$380B. That's what four companies - Microsoft, Google, Amazon, and Meta - will spend on AI infrastructure in 2026 alone. That's more than the GDP of most countries. Whether this is the greatest infrastructure buildout in history or the most expensive capex cycle of all time depends on one question: will it get monetized?
In 2021, the four major hyperscalers combined to spend $118B on capital expenditures. By 2025, that number hit $325B. By 2026, it's projected to reach $380B - a 3x increase in five years, driven almost entirely by artificial intelligence. The chart below shows the composition of that spending by company.
Microsoft, Google/Alphabet, Amazon/AWS, Meta - 2021–2026E. Hover bars for breakdown.
What's striking about this chart isn't just the magnitude - it's the acceleration. The 2024–2025 step change ($142B to $227B to $325B) is unlike anything in the history of corporate capital allocation. Amazon's AWS alone will spend over $120B in 2026, more than the entire combined hyperscaler capex in 2021.
The key driver: GPU clusters. A single NVIDIA GB300 cluster at 100,000 chips costs north of $5B - and hyperscalers are ordering multiple clusters per quarter. This spending is not discretionary. It's competitive survival. If Microsoft slows and Google doesn't, Azure loses enterprise AI deals. The dynamic is self-reinforcing.
Every 1 GW of AI compute requires ~50 MW of power - the equivalent of 50,000 homes. The power infrastructure buildout is running two to three years behind the compute buildout.
No company in history has grown revenue as fast as NVIDIA's data center segment. From $4.3B in Q1 2023 to $52.3B in Q1 2026 - a 13-quarter run that analysts consistently underestimated at every single step.
Data center segment only. Hover data points for values.
NVIDIA data center revenue grew 1,100% from Q1 2023 to Q1 2026 - the fastest revenue ramp of any company with more than $4B in quarterly revenue, ever.
The Blackwell architecture (launched Q3 2024) was the key inflection. Blackwell chips are roughly 4x more efficient at inference than Hopper - which accelerated deployment because the economics of running AI at scale finally worked. The next generation, Blackwell Ultra and GB300, is expected to ship through 2026 with demand already exceeding supply, according to SemiAnalysis research.
The bear case on NVIDIA is straightforward: at a P/E above 35x, the market is pricing in perfect execution for several years. Any hyperscaler capex pullback, any breakthrough by AMD or custom silicon, and the multiple compresses sharply. The stock is simultaneously the most important and most dangerous in the AI trade.
SemiAnalysis estimates Blackwell chip demand exceeds supply through at least 2026, with lead times on GB300 clusters extending past 12 months for new orders.
AI infrastructure is a layered value chain. NVIDIA gets the headlines, but there are investment theses at every layer. Below is the full stack from silicon to applications - click any layer to see the key companies and their competitive moat.
Click any layer to expand the investment moat.
The honest answer is: both, in different parts of the stack. The infrastructure itself - power, cooling, networking, data centers - looks increasingly foundational. The valuation of compute companies, particularly NVIDIA and its supply chain, has bubble pockets. Below is the evidence on both sides.
Capex growing 3x faster than AI revenue for most companies
ROI of AI investments unclear - most LLM use cases not monetized
NVIDIA P/E >35x requires perfect execution for years
Concentration risk: 4 companies = ~80% of AI chip demand
Electricity consumption from AI is structural and growing
Every tech co building proprietary AI = can't rely on one vendor
Inference demand growing even faster than training (recurring)
Sovereign AI spending beginning - every country wants own AI capacity
Verdict: Foundation, but with bubble pockets in compute valuations
Bernstein's AI infrastructure team noted in their March 2026 report that the key distinction is between training and inference. Training capex (building the clusters to develop models) is the high-risk, potentially over-invested part. Inference (running models for users) is growing even faster and has clearer revenue per unit economics. Chips optimized for inference - NVIDIA's next-gen products, as well as custom silicon from Broadcom - are the next growth leg.
The "picks and shovels" framing - investing in infrastructure rather than outcomes - is compelling in AI, but requires precision. Not all infrastructure is equal. Here are the companies beyond NVIDIA with the strongest structural positions:
Custom AI chips (XPUs) for Google, Meta, and Apple. Each customer represents a multi-billion dollar multi-year program. As hyperscalers seek to reduce NVIDIA dependency, Broadcom is the primary beneficiary. AVGO's networking ASICs (Tomahawk, Jericho) are also inside most AI clusters.
