Analysis
Morgan Stanley's North America Insight team, led by analyst Nicholas Lentini, estimates in [an Oct. 15 report cited by Fortune](https://fortune.com/2026/10/19/nvidia-concentration-hyperscaler-1-6-trillion-92-percent-exposure/) that Nvidia's 2027 revenue is 91.7% exposed to AI infrastructure demand -- the figure Fortune's headline rounds to 92% -- more concentrated than Vertiv (90%) or Broadcom (81%), the two other chip-adjacent companies the bank flagged.
The report's broader numbers show why that concentration matters:
- 2027 AI infrastructure spend: $1.58 trillion, up about 60% from roughly $980 billion in 2026, rising further to $1.8 trillion in 2028.
- "Synthetic capex": about $195 billion of the 2027 total is data-center lease commitments the report says can be unwound faster than a chip order once placed.
- Unstarted commitments: hyperscalers have already disclosed more than $1.1 trillion in lease commitments that haven't begun.
“- "Synthetic capex": about $195 billion of the 2027 total is data-center lease commitments the report says can be unwound faster than a chip order once placed.”
Where the growth is actually concentrated
Morgan Stanley's most pointed number is customer concentration, not market size: SpaceX and Amazon Web Services together account for about 54% of Nvidia's hyperscaler revenue in 2027 and 72% of its revenue growth. Nvidia's own share of other hyperscalers' spending, meanwhile, is projected to slip from 18% in 2026 to 16% in 2027, even as AI capex itself starts cannibalizing the revenue it was supposed to fund -- a dynamic Pulse flagged in a separate analysis this week. Nvidia trades at 15 times fiscal 2028 earnings, a multiple the bank's own model implies assumes the concentration holds.
The leverage math underneath the growth math
The report also runs a stress scenario on who gets hurt if spending slows. Interest coverage across Amazon, Alphabet, Meta and Microsoft falls from about 31x to 12x in that scenario -- still comfortable. For Oracle and SpaceX, coverage falls from 4x to 1.8x, a far thinner cushion, after hyperscalers have already sold roughly $255 billion in unsecured bonds this year, about 30% of estimated 2026 AI spending. Morgan Stanley also models a 32-gigawatt U.S. data center power shortfall through 2028, roughly a third of the 97 gigawatts the buildout needs -- a physical constraint no amount of capital can route around on its own.
What the headline misses
Morgan Stanley isn't calling for an AI capex collapse: the bank's base case still expects 25% to 50% returns on invested capital, and notes Wall Street has historically underestimated compute cycles by 38% on average across PCs, the internet, mobile and cloud. The risk the report surfaces isn't that AI spending stops -- it's that Nvidia's growth is now riding on a small number of highly leveraged counterparties, and a grid buildout that is behind schedule, rather than on broad-based hyperscaler demand.
