Public Kempax report

Daily Update Global Data Center and GPU capacity · 2026-07-28

Kempax Research

Loop report · 2026-07-28

Global Data Center & GPU Capacity — Daily Update

Published: 2026-07-28 12:20 UTC Next update: 2026-07-29

---

Executive Summary

The global AI infrastructure buildout continues accelerating with no signs of deceleration. Combined Big Tech capital expenditure for 2026 is now projected at $725–$775B , representing a near-doubling year-over-year. Tom's Hardware confirmed the combined Amazon Google Microsoft Meta capex figure at $725B — a 77% YoY increase — with an analyst dismissing the bear thesis as "garbage."

GPU cloud pricing has matured to a highly competitive market with 54+ providers and 5,200+ live listings tracked. Specialist clouds now undercut hyperscalers by 40–85% for identical hardware: Nebius offers the H100 at $2.00 hr 84% below Azure's $12.29 , Lambda lists the B300 at $6.50 hr 65% below AWS's $18.50 . The market is pricing H100 near EOL levels as Blackwell and Rubin dominate new deployments.

The NVIDIA R100 Vera Rubin is entering early customer shipments in H2 2026 with a generational leap: 16,000 FP8 TFLOPS 2.3× B300 , 50 PFLOPS FP4 3.3× B300 , 336B transistors +61% vs Blackwell , and HBM4 at 22 TB s bandwidth 2.75× B300 . Projected token costs of $0.005 1M represent a 40× reduction versus H100 . First cloud cohorts include AWS, Google Cloud, Azure, OCI, CoreWeave, and Lambda. However, the 2,300W TDP per GPU will stress data center power and cooling infrastructure at scale.

Power grid constraints remain the binding bottleneck , with new regulatory developments adding complexity: FERC ordered the six largest US grid operators to reform interconnection rules; Virginia enacted a first-of-its-kind $0.011 kWh data center consumption tax est. $600M year ; Texas ERCOT centralized load evaluation for sites 75 MW; and Monterey Park, California became the first US city to voter-enact a data center ban 86% vote . The "Bring Your Own Power" BYOP strategy — on-site gas, fuel cells, microgrids — is accelerating as the primary bypass for 4–10 year grid interconnection queues.

At least 20 new data center projects were detailed in today's research, including: Microsoft's $3.3B Mount Pleasant, WI campus now fully online; Amazon $10B and Google $15B campuses in rural Montgomery County, Missouri combined $25B ; Digital Edge's 60 MW AI-ready facility in South Korea with dual-feed 154 kV substations; and Google launched a new data center in Horndal, Sweden.

---

📊 Key Charts & Analytics

Chart 1: GPU Cloud Pricing — On-Demand by Provider July 2026

GPU Model Cheapest On-Demand $ hr Provider Most Expensive $ hr Provider Spread ----------- -------------------------- ---------- ---------------------- ---------- -------- H100 80GB $2.00 Nebius $12.29 Azure 6.1× H200 141GB $2.30 Nebius $13.13 GCP 5.7× B200 192GB $6.00 Hyperstack $14.38 AWS 2.4× B300 288GB $6.50 Lambda $18.50 AWS 2.8× MI300X 192GB $1.71 Crusoe $11.56 Azure 6.8×

Spot pricing on hyperscalers offers 63–70% discount: GCP B300 spot at $5.33 hr 70% off , Azure B300 spot at $5.44 hr 70% off , AWS B300 spot at $6.50 hr 65% off .

---

Chart 2: Hyperscaler Capex Trajectory 2022–2026

Company 2026 Capex vs 2025 Key Context --------- ----------- --------- ------------- Amazon $AMZN $200B +60% $200B confirmed in Q2 2026 earnings Alphabet $GOOGL $180–190B +98–109% $80B equity raise; Berkshire Hathaway buying $5B series Microsoft $MSFT $190B +111% $80B unfulfilled Azure orders; $3.3B WI campus online Meta $META $125–145B +74–101% Hyperion LA $500B 5 GW; Alberta C$13B 1 GW Oracle $ORCL $50B +136% $523B RPO; 147 active DCs + 64 developing Combined $725–775B +91–104% Tom's Hardware: $725B=77% YoY growth

