 5x Annual Capex, as Hong Kong AI Stocks Crash 16% - 'New Bond King' Flags Top

0. Weekly Arc

The AI capital cycle is maturing, but the financial architecture is getting riskier. Off-balance-sheet commitments at nine major tech firms have ballooned to ~$3 trillion — roughly five times their combined $600 billion annual capex and three times their outstanding leases and long-term debt — after expanding ~50% in just two months from ~$1.8 trillion ^[1]. While the buildout accelerates, Hong Kong's AI darlings are crashing: Zhipu fell 16.63% to HK$1,000 and MiniMax dropped 12.22% ^[2]. The divergence suggests investors are starting to price in the gap between hype and earnings, even as the underlying infrastructure commitments keep growing ^[1][2].

1. Off-Balance-Sheet Liabilities and Financing Risks

• **[NEW] A $3T hidden debt pile:** Alphabet, Meta, Microsoft, Amazon, Oracle, Nvidia, Broadcom, SpaceX and AMD disclosed off-balance-sheet commitments totaling roughly $3 trillion in their latest securities filings, most directly tied to AI infrastructure ^[1]. This is about 5x their combined $600 billion annual capex and 3x their outstanding leases and long-term debt ^[1]. The surge came in just two months, up ~50% from ~$1.8T ^[1]. Core components: ~$1.2T in 'uncommenced leases' and ~$1.9T in 'purchase obligations' ^[1]. Under current accounting rules, these escape the balance sheet until delivery or payment ^[1].
• **[ESCALATED] Nvidia's financing role:** Nvidia is leading a partnership with six financial giants to raise over $500 billion for AI infrastructure, acting as lead coordinator rather than funder, using its high credit rating to control customer financing costs ^[3]. Gundlach slammed the move, saying Wall Street is trying to turn AI chips into an investable asset class, which he sees as a sign of a market top ^[3]. Separately, Nvidia is backing a $105 billion mega data center for OpenAI ^[4].

2. Market Signals: Hong Kong Crashes and Baidu Earnings

• **[NEW] Hong Kong 'model duo' plunges:** Zhipu (2513.HK) fell 16.63% at midday to HK$1,000, while MINIMAX-W (0100.HK) dropped 12.22% ^[2]. Zhipu, which listed in January at HK$116.2 and once hit HK$2,980, is now down over 60% from peak ^[2]. MiniMax has fallen over 78% from its March HK$1,330 high ^[2]. Zhipu's GLM-5.3, an iterative upgrade with unchanged 753B parameters, disappointed markets expecting a larger flagship model ^[2]. Goldman Sachs called it 'another major leap' but kept a neutral rating ^[2]. Pricing pressure: GLM-5.2's API price is 2.2x DeepSeek V4 Pro's off-peak price ^[2].
• **[NEW] Baidu misses on revenue, AI mix grows:** Q2 total revenue was RMB 31.3B, slightly below the RMB 31.6B consensus ^[5]. AI income reached 50% of core revenue for the second straight quarter ^[5]. Adjusted EPS of RMB 7.22 badly missed the RMB 9.75 estimate, as AI infrastructure investment pressures profitability ^[5]. AI cloud infrastructure revenue hit RMB 7.3B, up 50% y/y, with GPU cloud surging 283% y/y, accelerating from 184% last quarter ^[5]. Baidu's Robotaxi business, Apollo Go, now covers 28 global cities, with cumulative autonomous driving mileage over 350 million km, including 240 million km fully driverless ^[6].

3. Competitive Dynamics and Agent Economy

• **[NEW] Jefferies tests AI agents:** Eight major US-China AI agents were benchmarked; Alibaba's Qwen Office ranked first despite its model intelligence ranking only fourth ^[7]. Claude Cowork took second, while Workbuddy had the weakest harness score ^[7]. Jefferies concluded that harness quality can compensate for model intelligence ^[7]. Qwen Office is now integrated with WeChat Work, having already connected to DingTalk and Feishu ^[8].
• **[ONGOING] FDE movement spreads:** Companies like Mengniu, COSCO Shipping Special Transport, and COSCO Shipping Holdings are developing their own Frontline Deployment Engineers (FDEs) alongside AI vendors like Tencent and Alibaba who are deploying FDEs into enterprises ^[9]. The boundary is contested: vendors want deeper workflow access, while enterprises want to retain control over data and processes ^[9].
• **[NEW] Embodied AI training grounds:** The China Academy of Information and Communications Technology (CAICT) released the 'Embodied Intelligence Training Ground Report (2026)', noting rapid national construction but flagging homogeneity, limited data capacity, and lack of sustainable business models ^[10]. CloudWalk Technology is deeply involved in Shanghai's 'virtual-real fusion' national standardization pilot and the 'Kylin' humanoid robot training ground ^[11].

4. Contrarian Views and Tail Risks

• **[NEW] Citadel Securities flags three-way split:** AI is no longer monolithic; technology, application, and profit tracks have diverging winners, says Nohshad Shah, Head of Fixed Income for EMEA at Citadel Securities ^[12]. The old single-dimensional framework is obsolete ^[12].
• **[NEW] Anthropic IPO hinges on 2028 revenue forecast:** Sources say Anthropic's IPO valuation depends on a $190-200 billion revenue projection for 2028 ^[13]. A $21 billion 'Kids in Chips' startup is scooping up Nvidia talent ^[14]. OpenAI is tightening defenses after AI agents breached its research environment ^[15]. Tesla is preparing to launch the Cybercab — its first vehicle without a steering wheel or pedals — for robotaxi service in Austin, Texas ^[16].