Wall Street isn’t worried about capital expenditures! Morgan Stanley: Next year, the five major cloud providers will spend $1.2 trillion in capital expenditures; by 2028, $1.4 trillion, with computing power increasing fourfold.

Wall Street isn’t worried about capital expenditures! Morgan Stanley: Next year, the five major cloud providers will spend $1.2 trillion in capital expenditures; by 2028, $1.4 trillion, with computing power increasing fourfold.

Capital expenditure figures are getting larger and larger, but Morgan Stanley believes this is not yet a problem.

According to Chase Wind Trading Desk, Morgan Stanley analyst Brian Nowak's team released the latest research report on July 12, raising the 2027/2028 capital expenditure forecasts for the five hyperscale cloud providers (Meta, Amazon, Microsoft, Google, SpaceX) by 9%/10% overall, reaching approximately $1.2 trillion and $1.4 trillion respectively.

Why is capital expenditure still rising? Dual drivers: cost inflation + shortage of computing power

The direct trigger for Brian Nowak's upward revision is rising costs.

GPU-related costs have increased by about 20%, mainly from two aspects:

Increase in rack internal costs: Memory prices have surged. In Blackwell (GB300, GB200) racks, memory's cost proportion has risen from the low single digits to the high single digits; in Vera Rubin racks, it goes from the high single digits up to about 25%.

Increase in rack external costs: Extended delivery times for electrical/mechanical equipment, scarcity of building materials, shortage of skilled labor, and the shift to “behind the meter power” solutions all collectively push up the cost of data center construction. Morgan Stanley currently estimates external rack costs for GB200, TPU, and Trainium at $11–13 million per megawatt, while the next-generation GB300, Vera Rubin, and Rubin Ultra rise further, to about $16–19 million per megawatt.

For each GW of data center construction cost: GB200 is about $3.5 billion, GB300 about $3.9 billion, Vera Rubin about $4.9 billion; for custom ASIC, TPUv7 about $2.7 billion, Trainium3 about $2 billion.

Besides the cost factors, Nowak also points out that the cycle from groundbreaking to opening of data centers has been extended to up to three years. Meanwhile, opposition to data center construction in society is rising, and political uncertainty ahead of the 2028 presidential election is prompting cloud providers to start building earlier, emphasizing job creation. This further increases inflationary pressure and accelerates front-loaded capital expenditure.

Computing power to quadruple: from 30GW to 120GW

The ultimate goal of capital expenditure is computing power.

Nowak forecasts that the available computing power of the five hyperscale cloud providers will grow from about 30GW in 2025 to about 116.6GW in 2028, nearly quadrupling.

Breakdown by company:

  • AWS: 35GW in 2028, largest computing power scale
  • Google: 31.6GW in 2028, largest increase in computing power (9GW/11GW added in 2027/2028 respectively)
  • Microsoft: 20.3GW in 2028
  • Meta: 21.2GW in 2028, a sharp leap from about 3.5GW at the end of 2025; of the additional computing power in 2026/2027, 55%/90% is self-built, 45%/10% from third-party

Meta: Five “call options” not yet priced by the market

Morgan Stanley's Nowak maintains Meta as the top pick with a target price of $775, representing about 15% upside and an Overweight rating.

Nowak believes that the market is currently “punishing” Meta for its high capital expenditure without giving proper valuation for its potential revenues. The report lists five earnings drivers not yet priced by the market, which, if realized, could add about $10 on top of baseline earnings per share of about $33, meaning Meta’s current stock price corresponds to a 2028 P/E of only about 15 times.

These five “call options” include:

  1. Neocloud monetization: about $2.97 per share upside
  2. Meta AI: about $2.89
  3. Search: about $1.91
  4. API revenue: about $1.90
  5. Subscription opportunity: about $0.94

API revenue: a new business worth close attention

Meta's recently launched Muse Spark 1.1 model API is priced 30%–85% lower than competitors. Morgan Stanley calculates that every 100MW of computing power allocated to the API business can generate about $8.588 billion in revenue, corresponding to about $1.91 earnings per share, about 6% of 2028 EPS.

Under the baseline scenario, Morgan Stanley assumes that 25% of Meta’s 15 million advertisers (about 4 million) pay roughly $200 a month for its products, which can similarly produce about $8 billion in revenue and about $2 in earnings per share.

Morgan Stanley points out that Meta only needs to allocate about 100MW of GB300 computing power to support this API business, while its expected computing power base at the end of 2028 is about 21GW, leaving ample room for expansion.

Valuation adjustment: Morgan Stanley raises Meta’s 2027/2028 capital expenditure by 29%/22% to $225 billion/$250 billion, respectively, and includes $40 billion in new debt issuance. Higher depreciation drags down 2027/2028 EPS by 3%/7% to about $32.99/$33.41. But Morgan Stanley also slightly raises 2028 revenue forecast (+1%) and rolls valuation to mid-year, using 23x P/E on average 2027/2028 EPS to maintain the $775 price target.

Amazon: AWS growth underestimated

Nowak raises Amazon’s 2027/2028 capex forecasts by 15%/29% to $308 billion/$318 billion respectively.

Unlike Meta, Amazon’s EPS forecasts are actually revised upward. The reason is a sharp upward revision in AWS revenue forecasts: AWS revenue in 2027/2028 is raised by 3%/7%, with YoY growth of 40%/36%, reaching $243.5 billion/$331.6 billion. Higher depreciation is offset by stronger AWS revenue, so 2027/2028 EPS is raised by 2%/3% to about $11.53/$15.05.

Nowak believes the market's current AWS revenue forecasts are too conservative. The report says even Morgan Stanley’s own 35%/40% AWS growth forecasts for 2026/2027 imply incremental revenue per watt of only $7/$9, “which can be considered conservative.”

Additionally, Nowak expects AWS backlog orders to rise by $110 billion quarter-on-quarter in Q2, reaching about $475 billion, mainly from large orders in private labs. This will enhance market confidence in sustained multi-year growth.

The target price of $330 is based on average 2027/2028 EPS of about $13 multiplied by a P/E of about 25x, corresponding to about 1.4x PEG, and is about 30% discounted compared with peers.

Google: Leading in full-stack AI, but faces short-term computing power constraints

Nowak assigns Google a target price of $415, about 16% upside, and an overweight rating.

Nowak expects Google Cloud’s growth rate for Q2/full year 2026 to reach 77%/78%, and Search’s Q2/full year growth rates to be 17.5%/16%, with 2027/2028 growth rates of 11%/8%.

But Nowak specifically warns of a tactical risk: Google is currently facing computing power constraints (as evidenced by its latest computing power leasing deal with SpaceX). Capacity constraints may drag down short-term revenue growth or product launches—a risk much smaller for Meta and Amazon.

In addition, the launch of Gemini 3.5 was delayed compared to Google’s I/O conference schedule. Morgan Stanley will be watching further productization of Gemini models later this year.

 

 

 

~~~~~~~~~~~~~~~~~~~~~~~~

The above exciting content is from Chase Wind Trading Desk.

For more detailed interpretation, including real-time analysis, frontline research, etc., please join [Chase Wind Trading Desk ▪ Annual Membership]

Risk Warning and DisclaimerThe market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions in this article suit their particular circumstances. Investing accordingly is at your own risk.