Global data center spending crossed the $1 trillion threshold in 2026 and is on track to more than double to $2.1 trillion by 2028, according to a Bank of America Global Research note published October 23. At a 37% compound annual growth rate from 2025 forward, the bank projects the broader market will reach $2.93 trillion by 2030 — a scale that has reframed AI infrastructure investment as a construction and energy challenge as much as a semiconductor race.
The Bank of America analysis makes concrete the per-megawatt cost structure of building that infrastructure. Standard data center facilities run approximately $49 million per megawatt to develop. AI-optimized “neocloud” facilities—built specifically for the dense GPU workloads that train and run large language models—run about $57 million per megawatt. Those numbers, and where the money flows within them, tell a clear story about why power grid access has become the binding constraint on how fast the market can grow.
What $49 Million Per Megawatt Actually Buys
The Bank of America note provides some of the most granular public cost breakdowns yet published on data center construction. At $49 million per megawatt for a standard facility, IT equipment accounts for roughly 78% of total project cost according to Bank of America. Servers are the single largest line item: $31.4 million per megawatt, representing 64% of total development cost. Networking equipment adds $4.8 million per megawatt.
The remaining 22% — roughly $10.7 million per megawatt — covers the physical plant: electrical switchgear, backup generators, chillers, and cooling infrastructure. Bank of America places these categories at between $100,000 and $1.1 million per megawatt depending on the facility type. For AI neocloud facilities, that physical-plant cost is higher: these buildings require liquid-cooling infrastructure and specialized power-delivery systems to manage the thermal output of GPU racks running at 100 to 300 kilowatts per rack, compared to 10 to 20 kilowatts for conventional compute racks. That engineering requirement — not real estate or permitting — is what drives the $57 million per megawatt price tag for AI-optimized facilities, a 16% premium over standard builds.
The cost structure matters because it reveals where the construction economy of the AI boom actually flows. Two-thirds of every dollar spent building a data center goes to IT equipment — primarily servers — but a significant and growing share flows to industrial manufacturers of switchgear, generators, and cooling systems. These are companies that had minimal exposure to the technology sector five years ago.
How Is it That AI Spending Helps Non-Tech Suppliers?
The build cycle’s industrial footprint has grown faster than its semiconductor footprint. Data center construction reached $58.1 billion in the first five months of 2026 alone — roughly four times the pace recorded in the same period in 2025, according to ConstructConnect. That construction spending flows to builders and manufacturers outside the traditional technology supply chain: the companies making the switchgear, generators, chillers, and precision cooling systems that keep GPU clusters operational.
Bank of America’s per-megawatt breakdown places cooling and power components — the physical infrastructure layer — between $100,000 and $1.1 million per megawatt depending on specification. At the scale of projects underway, even the low end of that range represents substantial aggregate demand for industrial suppliers. At $49 million per megawatt and with hundreds of gigawatts of new capacity under development or planned, the market for non-IT data center components alone runs into the tens of billions of dollars annually.
Hyperscaler Capex: Four Companies, $650 Billion in One Year
The four largest U.S. hyperscalers — Amazon, Alphabet, Meta, and Microsoft — are collectively on track to spend between $650 billion and $725 billion on capital expenditures in 2026, up roughly 77% from last year’s already record-breaking total. Amazon has committed approximately $200 billion in capital expenditures for the year. Alphabet is guiding $195 billion to $205 billion following multiple upward revisions in 2026. Meta has raised full-year guidance to $125 billion to $145 billion, citing higher component costs and expanded data center requirements. Microsoft is tracking toward approximately $190 billion for the calendar year.
According to TrendForce’s May 2026 analysis, including Chinese hyperscalers such as ByteDance, Tencent, Alibaba, and Baidu, combined capital expenditure across the top nine global cloud providers could approach $830 billion in 2026.
Goldman Sachs has projected that the four largest U.S. hyperscalers alone will spend approximately $5.3 trillion on capital expenditures between fiscal year 2025 and fiscal year 2030. The cumulative figure places AI infrastructure investment on a scale comparable to the interstate highway system as a share of GDP — a multi-decade physical buildout, not a technology product cycle.
