AI capex forecast 2026: Global infrastructure spending
Explore the 2026 AI capex forecast, detailing global infrastructure spending, datacenter construction costs, regional investments and architecture shifts.
In this article8

Global AI infrastructure spending projections
Goldman Sachs Research estimates that global AI-related investment will total around $1 trillion in 2026, including $581 billion in the United States, once private companies and non-US firms are added to the commonly cited hyperscaler capex figure[1]. On a narrower definition counted separately by IDC, spending on hardware infrastructure, covering compute servers and storage arrays, is forecast to reach $497 billion in 2026, representing an estimated 56% annual growth rate over 2025 levels[2].
Total AI infrastructure spending reached $89.7 billion in the first quarter of 2026 alone, expanding 33.1% year-over-year compared to the first quarter of 2025[2]. Long-term modeling indicates that global AI infrastructure markets will expand beyond $1.08 trillion by 2029 and reach $1.21 trillion by 2030, maintaining a five-year compound annual growth rate of approximately 30% from 2025[2].
| Metric | Projected value | Reporting window | Primary source |
|---|---|---|---|
| Global AI-related investment | $1.00 trillion | Full-year 2026 | Goldman Sachs Research |
| Hardware infrastructure spending | $497 billion | Full-year 2026 | IDC Worldwide Tracker |
| Q1 AI infrastructure spending | $89.7 billion | Q1 2026, up 33.1% year over year | IDC Worldwide Tracker |
| Cumulative AI investment since 2022 | $1.80 trillion | Through end-2026 | Goldman Sachs Research |
| Long-term AI infrastructure market | $1.21 trillion | Forecast 2030 | IDC Worldwide Tracker |
As capital commitments scale, macroeconomic analysis shows AI investment absorbing a growing share of output. In the United States, AI capital expenditure is forecast to rise from 1.8% of gross domestic product in 2026 to 2.5% in 2027 and 2.8% in 2028, levels Goldman Sachs Research describes as consistent with the 2% to 5% peak investment impulses observed in prior general-purpose technology buildouts[1].
Datacenter construction pipelines and scale
Datacenter construction sits in a separate, longer-dated capital pool from the annual hardware line: it funds land, shells, power and cooling over multi-year build cycles, whereas the $497 billion hardware figure is a single-year 2026 spend on servers and storage. At the end of September 2025, the latest global tracking cut, BloombergNEF identified 23.1 GW of critical IT capacity actively under construction across 831 physical facilities[3].
The broader pipeline of announced projects stands at approximately 190 GW across 777 large-scale datacenters and compute campuses announced since 2024[4]. Grid interconnect delays, transformer shortages and zoning approvals then slip the schedule. While approximately 16 GW was originally slated for 2026 delivery across roughly 140 projects, only about 5 GW was under active construction on site, with analysts forecasting that 30% to 50% of scheduled 2026 capacity will slip into subsequent years[4].
Execution constraints across the construction pipeline
- Lead times for standard large power transformers average 128 weeks, close to two and a half years, with generator step-up units averaging 144 weeks, according to Wood Mackenzie's Q2 2025 survey[5].
- Grid connection queues in the primary European hubs run far longer than the build itself: about eight years in London and ten years in Amsterdam, against physical construction times of roughly two years.[7]
- Average global shell-and-core construction costs stand at $11.3 million per megawatt in 2026, up from $10.7 million in 2025, while dense liquid-cooled AI facilities require $30 million to $40 million per megawatt once specialized hardware fit-out is included.[7]
- Attrition in transmission queues remains high: of the US capacity that requested interconnection between 2000 and 2020, only 13% had reached commercial operations by the end of 2025, while 75% had been withdrawn[6].
This divide between announced datacenter capacity and energised megawatts means that physical power access, rather than semiconductor production capacity, dictates operational deployment schedules globally.
Hyperscaler capex and regional capacity divides
Capital expenditure remains heavily concentrated among tier-one cloud providers and operators. Total company-level capital spending across the 14 largest publicly owned datacenter operators, which covers land, buildings, power, networking and chips rather than AI hardware alone, is projected to approach $750 billion in 2026, increasing by nearly two thirds from under $450 billion in 2025[3]. Because that is an annual outlay by a single cohort of operators, it is not directly comparable with the $497 billion global hardware forecast, which counts only servers and storage.
