The Hyperscaler AI Arms Race: Reshaping International Cloud Infrastructure

The main cloud hyperscalers—primarily Amazon, Google and Microsoft—are presently engaged in an infrastructure arms race of an unprecedented scale.

Pushed nearly completely by the explosive adoption and scaling of synthetic intelligence, this huge pivot is essentially altering capital allocation and the geographic footprint of world information facilities. Let’s check out AI-related investments, notable tasks, GPUs-as-a-Service, and why hyperscalers lease from neoclouds.

This evaluation is powered by proprietary information you may solely get in TeleGeography’s Cloud and WAN Analysis Service.

Hyperscaler AI investments

Trade projections point out a large surge in funding, with whole hyperscaler capital expenditure (CapEx) anticipated to achieve a staggering $600 to $700 billion in 2026. This represents a 40% to 50% improve over 2025 funding ranges.

Crucially, the character of this spending has shifted. Monetary analysts estimate that roughly 75% of this whole CapEx is being directed particularly towards AI infrastructure, closely outpacing investments in conventional cloud computing structure.

Whereas tech giants are actively retrofitting current cloud information facilities to accommodate baseline AI development, the overwhelming majority of heavy funding is being poured into big, new “greenfield” AI information facilities. Objective-built services are largely required as a result of AI workloads demand considerably greater energy densities, specialised liquid cooling, and strengthened structure for heavy GPU clusters.

A few of the largest tasks embrace:

• Amazon (AWS): Challenge Rainier (Indiana), Louisiana, Mississippi
• Google: Columbus (OH), Omaha (NE), Texas, Oklahoma, Visakhapatnam (India)
• Microsoft: Fairwater Campus (Mount Nice, WI), Atlanta (GA), Narvik (Norway), Loughton (U.Okay.)

Challenge Stargate

One other extremely publicized AI initiative is Challenge Stargate. Backed by an unlimited $500 billion funding, this three way partnership between OpenAI, SoftBank, Oracle, and MGX (an Abu Dhabi funding agency) goals to construct a community of information facilities particularly designed to coach and function superior AI fashions. Oracle is spearheading the flagship Stargate campus in Abilene, Texas, whereas OpenAI and its companions are creating a second web site in Port Washington, Wisconsin—about an hour north of Microsoft’s Fairwater campus.

This association typically raises a couple of questions: Is not Microsoft OpenAI’s major accomplice, and does not OpenAI run on Microsoft’s cloud? In that case, why is Oracle main the Stargate construct as an alternative of Microsoft, and why is Amazon concerned in OpenAI’s latest funding?

Whereas rumors in 2024 advised Microsoft would completely construct OpenAI’s information facilities, Oracle finally displaced them as the first infrastructure builder for the Stargate initiative. Regardless of this shift in bodily development, Microsoft Azure stays the unique cloud supplier for OpenAI’s first-party merchandise and its “stateless APIs” (the underlying know-how builders use to entry the fashions).

AI and GPU compute-as-a-service

Let’s present a bit extra element about this arms race. The 2020s have witnessed a surge of curiosity in AI, mirroring the preliminary hype and rise of cloud computing within the 2000s. Simply as cloud computing revolutionized how companies retailer, entry, and course of information, AI is being marketed for its potential to rework industries by automating duties, bettering decision-making, and enhancing total accuracy and precision.

On the coronary heart of this revolution are GPUs (Graphics Processing Items). Initially designed for rendering graphics, GPUs have grow to be the cornerstone of contemporary AI computation. They’re a necessary a part of an AI “cluster,” appearing as server accelerators that course of a number of calculations concurrently—typically one to 2 orders of magnitude quicker than a mean CPU. This processing energy is essential through the AI mannequin coaching section.

Neoclouds

Whereas GPU companies are hardly new—AWS and Microsoft have provided GPU compute companies for the higher a part of a decade, with Google becoming a member of barely later—the panorama is shifting. Immediately, all main Cloud Service Suppliers (CSPs), together with Oracle, IBM, Alibaba, and OVH, provide GPU compute. Nevertheless, a brand new wave of specialist cloud suppliers has emerged, providing GPUs-as-a-Service (GPUaaS). These “neoclouds” grant anybody entry to the {hardware} wanted to coach their very own fashions or run inferences.

Surprisingly, among the largest prospects for these GPUaaS suppliers are the hyperscalers themselves, particularly Microsoft and Google.

GPUaaS Supplier Cloud Areas

Observe: Knowledge embrace 18 main GPUaaS centered cloud suppliers reminiscent of CoreWeave, Nebius, and Nscale. This isn’t an exhaustive listing of specialist GPUaaS suppliers. Knowledge as of Q1 2026.

