Big 4 hyperscaler capex: $725B combined in 2026 (77% YoY) per Q1 earnings. Google 'compute constrained.' Microsoft cloud backlog $460B (doubled). 60%+ of spend goes to power/cooling/construction vs compute hardware. Growth decelerates to 13% in 2027 and 5% in 2028, but absolute level rises. The $830B (top-9) is not a spike — it's the new floor.

Autoresearch: CSP capex cycle — May 21, 2026 macro bucket scan

Generated by /autoresearch on 2026-05-21. Synthesized across 2 rounds (early exit — Round 3 sub-questions fully answered by Round 2). Search snippets only (WebFetch blocked). No Grokipedia anchor. Context: vault/projects/stock-market

Summary

All four major hyperscalers raised 2026 capex guidance during Q1 earnings (April 2026): combined Big-4 capex is now $725B (+77% YoY), and TrendForce's top-9 CSP figure of $830B (+79% YoY) remains intact. Google CEO explicitly stated the company is "compute constrained" and that cloud revenue would have been higher with more supply — directly answering the wiki's open question: the $830B is demand-driven, not discretionary. Microsoft's cloud contract backlog doubled to $460B QoQ. The key insight that updates the wiki: 60%+ of the $710-725B Big-4 capex is going to power infrastructure, cooling, and data center construction — not compute hardware — which substantially confirms the electrical/physical construction picks-and-shovels thesis. Growth rates will decelerate (77% → 51% in 2026 → 13% in 2027 → 5% in 2028), but the absolute spending level continues to rise. The $830B is not a one-year spike — it's a floor that expands modestly through 2028.

Findings

CSP Capex Is Accelerating, Not Normalizing — Q1 2026 Guidance Updates

Per Q1 2026 earnings calls (all companies reported late April 2026):

Hyperscaler	2026 Capex Guidance	YoY Growth	Notes
Microsoft	$190B	~130%	CFO attributed $25B to rising memory/component costs
Alphabet (Google)	$180-190B	~100%+	Revised up from $175-185B; CEO: "compute constrained"
Amazon (AWS)	$200-230B	~50%+	AI cloud services driving demand
Meta	$125-145B	~85%	Revised up $10B; cites higher component pricing, land, power, workers
Big 4 Combined	~$725B	77%	Analyst headline per Yahoo Finance / Tom's Hardware

TrendForce (top-9 CSPs including Oracle, ByteDance, Tencent, Alibaba, Baidu) revised upward to $830B (+79% YoY) as of May 6, 2026 (TrendForce), from an earlier estimate of 61% growth.

Google demand signal: CEO Sundar Pichai stated: "We are compute constrained in the near term" and "our cloud revenue would have been higher if we were able to meet the demand" (The Next Web). This is the clearest available evidence that the $830B is demand-constrained, not supply-ahead.

Microsoft lock-in: Microsoft's cloud contract backlog reached $460B in Q1 2026, up from $240B at Q4 2025 — nearly doubled in one quarter (Yahoo Finance). A $460B backlog locks in multi-year revenue and anchors capex commitments.

Key Insight: 60%+ of Capex Goes to Power/Construction, Not Compute

One of the most important compositional details from the Big-4 earnings synthesis (Hyperscaler Capex 2026 analysis):

"More than 60% of [the $320B identified for AI infrastructure] is going into power infrastructure, cooling, and data centre construction, not compute hardware. The binding limit, in an increasing number of markets, is electricity."

This means of the $725B Big-4 capex:

~$290B+ is power infrastructure, cooling, construction, real estate — flows to ETN, GEV, VRT, ABB, HUBB, PWR, and grid utilities
~$290B+ is compute hardware (GPU, networking, servers) — flows primarily through Nvidia → TSMC → HBM (SK Hynix, Micron) → semicap (ASML, AMAT, LRCX, KLA)

The flow diagram: hyperscaler capex → Nvidia (GPU) → TSMC (wafers) → equipment makers (ASML, AMAT, LRCX, KLA) is the minority beneficiary of the $725B. The majority goes to power and physical construction infrastructure that the existing thesis has been tracking through the datacenter-construction-electrical-picks-shovels question.

