Morgan Stanley: AI data centers now cost $49B a gigawatt

The price tag for the physical backbone of AI just got uglier. Morgan Stanley now pegs an all-in US data center built for the Vera Rubin era at roughly $49 billion per gigawatt of capacity, according to a July 2026 report led by analyst Brian Nowak. That lands $20 billion above where it sat three years earlier.

Compare it to the last generation. Facilities built around Nvidia's GB200 systems came in near $35 billion per gigawatt, the report says. The culprits are familiar: GPU and memory prices are climbing, delivery times are stretching, materials are scarce and skilled labor is hard to find.

The pain isn't confined to the racks. Off-rack costs — everything outside the computing racks — climbed to $16–19 million per megawatt, up from $11–13 million in the bank's analysis. Morgan Stanley also points to a shift toward behind-the-meter power, where electricity is generated on site or nearby instead of drawn from the public grid, as another cost driver.

The biggest buyers aren't blinking. Morgan Stanley raised its capital expenditure forecasts for the five largest hyperscale cloud providers, a group that includes Meta, Amazon and Microsoft, by 9–10%. It now projects combined capex of $1.2 trillion for 2027 and $1.4 trillion for 2028.

Power is the constraint that keeps resurfacing. In notes published between September and November 2026, the bank estimated US data center electricity demand could reach 97–116 GW by 2028. Without mitigation it sees a shortfall of up to 70 GW; even after accounting for fixes, a net gap of 33–43 GW remains.

Crypto watchers get a cameo here. Former Bitcoin mining sites appear among the mitigation strategies, since they arrive with grid connections and power infrastructure already in place — exactly what AI builders are scrambling to secure. Morgan Stanley treats them as a marginal piece of the puzzle rather than a fix. The bottom line: if construction keeps outrunning expectations, AI revenue has to grow faster to justify the bill, while GPU, memory, electrical equipment and construction suppliers are the ones capturing the higher per-gigawatt spend.