Goldman Sachs Raises 2030 Data Center Forecast to 273 GW, Implies $8.5T Investment
In a comprehensive research update from Goldman Sachs Global Investment Research, analysts raised their 2030 global data center power capacity forecast to 273 gigawatts (GW), representing a 26% upward revision (+56 GW) from their previous 217 GW estimate. Measured against an operating global baseline of approximately 103 GW in 2025, reaching this target requires adding roughly 170 GW of net new data center capacity over a five-year window. When evaluated at an estimated all-in capital intensity of $50 million per megawatt, encompassing both physical facility infrastructure and high-density AI server silicon, this expansion implies approximately $8.5 trillion in cumulative global infrastructure and hardware expenditure through 2030.
This analysis evaluates whether regional electrical transmission networks, municipal water resources, and global debt and equity markets can realistically support an infrastructure buildout of this magnitude.
Key Takeaways
-
Goldman Sachs raised its 2030 global data center power forecast from 217 GW to 273 GW, requiring roughly 170 GW in net capacity additions above the 103 GW baseline recorded in 2025.
-
Deploying 170 GW at an estimated full-stack cost of $50 billion per gigawatt implies roughly $8.5 trillion in cumulative spending across AI accelerators, networking fabrics, liquid cooling, and electrical substations.
-
Single-campus mega-sites rated at 75 MW or larger are projected to account for 66% of new global capacity additions by 2028, up from just 25% globally in 2025.
-
Four-to-seven-year utility interconnection queues, three-to-four-year high-voltage transformer lead times, and heavy reliance on 80%–90% project-level debt financing remain the primary execution risks.
Core Drivers: Why Goldman Sachs Raised Its 2030 Forecast by 26%
Goldman Sachs revised its forecast upward because capital deployment from specialized GPU cloud providers , wholesale colocation developers, and cryptocurrency mining companies converting facilities to High-Performance Computing (HPC) has accelerated faster than legacy models anticipated, accompanied by a structural shift toward massive single-campus "mega-sites" exceeding 75 MW.
Detailed Mechanism Breakdown
Historically, Wall Street infrastructure models assumed that global data center growth would be dictated almost entirely by the internal construction schedules of four traditional cloud hyperscalers: Microsoft, Alphabet, Amazon, and Meta Platforms. Analysts simply mapped global megawatt additions to the capital expenditure budgets of those four companies.
Between 2024 and late 2026, however, the supply side of artificial intelligence compute diversified across three major vectors. First, specialized GPU cloud operators, such as CoreWeave, Lambda, Nebius, and Crusoe, tapped tens of billions of dollars in private credit and asset-backed facilities to deploy bare-metal AI clusters at an aggressive pace, prioritizing rapid time-to-market for frontier model laboratories.
Second, publicly traded Bitcoin mining operators, including Core Scientific, IREN, Cipher Mining, TeraWulf, and Hut 8, pivoted a substantial share of their energized power portfolios toward AI and HPC hosting. Because industrial mining sites already possess active high-voltage grid interconnections, utility substations, and environmental permits, retrofitting a mining campus for liquid-cooled AI servers allows developers to energize capacity within 6 to 18 months, bypassing the multi-year utility queues facing greenfield builds.
Third, the physics of distributed AI training and low-latency reasoning inference forced a jump in individual site density. Because synchronizing tens of thousands of GPUs requires ultra-short optical fiber runs to minimize latency, developers shifted away from scattered 20 MW to 40 MW suburban data halls in favor of unified mega-campuses demanding 100 MW to over 1,000 MW (1 GW) of continuous power.
Empirical Market Indicators
Data compiled by Goldman Sachs underscores how rapidly these trends are reshaping the project pipeline, particularly in the United States, where projected 2030 capacity was raised from 125 GW to 152 GW. Across cumulative U.S. capacity additions from 2026 through 2030, analysts raised their estimate for the combined category of Neoclouds, crypto miner HPC conversions, and wholesale colocation providers by 11.8 GW, outpacing the 9.4 GW upward revision assigned to direct hyperscaler self-builds.
At the same time, project scale is consolidating dramatically. Globally, single-site facilities rated at 75 MW or larger accounted for just 25% of newly commissioned capacity in 2025, but Goldman Sachs projects that share to reach 66% of all new additions by 2028. Within the United States, the concentration is even more pronounced: mega-sites of 75 MW or above are forecast to represent 75% of new U.S. capacity additions by 2028, up from 37% in 2025.
Deconstructing the $8.5 Trillion Capital Requirement: Where Does the Money Go?
Translating 170 GW of net new electrical capacity into an $8.5 trillion investment figure requires breaking down the full-stack capital cost of a modern AI factory.
The $8.5 trillion estimate reflects an average all-in cost of $50 million per megawatt across 170 GW of new capacity, where advanced AI server silicon, HBM memory, and networking switches represent roughly 65% to 70% of total spending, while electrical substations, cooling systems, and physical buildings account for the remaining 30% to 35%.