The geographic distribution of active construction demonstrates a substantial structural imbalance between North America and other regions. The Americas account for 17.0 GW of ongoing construction across 311 sites, with the United States alone hosting 15.9 GW[3]. In contrast, the Europe, Middle East, and Africa (EMEA) region accounts for 2.9 GW across 258 sites, while Asia-Pacific (APAC) hosts 3.2 GW across 283 sites[3].
| Region | Active IT capacity, end-Sept 2025 (GW) | Active sites | Average capacity per site (MW) | Primary source |
|---|---|---|---|---|
| United States | 15.9 GW | Not reported | Not reported | BloombergNEF |
| Americas (Total) | 17.0 GW | 311 | 54.7 MW | BloombergNEF |
| Asia-Pacific (APAC) | 3.2 GW | 283 | 11.3 MW | BloombergNEF |
| Europe, Middle East & Africa | 2.9 GW | 258 | 11.2 MW | BloombergNEF |
North American campuses average over 54 MW per site, enabling large synchronous training clusters. In contrast, European and Asian builds average approximately 11 MW per site, reflecting local grid limits and prompting European buyers to examine EU sovereign alternatives to global cloud providers.
European gigafactories and sovereign funding
European policymakers have answered that concentration with public capital. The European Commission launched the €20 billion InvestAI initiative in partnership with the European Investment Bank (EIB) Group to finance large-scale artificial intelligence infrastructure across member states[8].
InvestAI was initially designed to co-finance up to five European AI gigafactories, expanding beyond the initial network of 19 EuroHPC AI factories[8]. Each planned gigafactory is to bring together over 100,000 advanced AI processors, the compute scale required for frontier model training inside European borders.[10] On 30 July 2026 the Commission opened the formal call, widening the target to up to seven gigafactories backed by up to €10 billion in EU and national funding and intended to unlock at least €20 billion in private investment[8].[9]
Strategic pillars of the European compute framework
- InvestAI blended capital: A memorandum of understanding signed in December 2025 pairs the €20 billion InvestAI facility with EIB Group advisory support and potential co-financing ahead of the formal gigafactory call, which opened on 30 July 2026[11].
- Gigafactory deployment: Financing up to seven major installations, each bringing together over 100,000 advanced AI processors for foundation model training in industry and science[8].[9]
- Application integration: Allocating resources to specialized initiatives such as GenAI4EU for biotech, energy optimization, and public sector workflows[12].
- Jurisdictional compliance: Maintaining infrastructure governance within EU regulatory boundaries to preserve legal control and data sovereignty.
That layer places compute capacity under European legal standards, insulating regional buyers from extraterritorial exposure and offshore allocation queues.
The shift to ARM-based accelerated servers
In the first quarter of 2026, server systems accounted for 97.6% of total global AI infrastructure spending, representing $87.6 billion out of $89.7 billion in quarterly expenditure[2].
The composition of accelerated servers has undergone a rapid architectural crossover. In Q1 2026, non-x86 ARM-based accelerated server value reached $53.0 billion, overtaking traditional x86 accelerated systems at $34.6 billion[2]. This marks a swift transition from the third quarter of 2025, when x86 held $51.9 billion compared to $29.8 billion for ARM platforms.
| Platform architecture | Q3 2025 spend | Q4 2025 spend | Q1 2026 spend | Market trajectory |
|---|---|---|---|---|
| ARM-based (Non-x86) accelerated | $29.8 billion | $47.5 billion | $53.0 billion | Dominant platform / expanding |
| x86 accelerated systems | $51.9 billion | $42.7 billion | $34.6 billion | Declining share / custom racks |
The crossover reflects tier-one buyers standardizing on liquid-cooled, rack-scale configurations pairing ARM host processors with high-bandwidth accelerator fabrics, which cuts CPU power overhead and raises interconnect throughput across distributed clusters.
The enterprise storage refresh cycle
During 2024 and 2025, enterprise IT departments redirected capital toward compute clusters and postponed routine storage refreshes; that deferral is now reaching its operational limit. In the first quarter of 2026, worldwide external enterprise storage revenue grew 22.7% year over year as delayed lifecycle replacements combined with active AI training and pipeline storage mandates[13].
Dedicated AI storage systems represented $2.2 billion of quarterly spending in Q1 2026, accounting for 2.4% of the $89.7 billion of AI infrastructure capital expenditure recorded in that quarter[2]. Modest against compute spend, it is nonetheless the layer that feeds checkpointing and multi-modal ingest pipelines.
Factors driving storage capital allocation
- Checkpoint throughput: High-performance parallel file systems and all-flash arrays prevent GPU idle time during distributed model training checkpoint writes.