Ethernet networking for AI clusters at scale. As AI clusters grow from thousands to hundreds of thousands of GPUs, the switching fabric becomes critical. Arista's EOS software platform and high-radix switches give it a durable edge. The shift from InfiniBand to Ethernet in some clusters directly benefits ANET.
Power management and liquid cooling for data centers. AI servers run at 10-30x the thermal density of traditional rack servers - conventional air cooling fails above ~100kW per rack. Vertiv's liquid cooling systems are already backlogged 18+ months. The power problem is structural and unsolved.
Nuclear power is being contracted directly for AI data center power. Microsoft's Three Mile Island restart deal (Constellation) and similar agreements signal that hyperscalers are going around the grid entirely. 24/7 carbon-free baseload power is exactly what AI data centers need.
Morgan Stanley's technology equity team (March 2026) highlighted Vertiv and Arista as their top infrastructure picks on the thesis that both have multi-year backlogs and pricing power that hasn't been fully appreciated by the market. AVGO trades at a discount to NVIDIA on a growth-adjusted basis despite a comparable exposure to AI capex.
The biggest underappreciated driver of AI infrastructure demand over the next five years may not be US hyperscalers - it may be governments. Every major economy is now building national AI compute capacity. The logic: AI is a strategic asset. Dependence on a single US cloud provider for critical AI infrastructure is untenable from a national security perspective.
$30B+ national AI compute program. G42 partnered with Microsoft in a landmark deal. Building sovereign LLM infrastructure.
NEOM AI cluster. $40B investment announced. Primarily through Project Transcendence - partnering with Humain (PIF subsidiary).
Mistral AI domestically funded. EU AI Act compliance driving national cloud compute. €1B+ in national AI infrastructure.
IndiaAI Mission: 10,000 GPU compute cluster funded by government. Bhashini AI program for local language models.
SemiAnalysis estimates sovereign AI spending will total over $100B over the next five years. This is incremental demand on top of US hyperscaler capex - and it almost exclusively flows through NVIDIA (the only supplier with chips at scale) and TSMC (the only foundry that can manufacture them). The geopolitical dimension also means this spending is sticky and insensitive to normal ROI calculations.
The most dangerous scenario for AI infrastructure investors is not a single shock - it's a capex coordination failure. If any two of the four major hyperscalers simultaneously pull back on AI spending (say, because of a recession, or because AI monetization disappoints), the entire semiconductor supply chain is exposed.
The math is unforgiving. NVIDIA's data center business is ~60% driven by hyperscaler purchases. If those orders decline 20%, NVIDIA revenues drop. Suppliers to NVIDIA - TSMC, SK Hynix, CoWoS packaging from ASE - all feel it within two quarters. The supply chain has expanded capacity on the assumption of continued hyperscaler demand. There is no graceful landing if demand disappoints.
Scenario to watch: Microsoft or Google guides capex down meaningfully in any single earnings call. That would be the signal that AI ROI expectations are being revised - and would trigger a re-rating of the entire infrastructure trade. Current consensus assumes capex continues growing through 2027.
Early monetization data is encouraging but thin. Microsoft Copilot is adding $12B ARR - ahead of expectations and the first AI product at genuine enterprise scale. Google AI Overviews serves over 1 billion users per month, though the revenue model remains indirect (search retention). AWS Bedrock is growing faster than AWS overall. These are data points that argue against the pure bubble interpretation - but three data points is not a trend.
Sources: Microsoft, Alphabet, Meta, and Amazon 10-K and 10-Q filings; NVIDIA quarterly earnings and investor materials; SemiAnalysis newsletter and datacenter capex tracker; Bernstein Research semiconductor coverage; Morgan Stanley Cloud Capex Tracker; Dell'Oro Group AI networking forecasts; IDC hyperscaler infrastructure reports.
This article is for informational purposes only and does not constitute investment advice. Estimates marked "E" are consensus projections and subject to revision. skopa does not hold positions in any securities mentioned.
Microsoft, Google, Amazon and Meta are projected to spend about $380B on capital expenditure in 2026, versus $325B in 2025, $227B in 2024 and $118B in 2021. Amazon’s AWS alone accounts for over $120B in 2026.
The spending is real and contractually committed rather than speculative — it is driven by competitive necessity, since a hyperscaler that slows while rivals do not loses enterprise AI deals. The genuine risk is not fake demand but physical constraint: power and grid capacity are running two to three years behind the compute build-out.
Roughly 50 MW of power per 1 GW of AI compute capacity — about the consumption of 50,000 homes. Power infrastructure is being built two to three years behind the compute it has to serve, which makes electricity the practical ceiling on AI deployment.