The hyperscaler capex cycle, explained simply — why it keeps coming up and what it means for your portfolio

Every earnings call in 2025 and 2026 has had the same question. An analyst asks Meta, Microsoft, Alphabet, or Amazon how much they plan to spend on AI infrastructure. The executive answers with a number that sounds impossibly large. The stock moves. The cycle repeats.

If you've been watching this from the outside and wondering what it all means for your US portfolio — here is the plain-language version.

What "hyperscaler capex" actually means

Hyperscalers are the four cloud giants building the backbone of AI: Microsoft Azure, Google Cloud, Amazon AWS, and Meta AI. They're called hyperscalers because they operate data centers at a scale nobody else does — tens of thousands of servers, entire city blocks of power infrastructure, global fibre networks.

Capex — capital expenditure — is money spent on physical stuff. Buildings, servers, power systems, cooling equipment. It is the opposite of opex (operating expenditure), which covers salaries, software, and day-to-day costs. When a hyperscaler says it's spending $75 billion in capex, it means $75 billion going into physical infrastructure this year.

The hyperscaler capex cycle is what happens when all four of them do this simultaneously, at historically unprecedented scale, because all four have concluded that AI is the defining infrastructure investment of the decade.

The actual numbers

Here is the combined FY2026 picture:

Company	FY2026 capex or backlog	What it represents
Meta	$125–145B (full-year guide, raised twice from original $60–65B)	Buildings, servers, GPU clusters for Llama training and AI advertising
Alphabet	$460B Google Cloud contracted backlog	Enterprise customer commitments — revenue already contracted, not yet recognised
Microsoft Azure	$80B backlog it literally cannot fulfill	Power constraints are preventing Azure from building fast enough to meet signed contracts
Amazon AWS	Not separately broken out	Visible in long-term purchase obligations in filings

Combined FY2026 hyperscaler capex: approximately $370 billion.

To put $370 billion in context: at roughly ₹31 lakh crore, it exceeds India's entire Union Budget for fiscal year 2026. It is larger than the GDP of over 150 countries. It is, by most measures, the largest peacetime capital deployment in corporate history — concentrated in a single technology theme, across a single calendar year.

Meta's guide tells the story clearly. The original 2026 capex guidance was $60–65 billion. It has since been raised twice. It now sits at $125–145 billion. That is not a rounding error. It is a doubling of commitment in under twelve months.

Why it keeps coming up

The reason analysts ask about capex every single quarter is that this number is the leading indicator for an entire industry.

Every dollar of hyperscaler capex flows somewhere. GPUs, networking, power systems, cooling, servers, data center construction. The companies that make those things — NVIDIA, ASML, TSMC, Vertiv, Eaton, Broadcom — have revenues that are direct functions of whether $370 billion of capex actually gets spent.

When Meta doubles its capex guide, NVIDIA's forward revenue estimate goes up. When Microsoft says it can't build fast enough due to power constraints, Vertiv's backlog expands. The chain is direct and quantifiable.

That is why the number matters beyond the hyperscalers themselves.

The transmission mechanism — layer by layer

Layer 1 — GPUs (most direct): NVIDIA sells the processors that go into every AI data center. $370B of hyperscaler capex means demand for tens of millions of Blackwell GPUs across FY26–27. NVIDIA's Q1 FY27 data center revenue was $75.2 billion — in a single quarter, up 92% year-over-year. That number is the direct output of this capex cycle.

Layer 2 — Chip manufacturing equipment: ASML makes the EUV lithography machines that produce the silicon that goes inside those GPUs. You cannot build a Blackwell chip without ASML machines. ASML's order book stood at €38.8 billion at year-end 2025. The order book doesn't care what quarter earnings prints at — those are committed orders.

Layer 3 — Foundries: TSMC manufactures the chips. It raised its own capex to $52–56 billion for 2026, in direct response to hyperscaler demand. TSMC's revenue grows when hyperscalers buy more chips.

Layer 4 — Power and cooling: Data centers consume enormous amounts of electricity and generate enormous heat. Vertiv (cooling systems, $15B backlog) and Eaton (electrical infrastructure, 228 GW order book) are direct beneficiaries. Their backlogs are orders already placed — not projections.

Layer 5 — Custom silicon: Broadcom and Marvell build custom AI chips (ASICs) that hyperscalers are developing in-house to reduce their dependency on NVIDIA. Google's TPU, Meta's MTIA — both are Broadcom's work. This is a layer that grows as hyperscalers diversify away from NVDA.

