Morgan Stanley는 미국 데이터센터 비용이 기가와트당 490억 달러까지 상승할 것으로 내다봤다.

Building the physical backbone of artificial intelligence keeps getting more expensive. Morgan Stanley now estimates that an all-in US data center built for the Vera Rubin era will cost approximately $49 billion per gigawatt of capacity.

That figure reflects a projected $20 billion increase over three years.

The estimate comes from a July 2026 report led by Morgan Stanley analyst Brian Nowak. It lands as the largest cloud companies keep committing staggering sums to AI infrastructure, and it suggests each dollar of that spending buys less capacity than it used to.

**What is driving the bill higher**

Facilities built around Nvidia's previous GB200 systems came in at about $35 billion per gigawatt, according to the report. The move to Vera Rubin-era hardware pushes that number meaningfully higher.

Morgan Stanley points to a familiar list of culprits: GPU and memory prices are rising, delivery times are stretching out, materials are in short supply, and skilled labor is hard to find.

The pressure is not confined to the servers themselves. Off-rack costs, meaning spending on everything outside the computing racks, rose from $11–13 million per megawatt to $16–19 million per megawatt in the bank's analysis.

The report also flags a shift toward behind-the-meter power as a contributor to higher costs. Behind-the-meter setups generate electricity on site or nearby rather than drawing it from the public grid.

**Hyperscalers keep spending anyway**

Rising costs have not scared off the biggest buyers. 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%.

The bank now projects combined capex of $1.2 trillion for 2027 and $1.4 trillion for 2028.

Electricity is the other constraint that keeps showing up. In notes published between September and November 2026, Morgan Stanley estimated that US data center power demand could reach 97–116 GW by 2028.

Without mitigation, the bank sees a potential shortfall of up to 70 GW. After accounting for mitigation measures, it estimates a net shortfall of 33–43 GW.

**Where Bitcoin miners fit in**

One detail in Morgan Stanley's power analysis will catch the eye of crypto watchers. Former Bitcoin mining sites are referenced as part of the mitigation strategies for closing the electricity gap.

The bank treats this as a marginal piece of the broader picture rather than a central solution. Still, it reflects a logic that has been circulating in the industry: mining facilities already come with grid connections and power infrastructure, the very things AI builders are struggling to secure.

**What this means for investors and builders**

The most direct implication is for the hyperscalers' return math. If construction costs keep outpacing expectations, the payoff from AI services has to grow faster to justify the spending.

Suppliers sit on the other side of that equation. Companies selling GPUs, memory, electrical equipment and construction services are the ones capturing those rising per-gigawatt costs.

Power is the variable to watch most closely. A net shortfall of 33–43 GW even after mitigation means energy access, not just chip supply, may decide which projects actually get finished on schedule.