The AI capex stress test: what actually breaks when Meta and Google cut capex 20%

The Mag-7 hyperscaler AI capex run rate hit approximately $370 billion in fiscal 2026. Microsoft $80B+, Google $75B+, Meta $72B (raised from $65B initial guide), Amazon $105B (mostly AWS-tied), Oracle $25B, and Apple about $15B. Add in the smaller hyperscalers (CoreWeave, Lambda, others), private cloud, and sovereign AI buyers, and the total addressable capex pool exceeds $450 billion globally.

This is the most concentrated capex cycle in modern stock-market history. The top 4 customers — Microsoft, Meta, Google, Amazon — account for somewhere between 40% and 60% of revenue for the AI infrastructure cohort. NVDA, AMD, AVGO, VRT, ANET, MU, SMCI, DELL, HPE — their fortunes are tied to whether four boards of directors keep approving annual capex increases.

The honest question every Indian retail investor with US AI exposure should be asking: what actually breaks when this cohort normalizes? Not "what if AGI doesn't happen" — that's the speculative version. The practical version: if Meta cuts FY27 capex from $90B guidance to $72B (a 20% cut), what does that mean for VRT, ANET, MU, SMCI?

This article does that math name by name. The result isn't a doom scenario. It's a calibration exercise: which AI-infrastructure names have operational flexibility to absorb a 20% hyperscaler capex cut, which have revenue cliffs that pass directly through, and which have multiple compression risk even if revenues hold.

The closing read: NVDA, AVGO, MU survive a 20% cut with multiple compression but no operational catastrophe. VRT, ANET, SMCI, DELL face revenue + multiple compression — the worst-of-both outcome. HPE and a handful of diversified names provide downside protection.

The capex math — where the $370B actually goes

Of the ~$370B in hyperscaler AI capex for FY2026:

Category	Approximate share	Primary beneficiaries
GPUs / accelerators	40-45% ($150-170B)	NVDA, AMD, AVGO custom silicon, MU (HBM)
Networking	10-12% ($40-45B)	ANET, NVDA (InfiniBand), Cisco
Power & cooling	12-15% ($45-55B)	VRT, ETN (Eaton), schneider electric (private), liquid cooling vendors
Servers & systems	15-18% ($55-65B)	DELL, HPE, SMCI, Lenovo
Storage	5-7% ($18-26B)	PSTG (Pure Storage), Dell, HPE
Land + construction	10-12% ($35-45B)	Real estate, construction (not stock-investable cleanly)

The total adds slightly above $370B because some categories overlap. The structure shows where every additional dollar of hyperscaler capex flows.

The customer concentration by stock:

Stock	Top 4 hyperscaler customer concentration	Sensitivity to MSFT/META/GOOGL/AMZN spend
NVDA	~45-50% (estimated; not separately disclosed)	High but diversified by application
AMD	~30-35% (MI300/MI355 ramp + EPYC server CPU)	High
AVGO	~50-55% (custom silicon for Google TPU, Meta MTIA, OpenAI, ByteDance)	Very high — direct
VRT (Vertiv)	~70%+ of AI/HPC revenue	Very high — direct
ANET (Arista)	~60% (META + MSFT primary)	Very high — direct
MU (Micron)	HBM3e exposure ~25-30% to AI chips; NVDA largest single	Moderate
SMCI	~40-50% concentrated in top hyperscalers + AI cloud	High
DELL	~30-35% of total revenue; AI server segment higher concentration	Moderate-High
HPE	~20-25% of total revenue; less hyperscaler-tied	Moderate
PSTG	~15-20% AI-tied; broader enterprise diversification	Lower

The pattern: VRT, ANET, AVGO are most exposed. MU and DELL moderate. HPE and PSTG least exposed. NVDA highly exposed but with workload-mix diversification.

The hyperscaler decision tree

To understand what triggers a capex cut, walk through each hyperscaler's capex driver:

Microsoft ($80B+ FY26 capex)

Drivers: Azure AI services + OpenAI partnership + Office 365 Copilot + enterprise customer commitments. The exposure: about 70% AI-related, 30% traditional cloud/data center.

Trigger for a 20% cut: OpenAI revenue growth slows materially, OR enterprise Copilot adoption disappoints, OR Microsoft chooses to focus on margins over volume. Recent Microsoft commentary has been consistent that capex tracks revenue growth — if OpenAI grows 60% YoY, capex follows; if growth slows to 30%, capex flexes.

Probability of FY27 capex cut: moderate. Microsoft has the cleanest unit economics among hyperscalers. The OpenAI exclusivity expires soon. Margin pressure is real.

