Goldman sees the five biggest hyperscalers adding $400 billion of AI capex in 2027

Crypto Briefing's account of the note lists energy first among the constraints and calls it "arguably the most fundamental constraint" ^[13]. That judgement is the publication's; the report does not attribute a ranking to Goldman. Memory is still on the bank's list, beside energy and labor ^[12]. High-bandwidth memory has been in tight supply, and SK Hynix and Samsung have added capacity without closing the gap, according to the same report ^[15].

The increase itself is $400 billion in a single year, half the estimated 2026 total ^[1]. Measured from the same $800 billion base, the $1.1 trillion consensus implies a $300 billion increase ^[3]. So the gap between Goldman and the Street is $100 billion. That is about 9% of the forecast and a quarter of the increase the bank projects ^[2][3].

The report's out-year numbers do not reconcile. Twelve percent growth on $1.2 trillion gives about $1.34 trillion for 2028, yet the same passage puts total 2028 spending at $2 trillion, a 67% rise ^[4][10]. Either the $2 trillion counts a wider set of spenders or one of the figures is wrong. The $7.6 trillion cumulative estimate favours the smaller number. With $2 trillion in 2028, the last three years of the 2026-2031 window would share $3.6 trillion, about $1.2 trillion a year and 40% below the 2028 level ^[5]. With $1.34 trillion, they share roughly $4.26 trillion, or about $1.42 trillion a year, a path of slowing growth that matches the deceleration Goldman describes ^[6].

Crypto Briefing argues the spending has a floor, because a hyperscaler that under-invests risks losing cloud customers to rivals ^[17]. By that account the floor holds even if returns take longer to arrive. So the five keep committing capex ahead of the revenue it is meant to earn. The same report says a move toward debt financing is bringing interest-rate sensitivity into what has historically been a cash-rich corner of the market ^[16].

The evidence supports three outcomes. Power and memory hold 2027 near the $1.1 trillion consensus ^[3]. Goldman's $1.2 trillion base case lands ^[1]. Or the $1.4 trillion upside does ^[5]. The energy thesis and the above-consensus forecast pull against each other. A power constraint that binds slows the spending, and new generation takes years to bring online ^[12][14]. I'd expect the forecast to hold up better than the bottleneck call for 2027. The upward revisions rest on reported second- and third-quarter capex, and the energy ranking rests on the publisher's framing ^[8][13]. That view is wrong if hyperscalers start deferring 2027 projects because they cannot get grid power. Deferrals would pull the year toward $1.1 trillion and make power supply the figure that sets capex ^[3].