Despite capital expenditure figures growing larger, Morgan Stanley does not view this as a concern.
A recent research report from a team led by Morgan Stanley analyst Brian Nowak, dated July 12, has raised the capital expenditure forecasts for the five leading hyperscale cloud providers (Meta, Amazon, Microsoft, Google, and SpaceX) by approximately 9% and 10% for 2027 and 2028, respectively. The revised projections stand at roughly $1.2 trillion and $1.4 trillion.
Drivers Behind Rising Capital Expenditure
The immediate trigger for analyst Brian Nowak's upward revision is cost inflation.
GPU-related costs have increased by about 20%, driven by two primary factors. Within the rack, memory prices have surged significantly. In Blackwell (GB300, GB200) racks, the memory cost share has risen from a low single-digit percentage to a high single-digit percentage. For Vera Rubin racks, the memory share has jumped from a high single-digit percentage to approximately 25%.
Outside the rack, costs are rising due to extended lead times for electrical/mechanical equipment, shortages of construction materials and skilled labor, and a shift towards "behind the meter power" solutions, collectively pushing up data center construction costs. Morgan Stanley's current estimates for rack-level costs for GB200, TPU, and Trainium are approximately $11 to $13 million per megawatt. The next generation, including GB300, Vera Rubin, and Rubin Ultra, is expected to see these costs rise further to about $16 to $19 million per megawatt.
Specific data center construction costs per gigawatt are estimated as follows: approximately $35 billion for GB200, $39 billion for GB300, and $49 billion for Vera Rubin. For custom ASICs, TPUv7 is estimated at around $27 billion, and Trainium3 at about $20 billion.
Beyond cost factors, Nowak points out that the timeline from groundbreaking to data center opening has stretched to up to three years. Concurrently, increasing public opposition to data center construction and political uncertainty ahead of the 2028 presidential election are prompting cloud providers to accelerate construction starts to emphasize job creation. This dynamic is exacerbating inflationary pressures and accelerating the front-loading of capital expenditures.
Computing Power Set to Quadruple
The ultimate goal of capital expenditure is computing capacity.
Nowak forecasts that the available computing power of the five major hyperscale cloud providers will grow from approximately 30 gigawatts in 2025 to about 116.6 gigawatts by 2028, representing a nearly fourfold increase.
A breakdown by company shows that AWS is projected to reach 35 GW by 2028, making it the largest in scale. Google is expected to reach 31.6 GW by 2028, adding the most new capacity (9 GW in 2027 and 11 GW in 2028). Microsoft is forecast to reach 20.3 GW by 2028. Meta is projected to surge to 21.2 GW by 2028 from around 3.5 GW at the end of 2025, with 55% and 90% of its 2026/2027 new capacity additions coming from self-builds, and 45% and 10% from third-party sources, respectively.
Meta's Undervalued Potential
Morgan Stanley's Nowak maintains Meta as a top pick with a $775 price target, implying roughly 15% upside, and an Overweight rating.
Nowak believes the market is currently "penalizing" Meta for its high capital expenditures without adequately valuing its potential revenue streams. The report lists five profit drivers not yet priced in by the market. If all are realized, they could contribute an additional approximately $10 in earnings per share on top of a base EPS of around $33, implying Meta's current stock price corresponds to an actual 2028 price-to-earnings ratio of only about 15 times.
These five potential catalysts include Neocloud monetization (about $2.97 EPS upside), Meta AI (about $2.89), Search (about $1.91), API revenue (about $1.90), and subscription opportunities (about $0.94).
Focus on API Revenue
The recent API pricing for Meta's Muse Spark 1.1 model is 30% to 85% lower than competitors. Morgan Stanley calculates that every 100 megawatts of computing power allocated to the API business could generate approximately $8.588 billion in revenue, equating to about $1.91 in EPS, or roughly 6% of the projected 2028 EPS.
Under a base case scenario, assuming 25% of Meta's 15 million advertisers (about 4 million) pay around $200 monthly for its products, this could also generate approximately $8 billion in revenue and about $2 in EPS.
Morgan Stanley notes that Meta would only need to allocate about 100 megawatts of GB300 computing power to support this API business, while it is projected to have a computing power base of approximately 21 gigawatts by the end of 2028, leaving ample room for expansion.
Valuation adjustments incorporate a 29% and 22% increase in Meta's 2027/2028 capital expenditure forecasts to $225 billion and $250 billion, respectively, along with an additional $40 billion in debt issuance. Higher depreciation reduces the 2027/2028 EPS forecasts by 3% and 7% to approximately $32.99 and $33.41. However, Morgan Stanley also slightly raised its 2028 revenue forecast by 1% and rolled the valuation forward to mid-year, applying a 23x P/E multiple to the average 2027/2028 EPS to maintain the $775 price target.
Amazon's Underestimated AWS Growth
Nowak has raised Amazon's 2027/2028 capital expenditure forecasts by 15% and 29% to $308 billion and $318 billion, respectively.
Unlike Meta, Amazon's EPS forecasts were actually increased. This is due to a simultaneous significant upward revision in AWS revenue projections: 2027/2028 AWS revenue was raised by 3% and 7%, implying year-over-year growth rates of 40% and 36%, reaching absolute values of $243.5 billion and $331.6 billion. The impact of higher depreciation is offset by stronger AWS revenue, leading to 2% and 3% increases in 2027/2028 EPS to approximately $11.53 and $15.05.
Nowak believes current market expectations for AWS revenue are too conservative. The report notes that even Morgan Stanley's own AWS growth forecasts of 35% for 2026 and 40% for 2027 imply incremental revenue per watt of only $7 and $9, which "could be considered conservative."
Additionally, Nowak expects AWS's backlog to increase sequentially by approximately $110 billion in the second quarter, reaching around $475 billion, primarily driven by large orders from private labs, which should bolster confidence in the sustainability of multi-year growth.
The $330 price target is based on applying a roughly 25x P/E multiple to the average 2027/2028 EPS of about $13, implying a PEG ratio of about 1.4x and representing an approximate 30% discount to comparable companies.
Google's AI Leadership and Near-Term Constraints
Nowak assigns Google a price target of $415, implying about 16% upside, with an Overweight rating.
Nowak forecasts Google Cloud growth rates of 77% for the second quarter and 78% for the full year 2026. Search business growth is projected at 17.5% for Q2 2026 and 16% for the full year, with 2027/2028 growth rates of 11% and 8%.
However, Nowak highlights a specific tactical risk: Google is currently in a compute-constrained state, as evidenced by its recent compute leasing agreement with SpaceX. This constraint could hinder near-term revenue growth or product launch cadence, a risk seen as relatively smaller for Meta and Amazon.
Furthermore, the release of Gemini 3.5 is already behind the schedule set at the I/O conference, and Morgan Stanley will monitor further productization progress of Gemini models in the second half of the year.
Comments