The financial strain from this commitment is already visible in cash flow statements. Amazon’s trailing twelve-month free cash flow has compressed sharply from approximately $26 billion a year ago to roughly $1.2 billion as capital expenditures consume nearly all operating cash flow. Microsoft’s capital expenditures exceeded operating cash flow in fiscal Q2 2026. Moody’s has reported that hyperscalers collectively hold approximately $662 billion in data center lease commitments that have been signed but not yet commenced — a figure that Moody’s analyst David Gonzales has described as reflecting obligations not yet triggered as liabilities “as of this time, but they will.”
Neoclouds: Built for Speed Where Hyperscalers Have Waitlists
Beyond the hyperscalers, a category of specialized AI infrastructure providers — commonly called neoclouds — has become an increasingly significant part of the capacity picture. CoreWeave, the largest neocloud operator, posted approximately $2.58 billion in revenue in the second quarter of 2026, representing year-over-year growth of 112%. CoreWeave is targeting 8 gigawatts of data center capacity by 2030.
In April 2026, CoreWeave announced an expanded agreement under which Meta will pay approximately $21 billion for dedicated AI cloud capacity through December 2032. The deal illustrates the structural role neoclouds now occupy: even the largest hyperscalers, spending hundreds of billions on their own infrastructure, are turning to specialized providers to fill gaps in their own build timelines.
The neocloud market is estimated at approximately $35 billion in 2026, growing from $24 billion in 2025, according to Mordor Intelligence’s neocloud report. Mordor Intelligence projects the segment will reach about $237 billion by 2031, at a compound annual growth rate of about 46%. The competitive advantage neoclouds offer is speed and specialization: while conventional hyperscale data centers can take years to bring online, neocloud providers have shown they can deploy GPU infrastructure within months by focusing on AI-specific workloads and avoiding the general-purpose architecture of hyperscale platforms.
Why Is Power the New Bottleneck for AI Data Centers?
The Bank of America analysis points to something that the per-megawatt cost structure alone cannot capture: the increasingly binding constraint on how fast all of this can be built is not chip availability, not capital, and not construction labor. It is electrical grid access.
Global data center electricity consumption is projected to reach 565 terawatt hours in 2026, a 26% increase from 447 terawatt hours in 2025, according to Gartner. Power demand is expected to reach 132 gigawatts this year, up from 104 gigawatts in 2025, with projections of 290 gigawatts by 2030.
Gartner’s Director Analyst Linglan Wang has described data center power security as “the new battleground for scaling and protecting margins in the global AI race.” The specific engineering reason is straightforward: large-load data center connections must queue for grid interconnection alongside renewable energy projects and other industrial loads. Interconnection queues in the United States already hold more than 2,000 gigawatts of pending requests, and new data center projects requiring hundreds of megawatts of dedicated grid capacity face multi-year waits in high-demand markets.
Wholesale electricity costs near major U.S. data center clusters have reportedly risen by as much as 267% in some markets as grid supply fails to keep pace with demand. Those cost increases do not fall only on the technology companies building the facilities: in regions without large-load tariff frameworks, grid upgrade costs triggered by new data center connections may be partially allocated to existing residential and commercial ratepayers. Oregon’s PGE has implemented a large-load tariff that shifts more of those grid upgrade costs to the large industrial users driving the demand.
In response to grid access constraints, hyperscalers are pursuing on-site generation at unprecedented scale. Microsoft has signed agreements for 10.5 gigawatts of new renewable energy capacity with Brookfield Renewable Partners. Google has partnered with NextEra Energy on gigawatt-scale clean-energy data center campuses. And xAI, Elon Musk’s AI company, moved road-portable natural gas turbines to its Memphis, Tennessee facility to power operations while permanent grid connections are secured — an improvised solution that drew environmental scrutiny and local opposition before some units were removed.
What the $2.1 Trillion Forecast Means for Investors Right Now
The Bank of America projection is a forecast, not a guarantee, and the research note acknowledges substantial execution risk. Securing electricity, obtaining construction permits, managing local opposition, and generating sufficient AI product revenue to justify the investment are all material uncertainties. Goldman Sachs’ $5.3 trillion cumulative figure through 2030 implies that the four largest U.S. hyperscalers will collectively spend more on infrastructure over the next four years than the annual GDP of most countries.