Why trillion-dollar titans lease from neoclouds

It could appear utterly counterintuitive that tech giants would want to lease compute from a lot smaller suppliers like CoreWeave or Lambda. Nevertheless, the generative AI growth created bodily and supply-chain bottlenecks that even the hyperscalers could not remedy alone. To maintain up with insatiable demand, Microsoft and Google adopted a symbiotic technique, counting on GPUaaS suppliers for a number of strategic, technical, and financial causes:

Velocity to deployment and energy bottlenecks

Constructing a large, conventional hyperscale information middle from scratch takes two to 4 years, and securing the large energy agreements and grid entry required for AI is extremely troublesome. Many GPUaaS suppliers bypass this by retrofitting services initially constructed for high-density purposes, like cryptocurrency mining. These websites already possess the 2 issues AI wants most: huge energy capability (typically a whole bunch of megawatts) and superior thermal administration. As a result of neoclouds focus solely on AI, they’ll deploy a cluster of 80,000 GPUs in weeks—a tempo hyperscalers can not match with their legacy infrastructure.

Objective-built AI structure vs. legacy overhead

Hyperscale clouds are constructed to do all the things, from internet hosting easy net apps to working huge enterprise databases. Consequently, their infrastructure depends closely on virtualization and complicated networking protocols, which provides a “tax” on efficiency. Coaching giant language fashions (LLMs) requires bare-metal efficiency, hyper-fast interconnects (like InfiniBand), and minimal latency. GPUaaS suppliers construct their community topology and storage architectures strictly for AI, yielding greater {hardware} utilization in comparison with the generalized architectures of Azure or Google Cloud Platform (GCP).

Strategic protection and consumer retention 

Microsoft and Google have huge commitments to premier AI companions like OpenAI and Anthropic. When Azure could not spin up GPU capability quick sufficient to fulfill OpenAI’s exploding wants for ChatGPT and GPT-4, Microsoft leased immense capability from CoreWeave and Lambda Labs to bridge the hole. Google equally has partnered with CoreWeave for OpenAI’s multi-cloud workloads. By leasing from neoclouds, hyperscalers can white-label this compute or cross it seamlessly to shoppers, guaranteeing their largest prospects do not defect to a rival cloud on account of capability limits.

Monetary de-risking (CapEx offloading) 

AI {hardware} evolves at breakneck pace; right this moment’s $30,000 GPU may be closely depreciated in only a few years. By leasing capability, hyperscalers shift billions of {dollars} from capital expenditure (CapEx) to operational expenditure (OpEx). If AI demand all of the sudden cools, the specialised neoclouds—not Microsoft or Google—can be left holding depreciating {hardware} on their steadiness sheets.

The NVIDIA allocation technique

NVIDIA holds the keys to the AI {hardware} revolution. To forestall the “Massive 3” (AWS, Azure, GCP) from monopolizing the market—and to hedge towards these hyperscalers creating competing customized silicon (like Google’s TPUs and Microsoft’s Maia)—NVIDIA actively diversifies its buyer base. NVIDIA strategically invests in and allocates its most superior chips (just like the H100, H200, and GB200) to neoclouds like CoreWeave. If Microsoft and Google need fast entry to this extremely sought-after silicon, they’re pressured to strike multi-billion-dollar leasing offers with the suppliers NVIDIA favors.

GPUaaS Supplier Deployment Map

Observe: Knowledge embrace 18 main GPUaaS centered cloud suppliers reminiscent of CoreWeave, Lambda, and Nebius. This isn’t an exhaustive listing of specialist GPUaaS suppliers. Knowledge as of Q1 2026.

By way of funding and infrastructure, CoreWeave is clearly main the pack, having raised over $13 billion in funding over the previous two years. A lot of the funding is reported to be going in direction of the growth of their information middle footprint. In a single yr, the corporate has practically tripled in measurement when it comes to areas. CoreWeave presently has 41 areas in service and yet another deliberate for 2026. The areas are situated within the U.S. (35) and Europe (6).

Lambda has raised $2 billion in funding and operates 16 areas. Lambda is barely extra various geographically than CoreWeave, with areas in Japan (2), Germany (1), India (1), Israel (1), in addition to the U.S. (11). Nebius and Crusoe are additionally notable, every with round $1 billion in funding and 6 and 5 areas in service, respectively. Fluidstack is within the hundreds of thousands when it comes to funding, with 6 deliberate areas.

GPUaaS Supplier Cloud Areas by Firm and Nation

 

Observe: Knowledge as of Q1 2026

Get extra AI market intelligence

There’s much more AI market information and evaluation out there in TeleGeography’s  Cloud and WAN Analysis Service, which delivers information, evaluation, and forecasts on worldwide cloud connectivity and WAN companies, and world WAN market measurement.