Source: Tom's Hardware, Futurum Group, SimianX, Reuters

---

Chart 3: NVIDIA GPU Generational Comparison — R100 vs B300 vs B200

Metric R100 Rubin B300 B200 R100 vs B300 -------- ------------- ------ ------ ------------- FP8 TFLOPS 16,000 7,000 4,500 2.3× FP4 TFLOPS 50,000 15,000 9,000 3.3× HBM 288 GB HBM4 288 GB HBM3e 192 GB HBM3e HBM4 2.75× BW Bandwidth 22 TB s 8 TB s 8 TB s 2.75× Transistors 336B 208B 208B +61% NVLink 3.6 TB s v6 1.8 TB s v5 1.8 TB s v5 2× TDP 2,300W 1,400W 1,000W +64% Token cost $0.005 1M $0.03 1M $0.05 1M 6× cheaper vs H100 inference 18–22× 5.5× 4.5× 3.3–4× B300

---

Chart 4: Megaproject Pipeline — Top 15 by Capacity

Project GW Investment Location Developer Operator Key Investor --------- ----- ----------- ---------- ------------------- ------------- OpenAI Piketon 10.0 $500B+ Piketon, OH OpenAI, SB Energy SoftBank, NVIDIA $250B backstop Stargate 7.0+ $500B TX, NM, OH, WI, MI, UAE OpenAI Oracle SoftBank SoftBank, MGX Damac Digital 6.0 Undisclosed 13 countries Damac Digital UAE private SoftBank France 5.0 €75B $85B Northern France SoftBank Group SoftBank Meta Hyperion 5.0 $500B Louisiana Meta Blue Owl Capital AirTrunk India 3.0 $21B Maharashtra, India AirTrunk Blackstone Oracle Jupiter FC 2.85 Undisclosed TBD Oracle Oracle Digital Realty KC 2.0 Undisclosed Kansas City Digital Realty Public REIT NVIDIA SK Group 2.0 $500B TBD NVIDIA, SK Group NVIDIA, SK Group Vantage Texas 1.4 $25B Texas Vantage DC Silver Lake Nebius PA 1.2 Undisclosed Pennsylvania Nebius Private Iren Australia 0.8 Undisclosed South Australia Iren — Meta Alberta 0.8→1.8 C$13B Alberta, Canada Meta Meta Pure DC Finland 0.55 €7.5B Finland Pure DC — Meta El Paso 0.5→1.0 $12B debt El Paso, TX Meta + BlackRock JV BlackRock, JPMorgan, MS

---

Chart 5: US Data Center Power Demand Trajectory Goldman Sachs

Year GW YoY Growth ------ ----- ------------ 2025 31 — 2026 41 +32% 2027 66 +61%

EPRI: DCs could consume 9–17% of all US electricity by 2030 ; Virginia alone 39–57% of state supply. Goldman Sachs: US DC power demand doubling from 31 → 66 GW in just two years.

---

Chart 6: GPU Accelerator Market Share 2022–2026E

Legend: ▬ NVIDIA ▬ AMD ▬ Broadcom ASICs ▬ Google TPU ▬ AWS Trainium

Key insight: NVIDIA share drops 12pp from 2024 peak 87% → 75% . Custom ASICs + in-house silicon now 31% combined up from 5% in 2022 .