What is not in the transmission: Most SaaS companies. Consumer internet names. E-commerce. Retail. The capex cycle benefits the physical infrastructure stack, not companies that sit above it as software.

The bear case — what if the cycle ends?

The honest version of this question.

The Google Cloud backlog of $460 billion is contracted revenue. Enterprise customers have already signed multi-year commitments. That number does not evaporate overnight. Microsoft's $80 billion backlog exists because customers have signed and Microsoft cannot build fast enough. These are not projections — they are obligations.

But the risk is not sudden stop. The risk is capex plateau — where the growth rate decelerates from 40–50% year-over-year to 10–15%. The spending continues but the acceleration slows. That scenario still compresses NVIDIA's earnings multiple meaningfully even if absolute revenues keep growing.

The historical analogy that gets cited is the 2000 fibre-optic overbuild. Telecom companies built fibre infrastructure far beyond near-term demand. Companies with two-year order books felt insulated — right up until the inventory correction hit everything simultaneously. Companies with contracted backlog still saw revenue collapse when those contracts were renegotiated or counterparties went bankrupt.

The reason that analogy doesn't map cleanly onto 2026: the $460B Google Cloud backlog represents demand from enterprise customers paying for cloud services, not supply-side speculation. AI is generating revenue for these hyperscalers right now. That is different from laying fibre in anticipation of demand that never materialised.

But valuation compression can happen even when fundamentals hold. If NVIDIA trades at 40x forward earnings and capex growth slows from 50% to 15%, the multiple compresses to 25–30x. Revenues may still grow; the stock may still fall. This is the risk, stated plainly.

What this means if you hold these names

If you own NVDA, MSFT, GOOGL, or META directly: your returns are partly a bet on the capex cycle sustaining. Q2 earnings starting July 22 will give the first H2 read on whether pace is accelerating or plateauing. Watch specifically whether hyperscalers raise, maintain, or cut their FY2026 capex guidance. A raise is positive for the whole semiconductor stack. A cut is a meaningful negative. In-line is already priced in.

If you own QQQ or a Nasdaq index ETF: you have meaningful indirect exposure. NVIDIA is QQQ's largest or second-largest holding. Microsoft, Alphabet, and Meta are all top-10 positions. The index is not diversified away from this theme.

If you own VTI (total US market): the exposure is smaller but present. NVIDIA, Microsoft, Alphabet, Meta, and Amazon together represent roughly 25–30% of VTI by weight.

If you hold PPFAS Flexi Cap or a Nasdaq FoF feeder fund: you have exposure through the fund manager's underlying positions, with an additional layer of TER drag that compounds over time. The fund takes the same risk; you pay for the management wrapper.

One tax point worth stating: if you are sitting on a large unrealised gain in NVDA, MSFT, or GOOGL and you have not yet crossed 24 months of holding, the capex cycle risk is not necessarily an argument to sell. Selling before 24 months crystallises short-term capital gains taxed at your income-tax slab rate. The decision to exit needs to weigh the capex risk against the cost of premature realisation. That calculation is different for someone at the 30% slab versus someone at a lower bracket.

The one number to watch in Q2 earnings

Q2 earnings start July 22. The specific thing to watch is not revenue — it is whether hyperscalers raise or maintain their FY2026 capex guide.

Meta guided $125–145B. If that goes to $150B+, the cycle is accelerating. If it stays flat, the cycle is sustaining. If it drops — that is the signal.

Microsoft and Alphabet report July 28–30. Meta reports July 29–31. NVIDIA's fiscal Q2 FY27 results arrive August 20–22.

The August 20 NVIDIA print is the terminal event for the near-term cycle read. Everything before it is context; that is the number that determines how semiconductor multiples trade into year-end.

---

FAQs

What is hyperscaler capex?

Capital expenditure by the four large cloud companies — Microsoft Azure, Google Cloud, Amazon AWS, and Meta AI — on physical infrastructure: data centers, servers, GPUs, power systems, and cooling. In FY2026, the combined total is approximately $370 billion. "Hyperscaler" refers to their ability to operate at scales no other company can match.

Why does hyperscaler capex matter for semiconductor stocks?

Because GPUs, chips, networking equipment, and power systems are what the capex buys. NVIDIA's data center revenue of $75.2 billion in a single quarter is the direct output of hyperscaler spending. ASML's €38.8 billion order book and TSMC's $52–56 billion own capex are downstream responses to the same demand. When hyperscaler capex grows, semiconductor revenue follows — with roughly one to two quarters of lag.