What makes that risk concrete is the pressure already visible in hyperscaler cash flows. Amazon’s free cash flow compression from approximately $26 billion to $1.2 billion in one year is not a temporary accounting artifact — it reflects a deliberate multi-year commitment to infrastructure spending that will not produce returns on the same timeline. The $662 billion in signed-but-not-commenced lease commitments documented by Moody’s represents financial obligations that will convert to balance-sheet liabilities as facilities are built out.
For investors, the Bank of America analysis suggests that the beneficiary class is broader than the AI software and chip layer alone. The per-megawatt breakdown points to switchgear manufacturers, generator suppliers, liquid cooling system makers, and data center construction contractors as a structural part of the AI infrastructure economy—companies not typically analyzed as technology-sector investments until the current buildout created sustained demand for their products.
The next phase of the AI boom may be measured as much in megawatts, concrete, and cooling capacity as in algorithms and chips. At $49 million per megawatt — and with hundreds of gigawatts under development — the infrastructure required to run AI at scale is being built right now, at a pace and cost that no previous technology buildout has matched.
Frequently Asked Questions
How much does it cost to build a data center per megawatt in 2026?
Bank of America Global Research estimates average data center development costs at approximately $49 million per megawatt of capacity, based on a research note published October 23, 2026. That figure covers IT equipment (servers at $31.4 million per megawatt, networking at $4.8 million per megawatt) and physical infrastructure (electrical switchgear, backup generators, chillers, and cooling systems at $100,000 to $1.1 million per megawatt depending on type). AI-optimized neocloud facilities — which require specialized liquid cooling and higher-density power delivery for GPU racks — run approximately $57 million per megawatt, a 16% premium that reflects the engineering requirements of running high-performance accelerators continuously rather than the cost of land or permits.
What are hyperscalers spending on AI infrastructure in 2026, and is it sustainable?
The four largest U.S. hyperscalers — Amazon, Alphabet, Meta, and Microsoft — are collectively on track for between $650 billion and $725 billion in capital expenditures in 2026, up roughly 77% from the prior year. Goldman Sachs projects $5.3 trillion in cumulative spending from this group through fiscal year 2030. The sustainability question is real: Amazon’s free cash flow has compressed from approximately $26 billion to $1.2 billion in trailing twelve months as capex consumes operating cash flow, and Moody’s has flagged $662 billion in signed-but-not-yet-commenced data center lease commitments as a forward financial obligation. Whether AI product revenues can justify those commitments over a four-to-six year return horizon is the central investment risk in the current buildout.
What is a neocloud and how is it different from a hyperscaler?
A neocloud is a specialized AI cloud provider that builds GPU-dense facilities optimized for AI training and inference workloads—as distinct from general-purpose hyperscale data centers operated by Amazon Web Services, Microsoft Azure, and Google Cloud. Neoclouds such as CoreWeave, Lambda, Nebius, and Crusoe compete primarily on deployment speed and GPU availability: where a hyperscaler data center can take years to bring online, neocloud operators have shown they can provision GPU capacity within months. CoreWeave, the largest neocloud, posted approximately $2.58 billion in revenue in the second quarter of 2026, up 112% year over year, and signed a $21 billion capacity agreement with Meta. The neocloud market is projected to grow from approximately $35 billion in 2026 to approximately $237 billion by 2031.
Why is the electrical grid — not chips — now the main constraint on AI data center growth?
GPU availability, which was the binding constraint on AI infrastructure in 2023 and 2024, has loosened as semiconductor production has scaled. The new constraint is electrical grid access. Data centers require large, dedicated power connections to the electrical grid, and grid interconnection queues in the United States and Europe involve multi-year wait times for projects requiring hundreds of megawatts. Wholesale electricity prices near major U.S. data center clusters have risen by as much as 267% in some markets as grid capacity fails to keep pace with demand. Hyperscalers are responding with on-site generation (Microsoft’s 10.5-gigawatt renewable deal with Brookfield, Google’s campus partnership with NextEra, xAI’s deployment of portable natural gas turbines in Memphis), but these are workarounds for a grid infrastructure gap that will take years to close through conventional transmission and generation expansion.
