Morgan Stanley has sharply increased its estimate for capital spending by the world’s largest cloud companies, projecting a historic AI infrastructure buildout that could push combined expenditures to $1.2 trillion in 2027 and $1.4 trillion in 2028.
The firm’s model suggests global spending tied to artificial intelligence infrastructure could approach $3 trillion by 2028, with about $2.9 trillion of that amount connected to data center construction. The forecast points to a rapid shift in the technology sector from software-led growth toward a physical race for chips, land, cooling systems, power contracts and grid access.
The scale of the projected spending has broad consequences. It could reshape earnings at major technology firms, intensify pressure on electricity markets, raise costs for AI computing, and force cloud operators to prove that expensive data center capacity can be converted into durable revenue from advertising, APIs, enterprise tools and cloud services.
Morgan Stanley expects available AI computing capacity to rise from roughly 30 gigawatts in 2025 to nearly 120 gigawatts by 2028. That would represent a fourfold increase in only three years, a pace that would require enormous purchases of advanced chips, networking equipment and supporting energy infrastructure.
Meta is projected to expand its capacity to about 21 gigawatts over that period, while Amazon could reach roughly 35 gigawatts. Google is expected to add the most new capacity among the major cloud peers, supported by continued deployment of its Gemini models and its in-house TPU architecture across cloud products.
The forecast highlights one of the central questions now facing the AI boom: whether the world’s largest technology companies can build fast enough, power those facilities reliably, and earn enough revenue from AI services to justify the spending.
AI buildout becomes a power and construction race
The AI infrastructure cycle is no longer only about buying graphics processors. It now includes the full physical system needed to operate advanced computing clusters at massive scale.
Data centers require land, power substations, transmission lines, cooling equipment, backup generation, water access in some regions, and long-term agreements with utilities. As AI models become larger and more power-intensive, the cost of adding each new gigawatt of capacity has climbed.
Morgan Stanley estimates the cost of building one gigawatt of computing power at about $35 billion for Nvidia GB200 systems and $39 billion for GB300 systems. The estimate rises to about $49 billion for Vera Rubin systems. For custom silicon, the report places Google’s TPU v7 at about $27 billion per gigawatt and Amazon’s Trainium3 at about $21 billion.
The variation reflects differences in chip design, memory needs, energy use and system architecture. But the direction is clear: AI capacity is becoming more expensive as the demand for high-bandwidth memory, advanced networking and power density increases.
Associated infrastructure costs are also rising. The report estimates that supporting infrastructure has moved from roughly $10 million per megawatt to a range of $11 million to $19 million per megawatt. That includes the non-chip spending needed to make AI servers usable at commercial scale.
This matters because power availability is becoming as important as chip availability. A company may secure servers and accelerators, but still face delays if it cannot obtain grid approvals, utility connections, construction labor or cooling systems on time.
Meta’s spending could pressure cash flow before revenue appears
Meta is one of the most closely watched companies in the AI capital spending cycle because of the size of its planned buildout and the uncertainty around near-term monetization.
Morgan Stanley projects Meta’s capital expenditures could rise to $225 billion in 2027 and $250 billion in 2028. Such spending could weigh on short-term free cash flow and earnings, especially if revenue from AI products takes longer to materialize.
The report also argues that the spending could create the foundation for several new business lines. These may include AI-powered advertising tools, Meta AI search, API access, cloud services and paid productivity subscriptions for businesses.
In earnings per share terms, Morgan Stanley’s model suggests potential AI-related revenue could add about $10 to Meta’s 2028 EPS, compared with a baseline estimate of $33.41. That additional contribution would depend on adoption across several products rather than one single source of revenue.
One modeled opportunity is API access. According to the report, each 100 megawatts of GB300 capacity dedicated to API operations could generate about $8.59 billion in revenue and add about $1.91 to 2028 EPS, assuming utilization of roughly 75%.
That assumption is important. Idle capacity would weaken returns. High utilization, stable pricing and strong gross margins would be needed for the economics to work at the scale being projected.
Meta’s large advertiser base could also support subscription-style AI tools. If one-quarter of Meta’s 15 million advertisers paid $200 per month for automation, content creation and productivity services, annual revenue could rise by about $8 billion. Morgan Stanley estimates that would add about $2 to 2028 EPS.
The opportunity is clear, but so is the risk. Advertisers would need to see measurable gains in efficiency, conversion rates, content output or campaign performance before paying for such tools at scale. If adoption is slower than expected, the payback period on infrastructure spending could lengthen.
Amazon and Google remain central to the AI cloud race
Amazon’s cloud division is also expected to benefit from the AI infrastructure cycle. Morgan Stanley forecasts Amazon Web Services revenue growth of 40% in 2027 and 36% in 2028. The firm also expects backlog to rise by about $110 billion quarter over quarter to roughly $475 billion.
That backlog figure would point to strong demand for future cloud services, including AI computing capacity. For Amazon, the challenge is to balance growth in AI workloads with the cost of building and operating the infrastructure needed to support them.
Amazon’s use of its Trainium chips may help manage costs, according to the estimates. Morgan Stanley places the cost of Amazon’s Trainium3 systems at about $21 billion per gigawatt, lower than several other advanced AI systems. If performance and software support remain competitive, custom chips could help Amazon defend margins while expanding capacity.
