· May 2026
Technology · AI Infrastructure · How to ETF it

The hyperscalers are spending
trillions —
who profits?

Microsoft, Google and Amazon are investing record amounts in data centres. But that money doesn't disappear — it flows to an entire ecosystem of companies. Who are the real winners, and how can you as an ETF investor benefit?

Combined CAPEX 2025
$325bn
Microsoft, Google, Amazon and Meta combined — almost doubled in two years
Expected Growth 2026
+38%
Hyperscalers increase budgets once more — no end to the cycle in sight
AI Datacenter Market 2030
$1.5tn
Expected market size, driven by GenAI and cloud migration

What is really happening

In the nineteenth century, thousands of people rushed to California to search for gold. Most barely earned enough to pay back their tools. But the men who made the jeans, forged the shovels and sold the provisions — they became rich. That age-old principle, picks and shovels investing, is today the key to understanding the largest infrastructure investment in human history.

Microsoft, Google (Alphabet), Amazon (AWS), Meta and a handful of other hyperscalers have embarked on a capital expenditure cycle without precedent. In 2025, the five largest tech giants combined spent more than $325 billion on data centres, servers, networking equipment and energy infrastructure. By comparison: the total annual investment budget of the entire European energy sector amounts to less than half of that.

The driver is clear: generative AI requires enormous amounts of computing power. Training a single AI model costs tens of millions of dollars in computing time. Deploying AI services to hundreds of millions of users requires infrastructure that must be continuously expanded. Demand is growing faster than supply can be built.

Company CAPEX 2025 Focus 2026 Announcement
Microsoft
$80bn
Azure + OpenAI infrastructure
AI computing, GPU clusters, cloud expansion "We see no end to demand" — CFO Amy Hood
Alphabet (Google)
$75bn
Google Cloud + TPU chips
Own AI chips (TPU), quantum computing, Search AI Budget increased after rising Cloud revenue
Amazon (AWS)
$105bn
World's largest cloud provider
AWS capacity, own Trainium and Inferentia chips Backlog of customer orders growing faster than construction
Meta
$65bn
AI Research + social platforms
LLaMA models, Reels recommendation systems "We are building the largest AI infrastructure ever"

The anatomy of a data centre

A data centre is not a building with a few computers in it. It is a complex ecosystem of specialised technology, energy infrastructure and cooling systems. If you understand where the billions are flowing, you also understand which companies benefit most.

🪙 Where does CAPEX money flow?
🖥
GPUs and AI chips — the most expensive item (~35–40% of CAPEX)
Each AI server contains dozens of GPUs at $30,000–$40,000 each. Demand structurally exceeds supply.
Nvidia · AMD · Intel · TSMC (fabrication) · ASML (chip equipment)
38%
Energy and cooling (~20–25% of CAPEX)
AI servers consume ten times more power than regular servers. Cooling is the biggest operational challenge — water-cooled systems, backup generators, transformers.
Eaton · Vertiv · Schneider Electric · Quanta Services · Emerson
22%
🌐
Networking equipment (~15% of CAPEX)
High-speed connections between servers, switches, fibre optic cables and connections to the internet. Bandwidth is becoming as critical as computing power.
Arista Networks · Cisco · Broadcom · Corning (fibre optics)
15%
🏗
Buildings and civil infrastructure (~15% of CAPEX)
Land, concrete buildings, security infrastructure and fit-out. Often outsourced to specialised data centre REITs or contractors.
Equinix · Digital Realty · Iron Mountain · Prologis
15%
💾
Storage and other hardware (~10% of CAPEX)
Solid-state drives, RAM memory, backup systems. AI workloads require much faster and larger storage than traditional workloads.
Samsung · SK Hynix · Micron · Western Digital · Seagate
10%

Picks & Shovels: the infrastructure layer

The hyperscalers themselves are of course direct beneficiaries of the AI wave — but they also bear enormous costs. The structurally most attractive position is often not the gold seeker themselves, but the supplier of the shovel. In investor terms: the infrastructure layer benefits from the investment cycle regardless of which AI model ultimately wins.

Nvidia is the most discussed example. The company sells its H100 and B200 GPUs for $30,000 to $40,000 each and delivers them in clusters of thousands. Gross margin exceeds 75% — a level of profitability that is virtually unmatched in the industry. But Nvidia is no longer a hidden gem: the stock is included in virtually every broad technology ETF.

