Public Kempax report

Daily Update Global Data Center and GPU capacity · 2026-07-26

Kempax Research

Loop report · 2026-07-26

Global Data Center & GPU Capacity — Publication Summary

Publication date: 2026-07-27 Dataset version: 1.0 Records: 44 normalized records across 8 categories Sources: 17 references Loop: loop-0c93c098-2728-4106-a8bf-b79ea6a93bdc → looprun-3520945e-edf3-4c27-882c-c78b906a83ad Interactive dashboard: dashboard index.md

---

Dataset Overview

The Global Data Center & GPU Capacity dataset is a structured intelligence collection compiled via Kempax Web Research CloakBrowser . It captures 44 records normalized into 10 fields name , type , capacity mw , location , company , investors , gpu specs , electricity capacity , source refs , last updated across the following categories:

Category Records Description ---------- :-------: ------------- hyperscaler capex 5 Big Four hyperscaler + Oracle capex plans and GPU fleets gpu cluster 13 Large GPU deployments xAI, Stargate, Anthropic, Tesla data center project 6 Major DC buildout projects Stargate, Vantage, Creekstone, Meta gpu specification 7 NVIDIA and AMD data-center GPU specs and pricing data center reit 2 Equinix and Digital Realty investment metrics regional capacity 5 Regional DC under-construction aggregates investment vehicle 2 Notable private neocloud DC investors risk bottleneck 3 Critical supply chain and infrastructure risks

A machine-readable JSON variant is available at data dc gpu dataset.json .

---

Key Data Highlights

Hyperscaler Capex 2026 — Record Infrastructure Build

The five largest hyperscalers AWS, Google, Microsoft, Meta, Oracle are projected to spend $630–690B combined in 2026, representing +66–82% year-over-year growth . Each of the Big Three operates GPU fleets of 1M+ accelerators:

Company 2026 Capex YoY GPU Fleet --------- :----------: :---: ----------- Amazon AWS $200B +60% 1M+ NVIDIA + Trainium 2 Alphabet Google $175–185B +92–103% 1M+ TPU v6 v7 + NVIDIA Microsoft $110–120B +22–33% 1M+ NVIDIA + Maia 2 + AMD Meta $115–135B +60–88% 600k+ NVIDIA + AMD MI450 Oracle $50B +136% —

Key signal: Google's cloud backlog surged 55% to $240B+ , while Microsoft reports an $80B backlog constrained primarily by power availability .

DC Construction Pipeline — 23+ GW Under Construction

The formal BNEF pipeline records 23.1 GW across 831 sites globally; a broader MSCI definition captures 47 GW+ $550B end value . The Americas lead at 17,000 MW under construction 311 sites , with 92% of US primary-market capacity pre-committed by hyperscalers. At a 77% global pre-commitment rate, these are largely spoken for before completion.

Major GPU Clusters — Million-Scale Deployments

Organizations with 1M+ accelerator fleets: Google TPU v6 v7 + NVIDIA , Microsoft NVIDIA + Maia 2 + AMD , Amazon NVIDIA + Trainium 2 . Meta follows at 600k+. Notable single-site clusters: xAI Colossus 1 200k H100 H200 GB200 at 200 MW in Memphis and the Stargate program aggregating 150k–200k operational GPUs across 7 US sites + UAE, ramping to "several 100k."

Market Share Dynamics — NVIDIA's Share Moderates

NVIDIA's DC GPU market share has declined from a peak of 87% 2024 to an estimated 75% in 2026 as custom ASICs Broadcom, Google TPU, AWS Trainium and AMD gain traction. Broadcom's custom ASIC share grew from 1–2% in 2022 to an estimated 12–15% in 2026E . NVIDIA's FY2026 DC revenue nevertheless reached $193.7B +68% YoY , with Q1 FY27 DC revenue accelerating to $75.2B +92% .

