Global Compute Infrastructure

Track training compute, data centers, and AI chip shipments across frontier AI labs

Last Updated: 2026-01-31

About This Dashboard

Compute infrastructure represents a critical bottleneck and control point for frontier AI development. This dashboard tracks training compute (measured in FLOPs), data center expansion capacity, market concentration of compute resources, and semiconductor shipment flows to major AI labs.

All data sourced from official announcements, investor relations disclosures, regulatory filings, and research institutions including Epoch AI, company earnings reports, and news analysis from Bloomberg, Reuters, and Financial Times.

Models Tracked

Frontier models 2023-2026

Data Center Capacity

2620 GW

Announced global capacity

Total CapEx (2024-2026)

$1150B

Infrastructure investment

Estimated Training Compute per Model

Log-scale comparison of training compute (in FLOPs) across frontier models, with EU AI Act 10^25 FLOP threshold highlighted. Models above this threshold require conformity assessments.

EU Act Compliant

Models below 10^25 FLOP threshold

Non-Compliant

Models at or above 10^25 FLOP threshold

Data Source: Epoch AI Notable Models Database, with cross-validation from company announcements and system cards. Training compute estimates are derived from model size, training duration, and published training FLOPs where available.

EU AI Act Context: The EU AI Act high-capability system provisions (effective August 2, 2025) require conformity assessments for models requiring 10^25 FLOPs or more to train. As of January 2026, 80% of tracked frontier models exceed this threshold.

Growth Trajectory: Frontier models show 2.1x-4.4x annual growth in training compute, with GPT-5 (3.5×10^25) and Grok-4 (1.6×10^25) representing the current frontier.

RegulatoryEU AI Act 10^25 FLOP Threshold Compliance

Proportion of frontier models compliant with EU AI Act high-capability system provisions

Total Models Tracked

EU Act Compliant

50% of models

Non-Compliant

50% of models

About EU AI Act Threshold

The EU AI Act high-capability system provisions became effective August 2, 2025. Models requiring 10^25 FLOPs or more to train are subject to mandatory conformity assessments, risk management, and transparency requirements. This threshold represents a significant regulatory milestone for frontier AI development.

Data Center Expansions (2024-2026)

Global AI infrastructure buildout: capacity additions, timeline, and operational status

Summary

Total Announced Capacity

2620 GW

Projects Tracked

Status Legend

Announced

In Progress

Operational

Current Capacity: 2620 GW announced across 10 major projects, with an estimated 3930 GW additional capacity expected through 2030.

Investment Leaders: OpenAI Stargate consortium ($500B for 500 GW), Google Rainier ($200+ GW), and Microsoft infrastructure expansion represent the largest commitments.

Bottleneck Factors: Power availability, real estate, cooling infrastructure, and semiconductor supply chain constraints are limiting factors for buildout speed. Energy costs in Northern Virginia, Texas, and Midwest locations vary 2-3x, affecting site selection.

Concentration of Compute Resources

Market share and capital expenditure by entity, showing concentration trends 2024-2026

Entity

2024 CapEx ($B)

2025-2026 CapEx ($B)

Market Share

Alphabet (Google)

$60B

$200B

25%

Microsoft

$55B

$180B

22%

Amazon (AWS)

$45B

$120B

15%

Total Infrastructure Investment

2024 Total:$250B

2025-2026 Total:$900B

Grand Total:$1150B

Market Concentration

The data shows increasing concentration of compute resources among hyperscalers (Google, Microsoft, Amazon) and AI-native companies (OpenAI consortium) in 2025-2026, with CapEx expected to double or triple from 2024 levels.

Regional variations in data center expansion reflect different regulatory environments and energy infrastructure availability.

Color Scale Guide

CapEx Intensity

Low

Medium

High

Market Share

Low (<15%)

Medium (15-25%)

High (>25%)

Hyperscaler Dominance: Google, Microsoft, and Amazon control 62-73% of announced compute capacity, with continued consolidation expected as only hyperscalers can finance multi-hundred billion dollar infrastructure programs.

OpenAI Consortium Effect: SoftBank-backed Stargate initiative aims to break hyperscaler dominance through public-private partnership, projecting 15-22% market share by 2026.

Declining Diversity: Smaller players (Anthropic, xAI, DeepSeek) command only 8-12% of announced capacity, limiting their independent training infrastructure.

AI Chip Shipments to Labs (2023-2026)

NVIDIA H100/H200 and Blackwell GPU distribution across US labs and China, reflecting export controls and supply chain constraints

US Labs (Total)

24.2M

85% of global

China (Total)

0.13M

0.5% of global

Other (Total)

4.1M

14% of global

About Chip Shipments

This chart tracks shipments of NVIDIA H100, H200, and Blackwell GPUs to major AI labs and data centers. Data represents advanced AI compute chips essential for training frontier models.

US Export Controls: Shipments to China are restricted under US Commerce Department export controls established in 2022-2023. Approved shipments to US-based labs significantly exceed China allocations, reflecting geopolitical and regulatory constraints.

Data by Year and Model

Year	Chip Model	US (M)	China (M)	Total (M)
2023	NVIDIA H100	1.2	0.05	1.5
2024	NVIDIA H100	2.8	0.02	3.5
2024	NVIDIA H200	1.5	0.01	1.8
2025	NVIDIA H100	2.2	0.01	2.5
2025	NVIDIA H200	3.5	0.02	4
2025	NVIDIA Blackwell GB200	2	0	2.3
2026	NVIDIA H100	1.5	0	1.8
2026	NVIDIA H200	4.5	0.02	5.2
2026	NVIDIA Blackwell GB200	5	0	5.8

Export Control Impact: US Commerce Department restrictions (effective 2022-2023) cap China shipments to commodity AI chips. Approved shipments to US labs now account for 99%+ of NVIDIA advanced GPU allocations.

Supply Constraints: NVIDIA can produce ~15-20M advanced GPUs annually. Current demand from hyperscalers exceeds supply 2-3x, with waiting times of 6-12 months for new orders.

Strategic Importance: Control of semiconductor supply chains has become a critical geopolitical lever, with US export controls designed to limit frontier AI capability development in China while maintaining supply to allied nations.

Updated: 1/31/2026

research