Full-Stack Cost Allocation and Depreciation Mismatches
When industry executives and analysts cite a $50 billion cost per gigawatt of AI data center capacity, that figure spans three distinct engineering layers with vastly different economic lifespans.
The largest cost component, accounting for roughly 65% to 70% of total outlay, or approximately $32.5 billion to $35.0 billion per gigawatt, consists of computational hardware and high-speed networking fabrics. This includes enterprise AI accelerators, high-bandwidth memory, enterprise solid-state storage, and non-blocking InfiniBand or Spectrum-X optical Ethernet switches.
The second layer, representing roughly 20% to 25% of total expenditure, or $10.0 billion to $12.5 billion per gigawatt, covers electrical distribution and mechanical thermal management. This encompasses on-site high-voltage substations, step-down transformers, switchgear, uninterruptible power supply (UPS) battery banks, backup natural gas or diesel generators, and direct-to-chip liquid cooling loops paired with closed-loop dry coolers.
The final layer, accounting for roughly 8% to 10% of total cost, or $4.0 billion to $5.0 billion per gigawatt, covers land acquisition, zoning and environmental remediation, reinforced concrete data hall shells, physical security perimeters, and utility grid interconnection fees.
For capital markets, the critical nuance in this breakdown is the depreciation mismatch between physical real estate and silicon. While concrete data halls, high-voltage substations, and transmission upgrades have useful economic lives of 20 to 30 years, AI GPUs and networking gear depreciate over three to five years as new chip architectures deliver dramatic jumps in inference efficiency per watt. Maintaining 273 GW of active compute therefore involves not just a one-time construction outlay, but recurring multi-trillion-dollar hardware refresh cycles.
Big Tech Cash Flows vs. the $7.55 Trillion Consensus Capex Curve
To evaluate who will pay for this buildout, Goldman Sachs compared the $8.5 trillion global estimate against consensus capital expenditure forecasts compiled by Visible Alpha for the eight largest publicly tracked infrastructure spenders: Alphabet, Amazon, Apple, Meta, Microsoft, Oracle, CoreWeave, and Baidu.
According to Visible Alpha consensus estimates, combined annual capital expenditures across those eight companies are projected to surge from roughly $456 billion in 2025 to $918 billion in 2026 and $1.52 trillion in 2027. Beyond 2027, however, consensus models project annual Capex growth to decelerate sharply into the single digits, reaching $1.67 trillion in 2028, $1.71 trillion in 2029, and $1.74 trillion in 2030, for a five-year cumulative total of roughly $7.55 trillion across the top eight operators.
While $7.55 trillion in projected corporate Capex covers the vast majority of the global requirement on paper, the sharp flattening of growth rates after 2027—combined with recent free cash flow deficits at major hyperscalers—shows that internal corporate cash flows alone cannot carry the entire expansion. Bridging the gap to 273 GW is pushing developers to rely heavily on private credit syndicates, commercial mortgage-backed securities (CMBS), asset-backed securitization (ABS), and sovereign wealth partnerships.
Structural Comparison: Hyperscaler Builds vs. Colocation vs. Crypto Miner Conversions
The table below contrasts the three primary development models driving the 2030 capacity forecast across timeline, financing, and operational dimensions:
| Operational Metric | Hyperscaler Self-Build (e.g., Microsoft, Alphabet) | Wholesale / Build-to-Suit Colocation (e.g., Digital Realty, Vantage) | Crypto Miner HPC Conversion (e.g., Core Scientific, IREN, Cipher) |
| Typical Site Scale | 150 MW to 1,000+ MW (Multi-phase mega-campuses) | 50 MW to 300+ MW (Dedicated campus shells) | 50 MW to 300+ MW (Retrofitted industrial power sites) |
| Time-to-Power Lead Time | 4 to 7 Years (New utility grid interconnection queues) | 3 to 5 Years (Substation permitting and grid upgrades) | Immediate to 18 Months (Pre-existing energized substations) |
| Primary Financing Structure | Corporate balance sheet cash flow & senior corporate bonds | 65% to 90% project-level debt (backed by investment-grade leases) | High-yield project debt, tenant Capex prepayments, and equity |
| Cooling & Shell Profile | Purpose-built direct-to-chip liquid cooling & closed-loop water | Modular high-density liquid-ready shells | Requires heavy retrofit of mining sheds into tier-3 data halls |
| Typical Lease / Offtake Term | Internal corporate consumption (15–30 year asset life) | 10 to 15 Year triple-net take-or-pay leases | 10 to 15 Year fixed hosting contracts (e.g., CoreWeave/Hyperscalers) |
| Primary Execution Risk | Multi-year utility transmission delays & Capex ROI scrutiny | Rising borrowing spreads & construction cost inflation | Retrofit execution complexity & tenant concentration risk |
Physical and Financial Bottlenecks Threatening the 2030 Timeline
While financial models demonstrate strong demand for 273 GW of capacity, converting paper commitments into energized megawatts faces three formidable real-world friction points.