Google/Alphabet ($75B+ FY26 capex)

Drivers: Gemini training + inference + TPU buildout + Search AI overlay + Google Cloud. Roughly 50% AI-specific, 50% broader Google Cloud + Search.

Trigger for a 20% cut: Gemini ROI not materializing in Google Search ad revenue, OR shareholder activist pressure on capex levels (recall Bill Nye-style activist letter could land), OR Sundar Pichai succession decision. Google has been the most aggressive capex grower (+78% YoY into 2026), creating the most room for normalization.

Probability of FY27 capex cut: moderate-to-high. Among Mag-7, Google has the most "catch-up" capex (was behind from 2023-2024 perspective). That dynamic has played out. Normalization is plausible.

Meta ($72B FY26 capex, raised from $65B)

Drivers: Llama training + advertising AI + Reality Labs. Meta is the most concentrated "training" buyer — uses GPUs primarily for LLM training rather than enterprise inference.

Trigger for a 20% cut: Reality Labs cost-cutting pressure (still loss-making $12B annually), OR Mark Zuckerberg's "year of efficiency 2.0" decision (recall 2023 layoffs), OR Llama 5 generation delayed/underperforming.

Probability of FY27 capex cut: moderate-to-high. Meta has the strongest history of capex flexibility. The 2023 efficiency narrative shows the playbook. Q4 2025 commentary already signaled "more disciplined growth."

Amazon ($105B FY26 capex)

Drivers: AWS expansion + advertising + retail logistics + Anthropic stake. AWS AI is the highest-growth segment but Amazon's capex is most diversified — only ~40% AI-specific, the rest is AWS general + retail logistics.

Trigger for a 20% cut: AWS growth deceleration (already showing in Q1 2026), OR retail logistics overbuild, OR Andy Jassy's profitability focus accelerating. AWS Q1 2026 grew 18% YoY vs market expectation of 22%; the slowdown is real.

Probability of FY27 capex cut: low-to-moderate. Amazon has the most flexible capex (can redirect to retail logistics). Less risk of a "stop AI capex" decision, more risk of "shift mix toward enterprise services."

The stress test — what happens at -20% capex

Modeling a 20% cut in FY27 hyperscaler capex (i.e., total goes from $370B → $296B):

NVIDIA (NVDA)

Direct impact: Top customer revenue declines proportionally. NVDA's H100/H200/B100/B200 are oversold; even if hyperscalers cut 20%, NVDA's data center segment revenue would decline meaningfully because hyperscaler share of total data center revenue is ~50%.

Quantification: Total NVDA revenue ~$220B FY26. Data center ~75% ($165B). Hyperscaler share of data center ~50% ($82B). 20% cut on $82B = $16.4B decline. NVDA revenue impact: ~7-8%.

Multiple compression risk: NVDA trades at ~40x forward earnings. In a "growth-decelerating" scenario, that contracts to 25-30x. Combined revenue + multiple compression = -30% to -40% potential downside.

However: NVDA has demand exceeding supply for 2+ years. Even with hyperscaler cuts, enterprise + sovereign + consumer GPU demand remains. NVDA survives this scenario but with significant multiple compression.

Verdict on NVDA: Most resilient AI-infrastructure name. Revenue impact modest; multiple compression real. Buy on -20% pullback for long-term thesis.

AMD

Direct impact: MI300/MI355/MI400 ramp depends on hyperscaler adoption. Less concentrated than NVDA (still small market share) but more sensitive to "will hyperscalers buy AMD's AI chips?" question.

Quantification: AMD AI revenue ~$15-20B FY26 (estimated). If hyperscalers cut capex AND choose to consolidate on NVDA, AMD AI revenue could drop 30-40%. Total AMD revenue impact: ~15-20%.

Multiple compression risk: AMD trades at ~30x forward. Could contract to 20x. Combined risk: -35% to -45% potential downside.

However: AMD is the second source for GPUs that every hyperscaler wants. Even in a slowdown, hyperscalers don't want NVDA-only dependency. AMD has structural protection but is operationally more vulnerable than NVDA.

Verdict on AMD: Higher downside than NVDA. Second-source thesis still intact. Add on pullback but smaller weight than NVDA.

Broadcom (AVGO)

Direct impact: Custom silicon for Google TPU, Meta MTIA, OpenAI custom chip. This is the most concentrated risk. Broadcom's AI revenue ($25B+ FY26 estimated) is primarily 3-4 hyperscaler customer contracts.

Quantification: Total AVGO revenue ~$60B FY26. AI ~40% ($24B). Hyperscaler share of AI ~95%+. 20% cut on $24B = $4.8B decline. AVGO AI revenue impact: ~20%. Total revenue impact: ~8%.