Growth Rate Trajectory: Deceleration, Not Decline

Per multiple analyst sources (DataCenter Dynamics/Moody's, Allianz-Trade):

Year	YoY Growth	Absolute Level (est.)
2025	Baseline	~$410B (Big 4)
2026	77%	~$725B (Big 4) / $830B (top-9)
2027	13%	~$820B (Big 4 est.)
2028	5%	~$860B (Big 4 est.)

The growth rate deceleration from 77% → 13% → 5% sounds bearish, but the absolute spending level continues rising every year. This is not a capex bubble pop; it's natural digestion of an extraordinary base. The annual increment still funds continued expansion of fab capacity, power grid, and packaging.

Bull thesis answer: The $830B is not a one-year spike — it's a new structural floor that will grow modestly through 2028. For long-duration semicap, HBM, and power-infrastructure plays, this confirms the 5-7 year secular hold thesis.

Bear risk (Moody's): "Hyperscalers perceive underinvestment as an existential threat, while investors worry aggressive spending could lead to overbuilding and weaker returns." If monetization of AI workloads disappoints (relative to the $460B+ in committed cloud contracts), future capex guidance could be cut. The current upward revision cycle assumes strong AI revenue growth through 2027.

Direct Beneficiaries — Updated Prioritization

Based on the 60%+ power/construction finding and continuing compute spending:

Power & physical infrastructure (60%+ of spend):

ETN (Eaton), GEV (GE Vernova), ABB, HUBB (Hubbell), PWR (Quanta Services) — confirmed picks-and-shovels
Grid utilities (Duke $102.2B, Southern $81.2B over 5 years)
VRT (Vertiv) — liquid cooling and data center thermal management

Compute hardware supply chain (remaining ~40%):

Nvidia (GPU demand directly; primary beneficiary)
TSMC (wafer economics; N2 ~$30K/wafer, CoWoS >60% expansion)
SK Hynix, Micron (HBM — pre-sold through 2026-2027)
ASML, AMAT, LRCX, KLAC (semicap equipment — buying locked in by fab expansions)

Hypothesis Surfaced: Microsoft Memory Cost Signal

Microsoft attributed $25B of its $190B capex increase specifically to rising memory chip and component costs (Yahoo Finance). This is a direct financial read-through of the HBM supply bottleneck — hyperscalers are paying materially more for memory, validating the HBM ASP thesis from hbm-supply-bottleneck. A $25B cost increase attributed to memory alone is an extraordinary figure.

Contradictions and Open Questions

Overbuild risk: Moody's flags overbuild/weak-return risk; Allianz says cycle is "war-proof for now." The $460B Microsoft backlog is the key counter-argument — if contracts are locked, the spend has committed buyers. But if AI revenue growth disappoints relative to the backlog, write-downs and future capex cuts are possible.
When does the power bottleneck bind? 60%+ of capex flows to power/construction, but grid interconnection queues and 5-year transformer lead times mean physical build will lag financial commitment. This creates a window where capex is committed but capacity can't be delivered — a risk for hyperscaler project timelines.
Non-US CSP share: The Big-4 $725B figure is heavily US-weighted. The $830B (top-9) adds ByteDance, Tencent, Alibaba, Baidu — all operating under export-control constraints on Blackwell/advanced GPUs. Their capex composition differs materially (more Huawei Ascend, domestic). The strategic implications of this split haven't been fully traced in the wiki.
2027 deceleration timing: The 77% → 13% jump is sharp. If Microsoft's backlog converts to revenue faster than expected, the deceleration may be milder. If cloud monetization disappoints, the deceleration may be sharper. This is the central risk for long-duration semicap positions.

Provenance

Rounds run: 2 (early exit — Round 3 sub-questions fully resolved by Round 2 data).

Sub-questions by round:

Round 1 (broad survey):

Has any hyperscaler revised 2026 capex guidance since early May?
Is the $830B figure holding or showing signs of normalization?
Which semiconductor and infrastructure companies are direct beneficiaries?

Round 2 (drill-down):

Q1 2026 earnings data — targeted updated Big-4 capex numbers
Overbuild / normalization risk — targeted the bear case

Anchor source: No Grokipedia entry fetched (HTTP 403).

URLs fetched: 0 successful (HTTP 403 environment).

Key sources (search snippets):

Tools used: WebSearch. Generated: 2026-05-21, remote execution environment.