More interesting are the less well-known layers of the ecosystem. Companies such as Vertiv (data centre cooling), Arista Networks (networking equipment) and Eaton (power management) structurally benefit from every dollar a hyperscaler spends — regardless of whether it is a Microsoft or Google data centre, and regardless of which AI model runs in it.

✓ The picks & shovels logic
The risk of the gold seeker: he can fail. The risk of the shovel seller: almost everyone buys his shovel. With AI the same applies. Which AI model dominates in ten years is uncertain — that GPUs, cooling systems and fibre optics are needed is a certainty.

The power problem nobody expected

The fastest-growing bottleneck in the entire AI infrastructure cycle is not a chip and not software — it is electricity. A 1-gigawatt data centre consumes as much electricity as a mid-sized city. The demand for power from data centres is growing so fast that existing electricity grids in the US and Europe cannot keep up with capacity.

Microsoft therefore struck a deal to restart the nuclear plant at Three Mile Island — closed for decades after an incident in 1979. Google purchased nuclear energy from startups building micro-reactors. Amazon has signed contracts for geothermal energy in Iceland. The energy transition and the AI infrastructure revolution have become intertwined.

This makes companies active in energy infrastructure — high-voltage cables, transformers, backup generators and cooling systems — unexpected beneficiaries of the AI boom. Vistra, GE Vernova and Constellation Energy all multiplied in stock market value in 2025, purely on the basis of expected power demand from data centres.

"We expect data centres to consume more than 8% of all electricity in the US by 2030. In 2023, that was still 2.5%."
Goldman Sachs — AI Power Demand Report, 2024

Five layers to invest in

The data centre gold rush offers multiple entry points for the ETF investor. The key is understanding which layer you want exposure to — and how that fits within a broader, diversified portfolio. Always use these themes as a satellite position (max. 5–10% per theme) on top of a broad core ETF.

Layer 1 — Chips & Semiconductors
The engine of the AI machine
SMH — VanEck Semiconductor ETF
Nvidia, TSMC, Broadcom, ASML, AMD
The most direct exposure to GPU demand. Top holding Nvidia weighs ~20%.
TER 0.35% · >$25bn AUM
SOXX — iShares Semiconductor ETF
More evenly spread across chip designers and manufacturers
Less Nvidia concentration than SMH, broader across the value chain. UCITS variant: SEMG on Euronext.
TER 0.35% · UCITS variant: SEMG on Euronext
ASML — Direct (no ETF needed)
Monopolist in EUV lithography machines — indispensable for chip manufacturing
Every data centre GPU in the world starts at ASML. Listed on Euronext Amsterdam.
Stock — included in AEX and MSCI World
WTAI — WisdomTree AI ETF
Broader AI exposure incl. chips, software and infrastructure
UCITS-listed, suitable for European investors.
TER 0.40% · IE-listed
Layer 2 — Energy & Cooling Infrastructure
The forgotten bottleneck of AI
POWR — iShares US Power Infrastructure
Energy infrastructure: utilities, transformers, networks
Directly exposed to rising power demand from AI data centres.
TER 0.30% · one of the fastest-growing energy ETFs of 2025
GRID — First Trust Nasdaq Clean Edge Smart Grid
Smart energy grids, transformers, infrastructure modernisation
Companies such as Eaton, Vertiv and Quanta Services — the "picks" of the energy boom.
TER 0.58% · US-listed
VDE — Vanguard Energy ETF
Broader energy sector as hedge on growing power demand
Less direct but cheaper and more broadly diversified. Includes nuclear energy companies.
TER 0.10% · lowest cost in the category
HEATR — Global X Data Center REITS & Digital Infra
Data centre real estate: Equinix, Digital Realty, Iron Mountain
REITs that rent power and space to hyperscalers — stable cash flows.
TER 0.50% · combination of infra + real estate
Layer 3 — Broad Technology (incl. hyperscalers)
The hyperscalers themselves in one ETF
QQQ / EQQQ — Nasdaq-100 ETF
Microsoft, Amazon, Alphabet, Meta, Nvidia — all in one
The most direct route to all hyperscalers combined. EQQQ = UCITS version.
TER 0.20% (EQQQ) · highest liquidity
IUIT — iShares S&P 500 Information Technology
S&P 500 technology sector — the IT column only
Broader than Nasdaq-100: also includes hardware, software and IT services. UCITS-listed.
TER 0.15% · IE00B3WJKG14
IGV — iShares Expanded Tech-Software
Enterprise software: the customer side of AI adoption
Microsoft Azure, Salesforce, ServiceNow — the companies that resell AI to businesses.
TER 0.43% · recovering after correction early 2026
SKYY — First Trust Cloud Computing ETF
Pure cloud exposure: AWS, Azure, GCP and satellite players
The revenue growers on the demand side of data centre CAPEX.
TER 0.60% · focus on cloud migration theme
Layer 4 — Networks & Connectivity
The highways between data centres
CIBR — First Trust NASDAQ Cybersecurity ETF
Network security: Palo Alto, CrowdStrike, Zscaler
More data centres = more attack surface. Security grows with every investment.
TER 0.60% · >$10bn AUM
ARGT / direct exposure
Arista Networks — data centre switches market leader
Broadcom — network chips and connectivity solutions
Both companies feature at the top of SMH and Nasdaq-100 ETFs.
Indirect via existing tech ETFs
📌 UCITS tip for European investors
Many of the ETFs mentioned are listed on US exchanges and may formally not be actively offered to European retail investors (MiFID II). Always look for the UCITS equivalent: it starts with IE (Ireland) or LU (Luxembourg) in the ISIN and is listed on Euronext, Xetra or the London Stock Exchange. Examples: EQQQ (Nasdaq-100 UCITS), SEMG (semiconductors UCITS), WTAI (AI UCITS).