Multiple compression risk: AVGO trades at ~30x forward. Contracts to 25x. Combined: -25% to -35% downside.

However: Custom silicon contracts are typically multi-year and difficult to cancel. Hyperscalers can shift toward AVGO custom silicon vs NVDA general GPUs — AVGO might be a relative beneficiary. AVGO has hidden upside in a margin-pressure scenario.

Verdict on AVGO: Counterintuitive winner in a capex-cut scenario. Custom silicon is the way hyperscalers reduce costs. Multiple compression real but operational resilience strong.

Vertiv (VRT)

Direct impact: Power/cooling for data centers is downstream of all capex. If hyperscalers cut 20%, VRT revenue follows.

Quantification: VRT AI/HPC revenue ~70% of total. 20% cut on 70% = ~14% total revenue impact. But VRT margins are operationally levered — fixed-cost manufacturing means margin compression exceeds revenue compression. Net earnings impact: ~25-30%.

Multiple compression risk: VRT trades at ~45-50x forward (one of highest multiples in cohort). In a cut scenario, contracts to 25-30x. Combined: -50% to -60% potential downside.

However: Data center cooling is a structural need that survives any single capex cycle. VRT will retain demand but at lower margin and lower multiple.

Verdict on VRT: Highest downside in cohort. Wait for pullback before adding. Structural story still intact; valuation cushion thin.

Arista Networks (ANET)

Direct impact: Ethernet networking for hyperscalers. ANET's growth has been driven by Meta + Microsoft expansion.

Quantification: ANET revenue ~$8B FY26. Hyperscaler concentration ~60%. 20% cut on $4.8B = $960M decline. ANET revenue impact: ~12%.

Multiple compression risk: ANET trades at ~30x forward. Contracts to 20-25x. Combined: -35% to -45% downside.

However: Ethernet vs InfiniBand competitive dynamic favors ANET in cost-conscious scenarios. Hyperscalers cutting capex might actually prefer ANET's cheaper Ethernet over NVDA's premium InfiniBand.

Verdict on ANET: Mixed picture. Revenue impact real; competitive positioning improves. Multiple compression dominates.

Micron (MU)

Direct impact: HBM3e/HBM4 demand depends on GPU shipments. If hyperscalers cut GPU buys, HBM demand follows.

Quantification: MU AI-tied revenue ~25-30%. 20% cut on AI portion = ~6-8% total revenue impact. MU is the most diversified of the AI cohort (still significant non-AI memory revenue).

Multiple compression risk: MU trades at ~12-15x forward (cyclical). Contracts to 8-10x. Combined: -25% to -35% downside.

However: Memory is cyclical and Micron has historically navigated cycles. The HBM premium pricing is the AI-specific lever. Loss of premium hurts margins but base business continues.

Verdict on MU: Cyclical name; manageable downside. Add on pullback for long-term memory thesis.

Super Micro (SMCI)

Direct impact: AI server systems (liquid-cooled GPU servers) — directly tied to hyperscaler buildouts.

Quantification: SMCI revenue ~$25-30B FY26. AI server concentration ~70%+. Hyperscaler share of AI servers ~50-60%. 20% cut implies meaningful revenue compression.

Multiple compression risk: SMCI trades at volatile multiples (10-25x forward depending on momentum). In a cut scenario, settles at 10-15x. Combined: -45% to -55% downside.

However: SMCI has shown ability to navigate manufacturing constraints. The accounting controversies of 2024 (Hindenburg report etc.) make this the most volatility-prone name.

Verdict on SMCI: Highest volatility + highest downside. Avoid or smallest weight only.

Dell Technologies (DELL)

Direct impact: AI servers via Dell PowerEdge XE platform — direct hyperscaler exposure.

Quantification: DELL AI server revenue ~$30B+ FY26. Hyperscaler share ~50%. 20% cut implies ~$3B revenue decline. Total DELL revenue impact: ~3% (DELL is more diversified with PC/storage business).

Multiple compression risk: DELL trades at ~12-15x forward. Contracts to 10-12x. Combined: -20% to -30% downside.

However: Dell's PC + storage + traditional enterprise diversification provides cushion. DELL is the most operationally diversified name in the AI cohort.

Verdict on DELL: Most operationally resilient. Underrated as defensive AI play. Reasonable valuation cushion.

Hewlett Packard Enterprise (HPE)

Direct impact: AI servers + GreenLake AI cloud services — moderate hyperscaler exposure.

Quantification: HPE AI revenue ~15-20% of total. Hyperscaler share ~30-40%. 20% cut implies modest revenue impact. Total HPE revenue impact: ~2-3%.

Multiple compression risk: HPE trades at ~10-12x forward. Limited compression risk. Combined: -15% to -25% downside.