- Data pipeline ingest: Unstructured data preprocessing for retrieval-augmented generation and domain fine-tuning requires scalable object and file tiers.
- Lifecycle refresh deadlines: Enterprise storage systems deployed between 2019 and 2021 have reached operational end-of-support, forcing infrastructure modernizations.
- Component cost dynamics: Memory and high-density solid-state drive pricing trends are elevating bill-of-materials costs across high-capacity enterprise storage configurations.
Capital allocation is settling into an integrated systems view that weighs accelerators against storage and facility limits. The trackers behind these figures - IDC, BloombergNEF and Sightline Climate - restate them quarterly, so each number above is worth re-checking against its retrieval date before it enters a budget.
Frequently asked questions
How much will be spent on AI infrastructure in 2026?
Global AI infrastructure hardware spending is projected to reach $497 billion in the single year 2026. Goldman Sachs Research separately forecasts around $1 trillion of total global AI-related investment in the same year, on a wider definition that also captures datacenter construction, power and non-hyperscaler buyers.
What is the total projected cost of the global AI datacenter buildout?
There is no single agreed total, because construction spending runs over multi-year cycles and covers land, power and cooling rather than hardware. The clearest annual marker is company-level capital expenditure at the 14 largest listed datacenter operators, projected to approach $750 billion in 2026 against under $450 billion in 2025. Shell-and-core construction averages $11.3 million per megawatt, rising to $30 million to $40 million per megawatt for dense liquid-cooled AI facilities.
How much datacenter capacity is currently under construction?
At the end of September 2025, the latest global tracking cut, 23.1 GW of datacenter IT capacity was actively under construction across 831 sites globally. The United States dominates the pipeline with 15.9 GW of that capacity, while the EMEA region lags at 2.9 GW.
What is the InvestAI facility?
InvestAI is the financing vehicle the European Commission and the European Investment Bank launched to build AI gigafactories in the EU. It was designed around a 20 billion euro facility for up to five sites; the formal call the Commission opened on 30 July 2026 widened that to up to seven, with up to 10 billion euro of EU and national funding behind it.
Which server architecture is dominating AI infrastructure spending?
ARM-based rack-scale platforms have overtaken x86 servers in the accelerated compute market. In the first quarter of 2026, ARM accelerated server spend reached $53.0 billion, outpacing the $34.6 billion spent on x86 alternatives.
Are enterprises continuing to delay storage infrastructure investments?
No, the period of deferred storage investment is ending. Enterprises that paused traditional storage refreshes to fund initial AI GPU purchases are now resuming these cycles, padding the broader datacenter infrastructure spending forecast for 2026.
Sources
- goldmansachs.com
- AI Infrastructure Spending Holds Near $90 Billion in Q1 2026 as ARM Overtakes x86 in Accelerated Servers… idc.com
- AI Data Center Build Advances at Full Speed: Five Things to Know | BloombergNEF about.bnef.com
- Data center outlook: half of 2026 pipeline may not materialize | Currence Blog sightlineclimate.com
- Transformers in 2026: Shortage, Scramble, or Self-Inflicted Crisis? powermag.com
- emp.lbl.gov
- AI Data Center Construction 2026: Capex, Cost per MW, Delays archdesk.com
- EIB Group and European Commission join forces to finance AI gigafactories eib.org
- EU launches AI Gigafactories call to boost Europe's computing capacity and unlock more than €30 billion in… digital-strategy.ec.europa.eu
- AI Factories digital-strategy.ec.europa.eu
- Commission and European Investment Bank Group team up to support AI Gigafactories digital-strategy.ec.europa.eu
- Europese Commissie - Invest AI - Kenniscentrum Data & Maatschappij data-en-maatschappij.ai
- AI Demand Puts Enterprise Storage Back in the Spotlight, Says IDC hostingjournalist.com
Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score
Related reading
All Markets →AI infrastructure spending races toward $1 trillion
Understand the global AI infrastructure market size, including a 31 percent CAGR projection, 2029 spending forecasts, and server shipment volume growth.

Where GPU prices go from here
An analytical breakdown of the GPU price trend and AI infrastructure costs, exploring macroeconomic forecasts, generation shifts, and capital expenditure.

Is the CHIPS Act still in effect?
The US CHIPS Act is still law, but its fab funding runs only through FY2026 and the 35% chip tax credit needs construction to start by 31 December 2026.
