〔Day Digest〕Goldman Sizes the $1.73T AI Capex Test: 15% ROIC Demands $1.42T of 2028-30 Revenue, While ~$3T Off-Balance-Sheet Items and a $4.14T Phase 3 Lurk Outside the Math - Returns Now Drive the Debate

0. Weekly Arc

The research board pivots from "how much" AI spend to "how much back": Goldman's hyperscaler capex frame demands $1.42T of 2028-30 revenue to justify the $1.73T 2026-27 spend at 15% ROIC ^[1], a Caixin column by Xu Xiaoqing argues AI investment is already weakening and US Treasury yields may turn down ^[2], and a CEPR/VoxEU column re-raises financial repression as a persistent debt-reduction tool ^[3]. On China, brokers hold the medium-term rebound call with the Shanghai Composite at 3850-4000 ^[4], and a quantum-classical fusion report marks accelerated deployment ^[5]. Nomi Prins keeps gold constructive despite any Fed hike ^[6].

1. The AI Capex Math (Goldman)

• **[NEW] Eric Sheridan, head of Goldman's internet & tech team:** reframes the AI capex debate as a returns question, not a spend question ^[1].
• **[NEW] The numbers:** six hyperscalers face roughly $1.73T of AI compute capex in 2026-2027; a 15% ROIC on that base requires about $1.42T of cumulative 2028-2030 revenue ^[1].
• **[NEW] The cushion:** combined AWS, Azure and Google Cloud backlog hit about $1.69T as of Q2 2026, up 152% y/y, providing a starting "book" against the revenue target ^[1].
• **[NEW] What is NOT in the denominator:** roughly $3T of off-balance-sheet commitments, ongoing GPU and compute pricing pressure, and the larger "Phase 3" cycle of about $4.14T for 2028-2030, which Goldman explicitly defers ^[1].
• **Source quality control:** the Phase 2/3 framework and the $1.42T/$4.14T splits are single-broker output; the backlog figure is reported as a Q2 2026 read, not yet audited to a calendar ^[1].

2. Rates, the Fed, and the Long Shadow of Repression

• **[ONGOING] Nomi Prins:** a Fed rate hike does not change gold's long-term outlook ^[6]. Single-source, op-ed-style framing; treat as view, not data.
• **[NEW] Xu Xiaoqing (Caixin weekly preview):** AI investment is weakening and US Treasury yields may turn down ^[2]. Headline is forward-looking, not a confirmed move.
• **[ONGOING] CEPR / VoxEU:** new quantity-based measures of financial repression across 17 advanced economies over a century show repression as a persistent post-WWII debt-reduction tool, rising again since the Global Financial Crisis ^[3].
• **[ONGOING] Caixin essay:** asks whether an "Engels pause" - the displacement-and-redistribution cycle that followed 19th-century industrial booms - is replaying under AI, with capital, labor and distribution all in motion ^[7]. Frame piece, no fresh data.

3. China A-Shares: Rebound Call Intact

• **[NEW] Weekly A-share report #111:** Shanghai Composite held the 3850-4000 range; STAR50's prior rebound was nearly 13%, this week's pullback framed as a "strong adjustment" rather than a break ^[4]. Hang Seng Tech shows clear bottom features ^[4].
• **[NEW] Allocation view:** keep mid-line positions through the holiday, add on post-holiday dips; favor previously hard-hit sectors - ChiNext/STAR, Hang Seng Tech, non-bank financials, media, computing - with balanced sizing ^[4].
• **[NEW] Sentiment reads:** Shanghai-Shenzhen turnover firmed, most index futures in contango discount, margin balance ticked up while margin-buying share fell, stock ETFs net inflow ^[4].
• **Source quality control:** broker-house strategy note, not a primary dataset; ranges and "bottom features" are framed in the report's own vocabulary ^[4].

4. Quantum-Classical Fusion: From Theory to Deployment

• **[NEW] Quantum-tech series report (3):** frames "quantum-classical fusion" (量超融合) as a hybrid model that uses classical compute to sidestep NISQ-era limits via VQA- and SQD-style algorithms, while quantum handles parallel and linear-algebra tasks ^[5]. The report positions the model as entering an accelerated deployment phase ^[5].
• **Source quality control:** sell-side thematic report; the "accelerated deployment" claim is qualitative, no shipped-unit or revenue numbers attached ^[5].

5. What Would Falsify It

• Goldman's frame breaks if 2028-2030 revenue does not approach the $1.42T cumulative target, or if backlog conversion slows from the 152% y/y pace, or if Phase 3 capex is forced into the ROIC math ^[1].
• Xu Xiaoqing's yield view breaks if AI capex does not actually roll over into the softer path the column implies ^[2].
• Prins's gold call is unverified beyond a single relay and should be read as one analyst's view, not consensus ^[6].
• The A-share rebound call breaks if the Shanghai Composite loses the 3850 floor or if STAR50's pullback deepens past the "strong adjustment" zone ^[4].