However: HPE's enterprise + government + GreenLake diversification provides strongest cushion. HPE is essentially insulated from hyperscaler capex cycle.

Verdict on HPE: Most defensive AI exposure. Boring but stable.

The portfolio implications

For Indian retail with existing US AI exposure, the framework:

If you own NVDA at >30% of US allocation:

• Trim 5-10% on any rally above current levels
• Redeploy to AVGO (custom silicon resilience) + HPE (defensive cushion)
• Keep core NVDA position long-term

If you own MSFT/META/GOOGL/AMZN as primary AI exposure:

• This is the safest place to be. Hyperscalers are the BUYERS, not the SELLERS.
• The risk to hyperscaler stocks is capex efficiency narrative (margin compression from underutilized capex), not capex level itself.
• Continue holding; consider trimming if you have >40% weight in this cohort.

If you own VRT/ANET/SMCI without significant hyperscaler offset:

• Highest downside cohort. Consider reducing 25-50% if these are >10% of portfolio.
• The "all the picks and shovels" thesis becomes problematic at hyperscaler concentration.
• Rotate proceeds to: NVDA (more diversified pick), AVGO (custom silicon counter-trade), or hyperscalers directly.

If you own AMD as second source play:

• Thesis intact but volatility elevated. Maintain but don't increase.

If you own MU:

• Cyclical name; less hyperscaler-direct. Reasonable defensive AI exposure.

What this stress test doesn't capture

Things I haven't quantified that should add to the analysis:

Sovereign AI buyers (positive offset): Saudi Arabia, UAE, India, Japan, Korea — sovereign cloud/AI investment is real and growing. NVDA has visibility on $40-50B in sovereign demand for FY27. This partially offsets hyperscaler cuts.

Enterprise AI infrastructure (positive offset): Banks, healthcare, manufacturing — non-hyperscaler enterprise spending on AI infrastructure is in early stages. NVDA + DELL + HPE all benefit. Hyperscaler cuts could shift some demand to enterprise AI capex.

AI chip supply constraints (positive for incumbents): Even at -20% capex, TSMC + Samsung HBM + Hynix HBM supply remains constrained. NVDA + AMD still allocated to highest-paying customers. Constrained supply protects pricing.

Tariff overlay: China tariffs at 60%+ create incentive for sovereign + non-Chinese cloud build-out. Tariffs are a positive demand catalyst for non-Chinese AI infrastructure.

Macro/recession: If recession arrives, hyperscalers cut not 20% but 40%+. The full downside scenarios are not in the -20% stress test. But this is the tail risk.

The closing read

The $370B Mag-7 AI capex cycle is real, structural, and the bedrock of the current AI infrastructure trade. The question is not "if" hyperscalers normalize but "when and how fast."

The honest assessment:

Safest in a -20% capex cut scenario: HPE (defensive), AVGO (custom silicon counter), DELL (operational diversification), NVDA (revenue impact modest), MU (cyclical resilience).

Most vulnerable in a -20% capex cut scenario: VRT (revenue + multiple compression worst-of-both), SMCI (highest volatility), ANET (revenue + multiple compression).

For Indian retail: if your US allocation is heavily weighted to "AI infrastructure" picks-and-shovels names without offset, rebalance toward hyperscalers directly + NVDA + diversified names (HPE, MU). The pick-and-shovels narrative is real, but customer concentration risk is meaningful at this stage of the cycle.

The single best protection: own the hyperscalers themselves (MSFT, META, GOOGL, AMZN). They are the BUYERS of capex, not the SELLERS. Their margins compress if AI ROI underwhelms; their revenue base grows regardless.

The single most important framework: a 20% hyperscaler capex cut is not a doom scenario, but it does compress multiples meaningfully across the cohort. Position accordingly. Buy the dip on quality (NVDA, AVGO), avoid the highest-momentum/highest-multiple names without conviction (SMCI, VRT), and tilt toward operational diversification (DELL, HPE) as defensive AI exposure.

Cross-references

• The trillion-dollar private valuation cluster — Apollo $36B + Anthropic context
• The Apollo-Anthropic $36B deal mechanics — what private AI funding tells you about public AI capex
• Iran war retrospective — parallel framework for how stress tests resolve

For Indian residents specifically:

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

This article reflects the AI capex landscape through early June 2026. Stress test scenarios are illustrative, not predictive. Actual hyperscaler capex decisions depend on factors not captured in this analysis.

Critical disclaimer: stress testing involves significant uncertainty. Past hyperscaler capex patterns don't guarantee future behavior. This article describes a framework for analyzing downside scenarios but does not substitute for personalized investment advice from a SEBI-registered investment adviser.