Hyperscale AI cloud platforms

Why the concentration exists

Hyperscale data centers are facilities exceeding 5,000 servers and 10 megawatts of critical IT load, designed for horizontal scaling across thousands of compute nodes. Traditional enterprise data centers typically run 1 to 5 megawatts with air cooling for general-purpose workloads. The scale difference reflects a fundamental architectural shift toward infrastructure that can train and serve large AI models.[12]

Capital intensity creates a formidable barrier to entry in this market. Combined capital expenditure for the five largest hyperscalers rose from approximately $150 billion in 2023 to $256 billion in 2024, then surged to an estimated $443 billion in 2025. Q1 2026 earnings confirmed the Big Five will collectively spend approximately $725 billion on AI infrastructure in 2026, with roughly 75 percent tied directly to AI workloads.[9][17][18]

McKinsey projects the AI capex cycle will require $6.7 trillion in global data center capital expenditure by 2030. AI workloads are expected to drive approximately 70 percent of total demand. This level of investment locks in market concentration because only incumbents with massive balance sheets can sustain the buildout.[18]

What the evidence shows

Who supplies it

AWS holds 28 to 30 percent global cloud market share as of late 2025, followed by Microsoft Azure at 20 percent and Google Cloud at 13 percent. These three providers control roughly 59 percent of global hyperscale capacity. Synergy Research Group counted 1,297 operational hyperscale data centers worldwide as of late 2025, with a pipeline of roughly 770 more in planning, construction, or fit-out.[1][13][17]

ByteDance represents a different category of hyperscaler focused on content distribution rather than enterprise cloud services. Its data center operations in China and overseas support video delivery and AI recommendation engines. The company has expanded into custom chip development to reduce compute cost and power draw.[3]

What it depends on, and what depends on it

AWS operates 39 geographic regions and 123 Availability Zones as of early 2026, with more than 30 Local Zones and 5G-integrated Wavelength locations. Microsoft Azure runs 62 cloud regions and 120 availability zones, with over 200 facilities linked by more than 175,000 miles of fiber optic lines. These networks provide the geographic reach enterprises require for low-latency AI inference.[13][15]

AWS designs its own servers through original design manufacturers based in Asia. In the past two years, 50 percent of AWS's new CPU capacity has been based on Graviton. Amazon used 150,000 Graviton chips during its annual Prime Day sale to meet e-commerce demand. This vertical integration into custom silicon reduces dependence on third-party chip suppliers.[8]

AWS AI Factories deploy EC2 UltraClusters interconnected using Elastic Fabric Adapter networking in a petabit-scale non-blocking network, scaling to thousands of GPUs. These clusters run workloads on Trainium accelerators including Trainium2 and Trainium3 as well as NVIDIA GPUs including B200, GB200, and upcoming B300 and GB300. AWS and NVIDIA are collaborating with HUMAIN to build an AI Zone in Saudi Arabia featuring up to 150,000 AI chips.[10][11]

Where it sits in the stack

What would break it

GPU supply constraints limit how quickly hyperscalers can expand capacity. Lead times for NVIDIA H100 and H200 GPUs stretched to 6 to 12 months during 2025, with similar constraints expected through 2026. When Azure could not spin up GPU capacity fast enough for OpenAI's ChatGPT and GPT-4 needs, Microsoft leased capacity from CoreWeave and Lambda Labs to bridge the gap.[6][2]

Permitting and local opposition slow new data center development. Between March 2024 and 2025, 16 data center developments were delayed or denied due to permitting restrictions, with local community pushback as a leading cause. Hyperscalers must navigate these constraints while demand continues to outpace available capacity.[16]

Neoclouds have emerged as purpose-built GPU infrastructure providers targeting AI workloads. CoreWeave claims to provide eight to 10 times faster container spin-up times and three to five times faster downloads than traditional providers. Nebius, founded in 2024 and headquartered in Amsterdam, owns and operates over 75 percent of its contracted data center capacity. These specialists can outmaneuver hyperscalers on performance for specific AI tasks.[4][5][14]

What to watch

AWS announced over $30 billion in combined investments in Pennsylvania and North Carolina for AI infrastructure, with $20 billion allocated to Pennsylvania and $10 billion to North Carolina. These facilities will take years to reach full operation and will face the permitting and grid connection challenges that have delayed other projects.[7]

Related nodes

Sources

1. holori.com · 2025-11-18T14:38:06
2. resources.telegeography.com · 2026-06-17T17:14:26
3. datacentremagazine.com · 2025-12-03T09:37:56
4. stlpartners.com · 2026-04-28T09:31:16
5. infoworld.com · 2025-08-25T14:54:44
6. datacentredigest.com · 2026-06-18T11:48:03
7. datacenterfrontier.com · June 10, 2025
8. journal.uptimeinstitute.com · 2025-02-12
9. shieldoperations.co.uk · February 23, 2026
10. aws.amazon.com
11. aboutamazon.com
12. aptlytech.com
13. digitalocean.com · Q4 2025
14. datacentremagazine.com
15. dgtlinfra.com
16. bvp.com · early 2026
17. irecruit.co · 2026
18. alcapitaladvisory.com · July 2026