AI Capex to Exceed $1.2 Trillion in 2027 and 2028?

Is the growing cost of Nvidia-based compute a business moat or a liability? And in either case, which contestants benefit or suffer? 

A new Morgan Stanley analysis suggests the cost of Nvidia's GB200 systems now cost about $35 billion per gigawatt (GW) of computing capacity, up 16 percent from prior estimates. 

GB300 clusters rise to $39 billion per GW, while Vera Rubin-based systems jump nearly 20% to $49 billion per GW.

Those estimates include networking equipment, storage, liquid cooling systems, electrical infrastructure, and power.

A single GW can power roughly 700,000 to 1 million U.S. homes. 

So are high costs a business moat or a sign of dangerous costs, or perhaps both?

Morgan Stanley also projected that the combined capital expenditure of the five largest AI-driven firms (Microsoft, Google, Amazon, Meta, and SpaceX) will reach approximately $1.2 trillion and $1.4 trillion in 2027 and 2028, respectively. 

By 2028, available computing capacity is expected to grow from around 30 GW in 2025 to nearly 120 GW, a fourfold increase.

Morgan Stanley expects the combined available computing capacity  of the five major hyperscalers to reach nearly 120 GW by 2028, a fourfold increase from approximately 30 GW in 2025. 

AWS will have the largest capacity in 2028 at 35 GW, followed by Google at 31 GW. Meta’s capacity is projected to grow from approximately 3.5 GW at the end of 2025 to 14 GW in 2027 and 21 GW in 2028.

But the AI compute arms race is shifting from "how much you build" to "how much you sell,” Morgan Stanley suggests. Who can convert compute into monetizable revenue streams?