Electrical Grid Queues and High-Voltage Transformer Shortages
The most binding constraint on reaching 273 GW by 2030 is physical electricity delivery. Across major U.S. grid operators such as PJM Interconnection, ERCOT in Texas, and MISO, the engineering study and approval process for connecting a new multi-hundred-megawatt load averages four to seven years.
Even after a project secures utility approval, supply-chain bottlenecks for heavy electrical equipment remain acute. Manufacturing lead times for Large Power Transformers (LPTs) and 345 kV to 765 kV switchgear have stretched from roughly two years in 2021 to between three and four years (160 to 210 weeks) in 2026, constrained by global shortages of grain-oriented electrical steel (GOES) and specialized winding labor. These delays are already causing real-world project friction: major cloud providers have paused or reallocated select regional lease commitments where utilities pushed power delivery dates past 2028, while large developments have encountered permitting and infrastructure hurdles over power and local resource requirements.
Municipal Zoning Moratoriums and Water Constraints
As campuses scale past 75 MW toward gigawatt size, local community pushback has intensified. Residential ratepayers and state utility commissions in hubs like Northern Virginia, Ohio, Georgia, and Pennsylvania are increasingly scrutinizing whether multi-billion-dollar transmission upgrades built for data centers will inflate household electricity bills.
Combined with concerns over evaporative cooling water consumption and continuous noise from chillers and backup generators, local boards across several U.S. states have enacted temporary moratoriums or stricter zoning rules on data center permits—delaying or blocking nearly $200 billion in planned U.S. projects according to industry trackers. While operators are pivoting toward water-free closed-loop cooling and behind-the-meter natural gas or nuclear power agreements, local permitting friction threatens to push many 2028–2030 deliveries to the right.
High-Leverage Project Debt and Credit Bifurcation
On the financial side, Goldman Sachs analysts highlighted the growing reliance on high-leverage project finance across the colocation and developer ecosystem. When a developer secures a 12-to-15-year lease with an investment-grade hyperscaler, private credit funds and structured debt desks frequently underwrite 80% to 90% of the project's construction cost using debt.
While 80% to 90% leverage works smoothly when construction finishes on time and tenants have Aaa/Aa credit ratings, it leaves minimal margin for error. If transformer shortages or municipal lawsuits delay a facility’s energization date by 12 to 18 months, interest expenses compound while lease revenues remain paused. Furthermore, capital markets have sharply bifurcated: projects backed by binding hyperscaler take-or-pay leases enjoy deep credit access, whereas speculative builds or facilities leased to heavily leveraged counterparties face much wider borrowing spreads and stricter underwriting criteria.
Conclusion
Goldman Sachs’ upward revision to 273 GW underscores the extraordinary scale of the global artificial intelligence infrastructure buildout, projecting 170 GW of net capacity additions and roughly $8.5 trillion in full-stack capital investment through 2030. While $7.55 trillion in projected five-year Capex across the top eight technology companies and rapid HPC conversions by cryptocurrency miners demonstrate intense demand, financial capital cannot instantly manufacture high-voltage transformers or bypass utility interconnection queues. Ultimately, whether the global market achieves 273 GW on schedule by 2030 will depend on grid modernization, behind-the-meter power generation, and disciplined debt underwriting across mega-site developments.
FAQs
Why did Goldman Sachs raise its 2030 data center forecast to 273 GW?
Goldman Sachs increased its 2030 global forecast by 26% due to faster-than-expected capacity commitments from specialized AI clouds (Neoclouds), cryptocurrency miners converting energized sites to AI/HPC hosting, and surging demand for single-campus mega-sites rated at 75 MW or larger.
What does the $8.5 trillion data center investment estimate cover?
The $8.5 trillion figure represents the estimated all-in capital required to add 170 GW of net new capacity between 2025 and 2030 at roughly $50 million per megawatt. About 65% to 70% of that cost goes toward AI GPUs, memory, and networking gear, while 30% to 35% funds electrical substations, liquid cooling, land, and building shells.
Why are Bitcoin miners a major driver of the upward forecast revision?
Connecting a brand-new data center to the electrical grid can take four to seven years. Because industrial Bitcoin miners already own energized high-voltage substations and multi-hundred-megawatt power contracts, retrofitting mining sites for AI hosting allows developers to bring compute online within 6 to 18 months, helping drive an 11.8 GW upward revision across U.S. Neocloud, colocation, and miner conversion projects.
What are the biggest risks to reaching 273 GW by 2030?
The primary bottlenecks are physical grid interconnection backlogs, three-to-four-year delivery delays for high-voltage transformers, local community moratoriums over electricity rates and water use, and financial risks tied to 80%–90% debt financing on third-party data center developments.
Disclaimer
This general market information, possibly from third-party, commercial, or sponsored sources, is provided for reference and educational purposes only and does not constitute financial or investment advice, an offer, solicitation, or guarantee. We disclaim liability for its accuracy, completeness, reliability, and any resulting losses. Investments/trading are risky; past performance doesn’t guarantee future results. Users should research, judge prudently, and take full responsibility. For further details, please consult KuCoin’s Terms of Use and Risk Disclosure.