Google is expected to expand new data center capacity faster than any other peer in the forecast. Its AI infrastructure strategy is built around both its Gemini model family and its long-running TPU chip program.
Google’s vertical approach could offer advantages if its custom silicon performs efficiently across internal products and external cloud services. The report estimates Google’s TPU v7 systems at about $27 billion per gigawatt, below some Nvidia-based systems but still part of a very expensive buildout.
For cloud companies, the central financial test is similar across the board. They must convert capital spending into recurring revenue. That could come through enterprise AI contracts, search and advertising improvements, software subscriptions, developer tools, APIs and data services.
Electricity becomes a limiting factor
The projected spending boom would have major consequences for physical energy markets. AI data centers consume large amounts of electricity, and the planned expansion could create a supply shock in regions where grid capacity is already tight.
Technology companies are moving faster than many utilities and local power systems can respond. Power plants, transmission lines and substations can take years to permit and build. Data center developers, by contrast, are trying to secure capacity on much shorter timelines.
Separate banking-sector estimates have warned that the domestic power grid could face a shortage of around 20% within two years if demand continues rising and alternative power sources are not added quickly enough. Some forecasts point to a potential shortfall of about 44 gigawatts by the end of 2028 if new supply, grid upgrades and power procurement fail to keep pace.
Those figures underline why power access has become a strategic asset. In some locations, the key bottleneck is not demand for AI services or availability of construction capital. It is simply whether enough electricity can be delivered to the site.
Political and environmental scrutiny may also increase. Large data centers can face criticism over land use, power consumption, water needs and emissions. Approval timelines could become more uncertain, especially as power demand becomes a public policy issue heading into the 2028 U.S. election cycle.
AI data centers and digital miners compete for the same resources
The AI buildout is also creating direct competition between machine learning data centers and digital token miners.
Both industries require large amounts of power, cooling infrastructure, electrical equipment and suitable real estate. Both also prefer locations with cheap electricity and fast interconnection access. As AI cloud operators search for capacity, some digital mining sites have become attractive because they already have power agreements, substations or large industrial buildings.
Several large digital miners have already agreed to lease or repurpose facilities for major cloud service customers under long-term arrangements. These deals can provide miners with steadier revenue than token production, while giving cloud companies faster access to powered sites.
For traders, these property and power deals have become important market signals. A mining company with underused buildings and secured electricity may be valued differently if those assets can be leased to AI cloud customers. The value of the site may depend less on token prices and more on power access, location and the ability to support high-density computing.
The overlap also creates pressure. If AI companies are willing to pay more for power than miners, some mining operations may be forced to relocate, reduce activity or renegotiate energy contracts. In regions with constrained grids, public officials may also prioritize projects based on jobs, tax revenue, strategic importance or political pressure.
Supply, regulation and demand remain the biggest constraints
Morgan Stanley’s projections depend on three major factors: supply, regulation and demand.
On the supply side, chip availability remains essential. Advanced AI systems require high-end accelerators, high-bandwidth memory, networking components and specialized servers. Any disruption in memory supply, packaging capacity or chip production could delay the buildout.
Labor is another constraint. Massive data center construction requires electricians, engineers, construction crews, cooling specialists and grid workers. A shortage of skilled labor could raise costs and extend timelines.
Regulation may be just as important. Grid approvals, environmental permits, zoning decisions and energy agreements can all slow projects. In areas where power demand is already rising, utilities may need to decide which projects receive capacity first.
Demand is the ultimate test. Cloud operators can build capacity, but they still need paying customers to use it consistently. AI services must become embedded in daily business operations, advertising systems, search products, software tools and developer workflows.
Meta’s lower-priced APIs may attract early users, but long-term profitability will depend on sustained usage and margin performance. If prices fall too quickly because of competition, revenue may not cover the cost of capacity as expected.
The same issue applies to Amazon, Google and other cloud operators. The market must show that each dollar of capital spending can generate recurring revenue, not just short-term excitement around AI.
Earnings reports could move the sector
Upcoming corporate earnings reports are likely to receive close attention from traders because they may reveal whether the AI infrastructure cycle is accelerating or slowing.
Key signals include changes to capital expenditure guidance, comments on data center timelines, cloud backlog growth, AI product revenue, gross margins and power availability. A single negative report from a major technology company could affect sentiment across chipmakers, utilities, data center operators, digital miners and cloud software firms.
Market strategists have warned that weaker share prices could eventually force large technology companies to delay some construction plans. If that happened, pressure on local utilities could ease and some hardware supply constraints could loosen.
For now, however, the direction of spending remains upward. The largest cloud companies appear to be positioning for a future in which AI computing becomes a core utility for advertising, enterprise software, search, content creation, automation and developer services.
The central question is no longer whether AI infrastructure will be built. It is whether it can be built fast enough, powered reliably and monetized at a level that supports the enormous cost.
Morgan Stanley’s $1.4 trillion capital spending estimate for 2028 captures the size of the bet. If demand for AI tools continues to grow and cloud firms maintain strong pricing, the buildout could support new revenue streams for years. If adoption slows, margins weaken or power constraints intensify, the same spending could become a drag on cash flow and earnings.
The next phase of the AI boom will be measured not only in model performance, but also in gigawatts, substations, cooling systems and long-term power contracts. For technology companies and traders alike, electricity has become one of the most important variables in the AI race.
Explore how AI-driven markets shape crypto trends in 2026 with our in-depth guide, read the full analysis here.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
