Morgan Stanley lifts AI cluster cost estimates to $49 billion per gigawatt

Morgan Stanley raised its bottom-up cost estimates for next-generation AI clusters, with Nvidia's Vera Rubin-based systems now priced at $49 billion per gigawatt of computing capacity — a nearly 20% increase from prior forecasts that only a handful of the world's most cash-rich technology companies can absorb.

"The cost of building frontier AI infrastructure is rising faster than many investors expected, and it's concentrating the market among a small group of hyperscalers," said Joseph Moore, an analyst at Morgan Stanley, in a research note published Tuesday.

The investment bank's updated estimates show Nvidia's GB200 systems cost about $35 billion per GW, up 16% from prior estimates, while GB300 clusters rose to $39 billion per GW. Those figures align closely with Nvidia's own guidance of $50 billion to $60 billion per GW for Rubin-era installations. The costs encompass not just graphics processors but networking equipment, storage, liquid cooling systems, and power delivery for facilities consuming as much electricity as 700,000 to 1 million US homes.

The rising price tag does not weaken Nvidia's outlook — it may strengthen it. Only companies generating hundreds of billions in annual operating cash flow, such as Microsoft, Amazon, Alphabet, and Meta Platforms, can comfortably finance projects at this scale. Smaller AI companies will increasingly lease capacity from cloud providers or specialists like CoreWeave rather than building their own campuses, shifting even more demand toward the largest operators while reinforcing Nvidia's dominant ecosystem of GPUs, networking hardware, and software.

The $50 Billion Barrier to Entry

OpenAI's Stargate initiative, backed by SoftBank and Oracle, plans to invest $500 billion through 2029 to build up to 10 GW of AI infrastructure. Meta is developing its Hyperion campus with plans to expand from 2 GW to 5 GW, while Microsoft and Google continue building multi-gigawatt data center campuses across the US. These projects require capital commitments that few companies outside the top-tier hyperscalers can match.

Morgan Stanley also noted that power availability — not financing — is increasingly becoming the biggest bottleneck. Utilities face multi-year delays adding new generation and transmission capacity, stretching construction timelines and increasing project costs. McKinsey estimates cumulative AI infrastructure spending could reach trillions of dollars by 2030, while Epoch AI projects multiple frontier AI clusters exceeding 1 GW this year alone.

What Rising Costs Mean for Investors

For Nvidia, more expensive AI factories translate into higher revenue per deployment because its chips, networking products, and software remain at the center of those installations. Suppliers of high-bandwidth memory, power management systems, and liquid cooling equipment also stand to benefit as clusters become larger and more complex.

Nvidia shares, trading at roughly 30x forward earnings, have already priced in much of this infrastructure buildout. The question for investors is whether the market has fully accounted for the concentration risk — that only a handful of companies can sustain this level of spending, and any pullback from Microsoft, Amazon, or Meta could ripple through the entire AI supply chain. Morgan Stanley's revised estimates suggest the AI buildout is not slowing down, but the cost of entry is creating a competitive moat that favors the biggest players and the chipmaker at the center of it all.

This article is for informational purposes only and does not constitute investment advice.