Can the capex cycle end suddenly?

Not cleanly. The Google Cloud backlog of $460 billion and Microsoft Azure's $80 billion unfulfilled commitments are contracted obligations, not projections. They don't disappear overnight. What can happen is a deceleration — where growth rate slows from 40–50% year-over-year to 10–15%. That is enough to compress earnings multiples on high-valuation names like NVIDIA significantly, even if revenues continue growing.

How does this affect an Indian retail investor specifically?

If you hold US tech stocks directly — NVDA, MSFT, GOOGL, META, AMZN — a meaningful portion of your return is tied to whether this capex cycle sustains. If you hold index funds like QQQ or VTI, you have indirect exposure through the weights of those names. If you hold Indian feeder funds tracking Nasdaq, you have the same exposure plus TER drag. On the tax side: gains from direct US stock holdings are taxed as foreign equity; short-term (under 24 months) at your income slab, long-term at 12.5% without indexation. Capex cycle risk should factor into whether you crystallise gains before reaching long-term status.

What happens if AI capex disappoints?

The transmission runs in reverse. Lower hyperscaler capex means lower GPU demand, which means lower NVIDIA revenue and significant multiple compression. The names most exposed are those with highest hyperscaler customer concentration: Vertiv (70%+ of AI revenue from hyperscalers), Arista Networks (≈60%), and Broadcom's custom silicon division (≈95% hyperscaler revenue). NVIDIA is less exposed than it appears — sovereign AI buyers, enterprise customers, and consumer GPU demand provide partial offset. DELL and HPE, which are more diversified across enterprise and consumer, have the most protection. Index holders take the weighted average of all of this.

---

Cross-references

• The AI capex stress test: what actually breaks at -20% — name-by-name math on the downside scenario
• Semiconductors entering earnings season: 30 stocks, 6 layers — the order-book data and model portfolios going into Q2 results

For Indian residents on tax and disclosure:

• How RSU double-taxation actually works — tax framework for US equity income
• Schedule FA disclosure guide — disclosure obligations for these positions

---

This article reflects the AI infrastructure landscape as of July 13, 2026. All figures are sourced from company earnings calls and filings through Q1 2026. Stock multiples and prices move daily; forward P/E references are illustrative. This article does not constitute investment advice. Consult a SEBI-registered investment adviser before making portfolio decisions.

The hyperscaler CapEx figures in detail: AWS, Azure, GCP, OCI

The four hyperscalers — Amazon AWS, Microsoft Azure, Google Cloud (GCP), and Oracle Cloud Infrastructure (OCI) — are spending at different absolute scales but with one common direction: up.

Amazon AWS does not break out CapEx separately from the parent company, but Amazon's total CapEx for 2024 came in at approximately $75 billion, with AWS infrastructure representing the majority. For 2025–2026, Amazon has signalled continued acceleration, with total CapEx expected to exceed $100 billion across data centres, servers, and custom silicon (Trainium, Inferentia).

Microsoft Azure announced approximately $60 billion in AI infrastructure CapEx for fiscal year 2025 (ending June 2025), with further increases expected in FY2026. The $80 billion Azure backlog — demand it literally cannot fulfill due to power constraints — indicates supply-constrained growth, not demand-constrained.

Google Cloud (GCP) guided $75 billion in CapEx for 2025 alone, reflecting a step-change acceleration from prior years. The contracted backlog of $460 billion represents multi-year enterprise commitments already signed — the largest revenue visibility of any hyperscaler.

Oracle Cloud Infrastructure (OCI) is the fastest-growing in percentage terms, starting from a smaller base. Oracle's remaining performance obligations (RPO) — its contracted backlog — exceeded $130 billion as of early 2026, driven by large AI training contracts including deals with OpenAI and government entities. OCI's CapEx growth rate has exceeded 100% year-over-year in recent quarters.

The total combined CapEx across the four hyperscalers for FY2026 is approximately $370 billion. Oracle's growth rate is the fastest in percentage terms; Meta's absolute raise (from $60B to $125–145B) is the most dramatic single-year revision.

Why this matters for Indian RSU holders specifically

Indian employees holding RSUs at Microsoft, Amazon, Google, or Oracle are directly exposed to this cycle — not as investors choosing to participate, but as employees whose equity compensation is tied to their employer's stock price.