Power — The Binding Constraint

Global DC electricity consumption is forecast at 565 TWh in 2026 +26% YoY per Gartner , projected to reach 219 GW by 2030 . Grid interconnection is the critical bottleneck: US wait times average 5 years with a 77% withdrawal rate, and 54 US moratoriums block an estimated $64B in projects. Transformer lead times extend to 128–144 weeks with prices up 77% since 2019. Grid investment needs through 2030 are estimated at $720B .

Construction Economics — Rapidly Escalating

Global average shell-and-core: $11.3M MW +6% YoY
All-in AI build including GPU fit-out: $30–40M MW
Hyperscale AI campus fully built-out: $45–55B per GW
Build times have stretched from 12 months to 18–24 months ; full site-to-COD cycles run 3–6 years

---

Dashboard View

The interactive dashboard at dashboard index.md presents the same data in a page-oriented markboard format with 10 sections:

Section Content :-: --------- --------- 1 Overview Statistics 14 key headline metrics at a glance 2 Hyperscaler Capex Plans 2026 Company-level capex, YoY change, GPU fleet size, backlog 3 DC Capacity Table Major buildout projects + regional construction pipeline by region 4 GPU Inventory Table GPU specifications 7 models , cloud pricing 5 tiers , cluster table, Stargate site estimates, market share trend 5 Electricity Capacity Highlights Global consumption, regional power constraints, notable power-supply projects 6 Investor & Backer Map Hyperscalers with tickers, DC REITs, Stargate project backers, sovereign private investors 7 Key Risks & Bottlenecks 8 risk items with severity indicators 🔴 🟠 🟡 8 Construction Economics Cost tables and timeline data 9 NVIDIA Financials FY2026 Revenue, net income, margins, guidance 10 Data Center REITs 2026 Equinix and Digital Realty with ticker, market cap, yield

The dashboard faithfully mirrors the normalized dataset from data dc gpu dataset.md and data dc gpu dataset.json 44 records, 10 normalized fields .

---

Methodology Notes

Data Collection - Content was gathered via Kempax Web Research CloakBrowser , compiling public information from 17 sources including: JLL, Data Center Dynamics, Archdesk, Datacenters.com, Futurum Group, IntuitionLabs, Presenc.ai, SiliconAnalysts, Deploybase, Opslyft, Programs.com, REITRankings, Gartner, and Knight Frank. - Source URLs are enumerated in data dc gpu dataset.md § Sources and referenced per record via source refs .

Data Quality & Limitations - Capacity values labeled "up to" or "planned" are forward-looking. 30–50% of the 2026 pipeline is expected to miss schedule per Sightline Climate . - GPU counts are estimates from public reporting; actual deployed counts, especially for non-public disclosures, may differ materially. - Capex figures are company-announced or analyst-estimated ranges; actual spend may differ. - Market share percentages are from third-party analyst estimates SiliconAnalysts, Presenc.ai and should be treated as directional, not audited.

Update Cadence - This dataset is updated on a daily cycle via Kempax Loop loop-0c93c098-2728-4106-a8bf-b79ea6a93bdc . - Last-updated dates reflect the research generation timestamp, not individual fact freshness.

---

Key Risks & Bottlenecks Summary

Risk Severity Signal ------ :--------: -------- Power availability 🔴 Critical Microsoft backlog "largely a function of power" Transformer lead times 🔴 Critical 128–144 weeks; +77% cost since 2019 Grid interconnection US 🔴 Critical 5 yr wait, 77% withdrawal, 54 moratoriums Construction delays 🟠 High 50% of 2025 projects delayed 3+ months HBM supply pressure 🟠 High H100 contract pricing +40% Oct 2025–Mar 2026 TSMC concentration 🟠 High Single point for advanced packaging Export controls China 🟡 Medium Black market pricing doubled REIT disintermediation 🟡 Medium Hyperscaler self-build trends

---

Reference Links

Resource Path ---------- ------ Interactive Dashboard dashboard index.md Dataset Markdown data dc gpu dataset.md Dataset JSON data dc gpu dataset.json

---

This document is the canonical publication candidate for the Loop publication adapter. It was generated from the structured dataset and dashboard view on 2026-07-27.