---

Key Developments — July 28, 2026

Date Company Project Type Key Details Source ------ ------------------- ------ ------------- -------- Jul 28 GPU Cloud Pricing Market Data Granular per-provider pricing published for B300 B200 H200 H100 MI300X across 54 providers; 5,200+ listings GPU Advisor, SiliconAnalysts Jul 28 Microsoft Mount Pleasant, WI DC $3.3B campus — first facility fully online DCK Jul 28 Amazon + Google Missouri DC $10B Amazon and $15B Google campuses in Montgomery County, MO; combined $25B DCK Jul 28 Google Sweden DC DC Launched in Horndal for Search Cloud YouTube DCK Jul 28 Digital Edge South Korea DC 60 MW AI-ready ; 90 MVA power agreement; dual-feed 154 kV substations DCK Jul 28 Samsung floating DCs Innovation Plans to develop floating data centers for land scarcity DCK Jul 28 Chevron + Microsoft Pecos TX Power 20-year power agreement; on-site generation for planned campus DCK Jul 28 FERC interconnection reform Regulation Show-cause orders to 6 largest US grid operators DCK Jul 28 Virginia DC tax Regulation $0.011 kWh consumption tax; est. $600M year ; effective Jul 1, 2026 DCK Jul 28 Texas PUC ERCOT order Regulation Centralized evaluation for loads 75 MW DCK Jul 28 Monterey Park, CA DC ban Regulation First US voter-enacted ban 86% vote DCK Jul 28 Alphabet $80B equity raise Financing Berkshire Hathaway buying $5B each Class A and C shares DCK Jul 28 Hassan Allam Egypt DC $400M first-phase hyperscale development; license approved DCK Jul 28 Stark Power acquires Sagebrush M&A Israeli firm acquiring 100% of Sagebrush Infrastructure US DCs + gas plants DCK Jul 28 Galaxy Data Center $250M raised Financing Initial strategic round for SE Asia platform development DCK Jul 28 Samsung + Broadcom $200B collab Supply Chain Memory, foundry, packaging — worth $200B by end of decade DCD Jul 28 NVIDIA $250B backstop for OpenAI Financing NVIDIA weighing $250B backstop for OpenAI 10GW Ohio DC on DOE land DCD Jul 28 Utho Cloud India 10K GPUs Procurement Indian cloud firm aims to buy 10,000 GPUs over 2 years DCD Jul 28 CoreWeave EdgeConneX lease DC Leasing capacity at Bastrop County, TX campus 4 DCs DCD

---

GPU Cloud Pricing — Complete Reference

Cheapest On-Demand per GPU July 2026

GPU Lowest $ hr Provider Highest $ hr Provider Spread ----- ------------ ---------- ------------- ---------- -------- R100 proj. $8–12 2027 Marketplace $18–25 H2 2026 Hyperscaler 2–3× B300 288GB $6.50 Lambda $18.50 AWS 2.8× B200 192GB $6.00 Hyperstack $14.38 AWS 2.4× H200 141GB $2.30 Nebius $13.13 GCP 5.7× H100 80GB $2.00 Nebius $12.29 Azure 6.1× MI300X 192GB $1.71 Crusoe $11.56 Azure 6.8× A100 80GB $1.09 Thunder Compute $5.03 GCP 4.6×

Selected Provider Pricing On-Demand, $ GPU-hr

Provider H100 H200 B200 B300 MI300X Spot Available ---------- ------ ------ ------ ------ -------- ---------------- Nebius $2.00 $2.30 — — — — Lambda $3.14 $3.75 $6.99 $6.50 $3.06 — Crusoe $2.50 — — — $1.71 H100 $1.50 Hyperstack $2.50 $3.99 $6.00 — — 20–30% off RunPod $2.69 $5.50 — — — H100 $1.25 54% off CoreWeave $4.76 $5.20 $6.50 $8.50 $4.10 — GCP $10.98 $13.13 $13.75 $17.75 — 70% off Azure $12.29 $13.78 $14.00 $18.13 $11.56 70% off AWS $6.88 $13.00 $14.38 $18.50 — 63–65% off

---

NVIDIA R100 Vera Rubin — Deep Dive

Confirmed Specifications

Spec R100 vs B300 vs B200 ------ ------ --------- --------- FP8 TFLOPS 16,000 2.3× 3.6× FP4 TFLOPS 50,000 inference 35,000 training 3.3× — 5.6× — HBM 288 GB HBM4 Same capacity, 2.75× bandwidth +96 GB, 2.75× bandwidth Bandwidth 22 TB s 2.75× 2.75× Transistors 336B TSMC N3 +61% +61% NVLink 3.6 TB s v6 2× 2× Networking ConnectX-9 1.6 Tb s 2× 4× TDP 2,300W +64% +130%

