Can hyperscalers clear $1.42T / $11.6B per GW for 15% ROIC?

Opening

Goldman’s clearability math — ~$1.42T cumulative revenue by 2030 / ~$11.6B per GW for 15% ROIC, with 59% already in cloud backlogs — has the better of pure “capex is doomed” overnight takes, but the same thread’s sensitivity table ($920B/yr even at 0 ROIC) and rate path keep the break conditions live. @zerohedge overnight TL/DR of a Goldman spreadsheet on justifying ~$1.7T hyperscaler capex (post; earlier link post); counter in the same note lists the ifs (compute stays scarce, GPU/token prices don’t deflate, 5-year useful life, labs keep paying, $3T off-balance-sheet needs its own return). A paired ZH post says Goldman’s sensitivity table shows that if hyperscaler ROIC on capex is 0, they still need $920BN every year for depreciation and running costs (post). Tier-1 replies reframes: @Klaudnin3 maps the same math to an Apple “rent the GW” bull vs shovel owners; @jukan05 treats elevated rates as a major obstacle to AI investment even while quoting a Fed Jul-2000 line that tech capex can hold up vs borrowing costs; @TheStalwart asks why META would pay more interest to stay nominally asset-light. Deep reply tails on Stalwart/jukan unread (top-replies). Numbers are Goldman-via-ZH / sell-side-via-X — not a primary GS PDF in this capture.

Sides

FOR — revenue hurdle is clearable on backlog + claimed paybacks

• @zerohedge [pseudo] relaying Goldman — $1.42T rev over 2028–30 / $11.6BN per GW clearable for 15% ROIC; 59% already in cloud backlogs; managements claim 1–3 year paybacks
• @Klaudnin3 [pseudo] — same $1.42T / $11.6B/GW math; bull case is Apple not buying the shovel (capex <$15B/yr, Gemini license ~$1B/yr cited) while builders need the revenue path

AGAINST — sensitivity + rates + off-BS break the easy clear

• @zerohedge [pseudo] — if ROIC on capex is 0, still $920BN/yr depreciation + running costs; counter ifs on scarcity, token/GPU deflation, 5-year life, lab credit, $3T off-balance-sheet return
• @jukan05 [pseudo] — “high interest rates will be a major obstacle to AI investment”; Fed shouldn’t hike further (own view); pairs Ackman rates→inflation thread with a Fed Jul-2000 quote that tech investment can hold vs borrowing costs
• @JonSteinsson [named] (quoted by Stalwart) — Van Nieuwerburgh BPEA: AI buildout $10.3T over next 8 years; $3.7T annual revenues by 2032 for 10% return (~9% of GDP)

THIRD — wrong question: ownership / financing form, not whether GW get built

• @Klaudnin3 [pseudo] — scoreboard YTD (cashtags in-post): AAPL +23.9% vs META/GOOGL/AMZN/MSFT; “market stopped paying for the biggest shovel and started paying for the biggest toll booth”
• @TheStalwart [named] — datacenter-finance writeup is clear; asks for an EMH-compatible reason META pays more interest just to remain nominally “asset-light”

Room: Goldman math arrives via ZH screenshots/links, not a primary deck; Klaudnin3 and Stalwart are named book/desk voices arguing structure (rent vs own, SPV interest). Reply volume is not evidence.

Quotes

TL/DR: Goldman thinks $1.42 trillion in revenue over 2028-30, or $11.6BN per GW - is clearable (required for 15% ROIC), with 59% of it already sitting in cloud backlogs and management teams claiming paybacks of anywhere from one to three years

And here is the counter: the hurdle is clearable if compute stays scarce, if GPU and token prices don't deflate, if a 5-year useful life is real, if a handful of AI labs keep paying their bills, and if the $3 trillion off-balance sheet iceberg doesn't need a return of its own. — zerohedge (@zerohedge)

This could be a problem: Goldman sensitivity table shows that in a worst case scenario where hyperscaler ROIC on capex is 0 (token costs collapse, token demand goes to open models, etc) they still need to spend $920BN every year just to cover depreciation and running costs — zerohedge (@zerohedge)

goldman just ran the math on the $1.2T buildout and the answer rhymes with one ticker: $AAPL.

six hyperscalers now need ~$1.42T in cumulative revenue by 2030 just to earn 15% on their 2026-27 capex. that is $11.6B of annual revenue per gigawatt. every year.

apple's answer: don't build the gigawatt. rent it. — Bottleneck Capital (@Klaudnin3)

This is a very good and clear writeup on datacenter finance. Can someone explain to me in an EMH-compatible way why it’s worth $META paying more in interest just in order to nominally remain an “asset-light” business? — Joe Weisenthal (@TheStalwart)

No matter how I look at it, I think high interest rates will be a major obstacle to AI investment. The Fed shouldn’t raise rates. — Jukan (@jukan05)

New BPEA paper by @SVNieuwerburgh argues that AI buildout will cost $10.3T over next 8 years. Furthermore, to get a 10% return on this, those investing will need annual revenues of $3.7T by 2032! (Approx 9% of GDP!) — Jón Steinsson (@JonSteinsson)

What would settle it

Hyperscaler / neocloud FY27–28 guidance that keeps or cuts the multi-year capex path Klaudnin3 flags (~$1.2T 2027 number in-post) with explicit AI revenue or ROIC language Knowable — Mag7 / Oracle / neocloud quarterly prints — through 2026-12-15 earliest checkpoint

Primary Goldman (or peer) note text stating the $1.42T / $11.6B-per-GW / 59% backlog / $920B sensitivity assumptions without ZH mediation Knowable — GS client PDF or confirmed excerpt — through 2026-11-30

ORCL / other DC-linked credit spreads and CDS vs the “neutral credit” Barclays read after Jupiter FM Knowable — market prints already in motion (ZH: ORCL CDS 237bps record Sep 25) — ongoing through 2026-10-31

Fed funds / 10y path vs hyperscaler announced project FID / cancellations over the next two quarters Knowable — FOMC + company capex commentary — through 2027-03-31

Positioning

AI capex durability — Clearability-at-15%-ROIC with backlog cover is the durable-spend case; the $920B opex floor at 0 ROIC and rate-as-obstacle camp is the break path if token prices deflate into open models or credit rationing hits off-balance-sheet shells.