What can go wrong?

The data centre investment cycle is real and structural — but that does not mean investing in it is risk-free. There are three risks that are systematically underweighted in the enthusiasm surrounding AI.

Risk Why it matters How to mitigate it
Overcapacity If AI demand disappoints or remains concentrated in one model, there will be billions in unused data centres. This happened to the telecom sector after the Y2K era. Invest in infrastructure layers that are used regardless of overcapacity: energy, cooling, networks.
Concentration Many AI ETFs have 20–30% in Nvidia alone. If the stock falls, the fund falls with it — even if the rest of the market rises. Spread across multiple layers of the ecosystem. Choose equal-weight ETFs or combine multiple funds.
Valuation AI-related stocks have already risen sharply. The market has priced in high growth. Disappointments — in revenue, margins or adoption — can lead to sharp corrections. Use dollar-cost averaging: invest monthly rather than all at once. Keep the position small (5–10% of portfolio).
Regulation AI legislation, chip export restrictions (US–China) and energy regulation can delay or reverse investment plans. Broad ETFs automatically spread this risk; avoid pure single-country exposure to geopolitically sensitive markets.
⚠ Don't forget the core ETF
All thematic exposure to the data centre cycle works best as a complement to a broad core ETF — such as iShares Core MSCI World or Vanguard FTSE All-World. That broad ETF already contains exposure to Microsoft, Amazon, Alphabet and Nvidia. Thematic ETFs add concentration, not automatically extra return. Keep the satellite position small and the horizon long.

The conclusion: shovel, not gold

The hyperscalers are spending hundreds of billions on data centres. That money flows to an entire ecosystem of companies — chipmakers, cooling specialists, energy companies, network suppliers and real estate parties. That ecosystem is broader, less well-known and in many cases more attractively priced than the hyperscalers themselves.

As an ETF investor you have multiple entry points: from broad technology ETFs that contain the hyperscalers themselves, to targeted semiconductor and energy ETFs that capture the infrastructure layer. The picks-and-shovels principle suggests that precisely this latter category is structurally interesting — less dependent on who wins the AI race, more dependent on the fact that the race is being run at all.

And that conclusion is certain: the race has begun, the budget has been approved, and the shovels have been ordered. Seeking the gold is risky. Selling the shovel is the business model of the century.

Disclaimer — This article is intended solely for educational purposes and does not constitute investment advice. ETFs and stocks mentioned are illustrative; always check the current TER, ISIN and suitability for your situation. Many US ETFs are not available to European retail investors — look for the UCITS equivalent. Past performance is no guarantee of future results. Consult a financial advisor for personal advice.

Sources: Goldman Sachs AI Power Demand Report (2024), Microsoft/Alphabet/Amazon/Meta Q4 2025 earnings reports, VanEck, iShares, Global X fund documentation, McKinsey Global Institute, May 2026.