Microsoft RSU holders: Azure CapEx drives Microsoft's revenue growth expectations. If Azure revenue misses — because enterprises pause AI commitments — Microsoft's stock reprices. Microsoft RSU holders at Indian offices (Hyderabad, Bengaluru, Noida) are directly exposed.

Amazon RSU holders: AWS is Amazon's profit engine. The CapEx cycle risk translates directly into RSU mark-to-market. Indian Amazon employees (primarily Hyderabad, Bengaluru) have material RSU exposure.

Google RSU holders: Alphabet's $460B backlog provides substantial revenue visibility, but the stock trades on AI execution — whether GCP takes share from Azure and AWS. Indian Googlers (Hyderabad, Bengaluru, Gurugram) hold equity directly tied to this outcome.

Oracle RSU holders: Oracle's fastest percentage CapEx growth and the OpenAI deal make OCI the most momentum-driven of the four. Indian Oracle employees may hold significant unvested equity tied to OCI execution.

For RSU holders at these companies: the capex cycle risk is not theoretical. It is priced into your unvested equity. Understanding which way the cycle turns — and when to diversify away from concentrated single-employer exposure — is a material financial decision.

How to position: NVDA for direct exposure, CSPX for diversification

For direct hyperscaler-cycle exposure: NVIDIA (NVDA) is the most efficient single-stock expression of the CapEx cycle. NVIDIA's data center segment revenue is the direct output of hyperscaler GPU purchases. When hyperscalers raise CapEx guides, NVIDIA's forward revenue estimates go up within days. The stock is high-beta to the cycle — it outperforms when the cycle accelerates and underperforms when it slows.

For suppliers further down the stack: ASML (lithography), TSMC (foundry), Broadcom (networking ASICs), and Vertiv (power and cooling) all have significant hyperscaler revenue — but each has diversification across other customers and use cases. They are lower-beta expressions of the same theme.

A note on TQQQ: TQQQ (ProShares UltraPro QQQ) is a 3× daily leveraged ETF tracking the NASDAQ-100. It is not an appropriate vehicle for Indian investors doing buy-and-hold US equity investing. Daily rebalancing causes compounding decay (volatility drag) that destroys returns over multi-month periods in volatile markets. In a sideways but volatile market, TQQQ can lose value even while QQQ is flat. TQQQ is designed for very short-term traders with daily holding horizons, not for long-term wealth building in an LRS portfolio.

For diversified tech-weighted exposure: CSPX (iShares Core S&P 500 UCITS ETF, Ireland-domiciled, LSE-listed) gives you S&P 500 exposure including NVIDIA, Microsoft, Alphabet, Amazon, and Meta as top holdings. It captures the hyperscaler cycle through index weights without single-stock concentration risk. No US estate tax exposure (Ireland-domiciled), accessible to UAE NRIs, UK NRIs, and Singapore NRIs through their local brokers.

The CapEx cycle risk: if revenue doesn't materialise

The bear case for the hyperscaler cycle is not that the spending stops — it is that the revenue does not materialise to justify the investment. The distinction matters.

Google has $460B in contracted cloud revenue. Microsoft has $80B in unfulfilled orders. These are signed contracts, not aspirations. The spending has real demand behind it.

The risk scenario is slower-than-expected AI adoption at the enterprise level. If Fortune 500 companies sign enterprise AI contracts but don't actually deploy at scale — if the business case for AI doesn't close in 2026–2028 — contract renewals slow, new signings slow, and the backlog growth rate collapses even while existing commitments are honored. Hyperscaler CapEx decelerates, GPU orders slow, and NVIDIA's forward revenue estimates come down.

This is the scenario that would cause a significant correction in semiconductor stocks without a "crisis" — just a quieter demand plateau. NVIDIA at 30–40x forward earnings absorbs this plateau badly even if revenues still grow at 20–25% instead of 50–60%.

For Indian investors holding these names: the risk is not binary (crash vs fine). It is valuation compression on sustained but slower growth — which can mean a flat or declining stock price even while the company reports record revenues. Understanding this distinction changes how you think about position sizing and time horizon.

Cross-references

• The AI capex stress test: what actually breaks at -20% — name-by-name math on the downside scenario
• RSUs and ESPP tax in India — for RSU holders at hyperscaler companies
• Schedule FA disclosure guide — disclosure obligations for US stock positions

---

This article reflects the AI infrastructure landscape as of July 2026. Figures sourced from company earnings calls and filings. This is not investment advice. Consult a SEBI-registered investment adviser before making portfolio decisions.