Architectural breakthroughs: - HBM4 enables models requiring 4× H100 or 2× B200 for 128K context to fit on a single GPU - NVLink 6 doubles inter-GPU bandwidth, cutting gradient sync time for communication-bound training - FP4 at 50 PFLOPS means a 400B-parameter model at FP4 fits on one R100 vs 2× B300 or 3× B200 for FP8 - NVIDIA claim: "Up to 10× reduction in inference token cost at system scale vs Blackwell platform"

Pricing Projections

Tier On-Demand hr Timing ------ ------------- -------- Hyperscalers AWS, GCP, Azure, OCI $18–$25 H2 2026 early access Specialist clouds CoreWeave, Lambda, Nebius $10–$15 H2 2026–2027 Marketplaces Spheron $8–$12 2027 18-month projected decline $5–$10 By 2028

Availability Timeline by Provider

Provider Target Status ---------- -------- -------- AWS, GCP, Azure, OCI, CoreWeave, Lambda H2 2026 Early access confirmed Groq 3 LPX Rack replaces Rubin CPX Q3 2026 Shipping alongside NVL72 Nebius, Nscale, Spheron 2027 Broader marketplace availability Rubin Ultra 100 PFLOPS FP4, 1TB HBM4e, NVL576 2027+ "Kyber" rack

---

Data Center Construction — Updated Pipeline

Global Capacity Snapshot

Metric Value Source -------- ------- -------- Active data centers worldwide 11,000–12,000 MedhaCloud Hyperscale DCs operational 1,297 +770 pipeline Synergy Research Global installed capacity 103 GW JLL ; → 200 GW by 2030 JLL 2026 Under active construction 23.1 GW across 831 sites BloombergNEF US pipeline under construction 5 GW of 12 GW announced Coradvisors YTD US DC construction spending Jan–May 2026 $58.1B 4× 2025 record ConstructConnect Preconstruction pipeline $101B across 100 projects ConstructConnect

20+ New Projects Detailed Today

Project Capacity Investment Location Type --------- ---------- ------------ ---------- ------ Microsoft Mount Pleasant, WI Not specified $3.3B WI, USA Operational — NEW Amazon Missouri Not specified $10B Montgomery County, MO New campus Google Missouri Not specified $15B Montgomery County, MO New campus Google Sweden Not specified Not specified Horndal, Sweden Launched Chevron + Microsoft Pecos TX Not specified Not specified Pecos, TX 20yr PPA + campus Digital Edge Ansan, South Korea 60 MW AI-ready Not specified Ansan, South Korea Dual-feed 154 kV Samsung floating DCs Not specified Not specified — Innovation Hassan Allam Egypt Hyperscale $400M first-phase Egypt License approved Stark Power Sagebrush Not specified Not specified US DCs Acquisition Galaxy DC SE Asia Not specified $250M raised SE Asia Platform development

---

Power Grid & Regulatory Landscape

Regulatory Developments NEW this session

Action Detail Impact -------- -------- -------- FERC show-cause orders 6 largest US grid operators ordered to reform interconnection rules Could shorten 4–10 year queues Virginia DC tax $0.011 kWh consumption tax; est. $600M yr; effective Jul 1, 2026 First state-level DC-specific tax Texas PUC order Centralized ERCOT-led evaluation for loads 75 MW Regulatory burden increase Monterey Park, CA ban First US voter-enacted DC prohibition 86% vote Precedent for local opposition

Grid Constraints — Worsening Trajectory

Constraint Current State Trend ----------- -------------- ------- US interconnection queue 4–10 year avg wait; 77% withdrawal rate 🛑 Worsening Transformer lead times 3–5 years worsened ; +77% cost since 2019 🛑 Worsening PJM Interconnection "Cannot meet expected energy requirements" Jul 3, 2026 🛑 Critical Northern VA 70 GW queue vs 25.4 GW peak load; 3 GW fault drop Jul 23 🛑 Critical BYOP Bring Your Own Power On-site gas fuel cells microgrids — primary bypass strategy 🔄 Accelerating

Nuclear SMR Portfolio Unchanged

Hyperscaler Partners Total Capacity Key Detail ------------- ---------- --------------- ------------ Meta $META Constellation, TerraPower, Oklo, Vistra, Dominion 6.6 GW TerraPower 4.0 GW 8 Natrium ; Oklo 1.2 GW Amazon $AMZN Talen Energy, X-Energy, Dominion 2.4 GW 960 MW X-energy 12 SMRs ; target 2028 Microsoft $MSFT Constellation TMI restart 837 MW 20-year PPA $1B; target H2 2027 Google $GOOGL Kairos Power 500 MW 50 MW → 500 MW; 2030 first reactor Equinix $EQIX Radiant Industries 20 MW Kaleidos microreactor

---

Key Risk Dashboard

Risk Update Jul 28 Severity ------ --------------- ---------- Power grid interconnection FERC orders reform; 4–10 year queues; VA tax; CA ban 🔴 Critical Transformer equipment shortage 3–5 year lead times; BYOP accelerating; 50% US builds delayed 🔴 Critical Regulatory fragmentation FERC + VA tax + TX ERCOT + CA ban — growing patchwork risk 🔴 Critical NEW Hyperscaler capex sustainability $725B+ combined; analyst defends as sustainable 🔴 Critical Circular financing NVIDIA↔OpenAI $250B backstop talks confirmed — NVIDIA for OpenAI 10GW Ohio DC 🟠 High NVIDIA→R100 generation transition 2,300W TDP per GPU; new cooling infrastructure required at scale 🟠 High HBM4 memory cost inflation Samsung+Broadcom $200B deal highlights intensity; $40-50B market by 2027 🟠 High TSMC N3 concentration Rubin capped at 250–300K units by N3 wafer-out in 2026 🟠 High AMD MI400 pricing volume TBA Announced Jul 27; no pricing, TDP, volume, or ship dates 🟡 Medium GPU useful life uncertainty CoreWeave 6yr vs Lambda 5yr vs Nebius 4yr depreciation 🟡 Medium

---

Sources

• GPU Advisor — Cloud GPU Pricing July 2026 — 2026-07-28
• SiliconAnalysts — Cloud GPU Pricing Comparison — 2026-07-01
• Spheron — NVIDIA Rubin R100 Guide — 2026
• Spheron — R100 Rental Page — 2026
• Spheron — B300 Blackwell Ultra Guide — 2026
• GPU Tracker — Rubin R100 — 2026
• Data Center Knowledge — New DC Developments July 2026 — 2026-07
• DatacenterDynamics — Jul 27-28 Headlines — 2026-07-27 28
• Tom's Hardware — Big Tech AI Spending $725B — 2026-07
• Dedirock — New DC Developments July 2026 — 2026-07
• Coradvisors — Energy Grid DC Capacity Bottlenecks — 2026-06
• GitHub — GTC2026 Supply Chain Alpha Report — 2026
• Futurum Group — AI Capex 2026 — 2026-Q2
• SimianX — Amazon Q2 Earnings 2026 — 2026-07-27
• SimianX — Microsoft Q4 Earnings 2026 — 2026-07
• Reuters — Google Cloud Revenue Beats — 2026-07-22
• JLL — 2026 Global Data Center Outlook — 2026-Q2
• OpenAI — Stargate Project — 2026-07
• ConstructConnect — July 2026 DC Report — 2026-07
• Speediyo — Vera Rubin Availability Matrix — 2026
• TrendForce — CSP CapEx 2026 — 2026-05
• Gartner — DC Electricity Demand 2026 — 2026-06-10
• CBRE — Global Data Center Trends 2026 — 2026
• MedhaCloud — Data Center Statistics 2026 — 2026
• Programs.com — Data Center Statistics 2026 — 2026

---

<small Report details: Loop: Daily Update Global Data Center and GPU capacity loop-0c93c098-2728-4106-a8bf-b79ea6a93bdc Run: looprun-89407803-bad5-40cd-928d-78178d6a9c90 Session: loopsession-5eb1ea6f-64fe-45ea-b080-ce10aa3198e6 Workspace: ws-01 Compiled: 2026-07-28 12:20 UTC. All claims traceable to cited sources. Data quality: High for GPU pricing live scraping 54 providers , NVIDIA AMD specs; Medium for capex projections and construction timelines; Low–Medium for R100 pricing projections and GPU shipment forecasts. Market conditions change rapidly — verify with current sources for procurement decisions.< small