1 hour ago
22
The cloud computing hierarchy is getting reshuffled. A new breed of specialized data center operators, known as neoclouds, is finding itself in an increasingly powerful negotiating position with the very hyperscalers that have dominated the industry for over a decade.
The reason is straightforward: everyone wants Nvidia GPUs, and neoclouds have them.
The deals reshaping AI infrastructure
Microsoft alone has committed more than $33 billion in capacity agreements with neocloud providers. The largest single arrangement is a $19.4 billion deal with Nebius, a specialized GPU cloud firm, which will give Microsoft access to over 100,000 Nvidia GB300 chips.
The neocloud category includes companies like CoreWeave, Nebius, Lambda, and Crusoe, all of which have built their businesses around a single thesis: purpose-built GPU infrastructure for AI workloads, nothing else. No general-purpose compute, no sprawling product catalogs, no legacy architecture weighing down their cost structures.
CoreWeave’s recent financial performance illustrates how well the thesis is playing out. The company reported Q2 2026 revenue of $2.575 billion, representing a 112% increase year-over-year.
Why neoclouds can undercut the giants
The pricing advantage these firms enjoy isn’t marginal. Neocloud GPU instances typically cost 60-70% less than equivalent offerings from traditional hyperscalers like AWS, Azure, and Google Cloud.
Their infrastructure is designed from the ground up for GPU-intensive AI workloads, which means they don’t carry the overhead of supporting the thousands of other services that hyperscalers bundle into their platforms. Leaner operations translate directly into lower prices for customers who only need one thing: raw GPU power.
Nvidia’s role as kingmaker
None of this would work without Nvidia’s active involvement. The chipmaker has positioned itself as something between a supplier and a strategic patron to the neocloud ecosystem, providing priority access to hardware, making equity investments in neocloud companies, and even offering revenue backstops for unsold capacity.
Nvidia is essentially guaranteeing that neoclouds won’t get stuck with idle GPUs. It’s a financing mechanism that makes it dramatically easier for these younger companies to raise the substantial debt needed to buy server fleets in the first place, because lenders see customer contracts and Nvidia guarantees as solid collateral.
For Nvidia, the motivation is straightforward. Diversifying its customer base beyond the handful of hyperscalers that currently buy most of its chips reduces concentration risk and creates competitive tension among buyers.
The risks lurking beneath the growth
Customer concentration is the most obvious risk. When a significant portion of your revenue comes from one or two hyperscaler contracts, losing even one relationship could be devastating. Microsoft and OpenAI represent massive revenue streams for several neocloud providers, and that kind of dependency cuts both ways.
GPU depreciation is another concern. Nvidia releases new chip architectures on roughly annual cycles now, and each generation delivers substantial performance improvements over the last. A fleet of GPUs purchased today could be worth meaningfully less in 18 months, not because the hardware breaks, but because newer, faster chips make the older ones less competitive on a price-per-performance basis.
That dynamic creates a treadmill effect. Neoclouds need to continuously reinvest in the latest hardware to maintain their edge, which requires sustained access to capital markets. The significant debt levels many of these companies carry make that reinvestment cycle a high-wire act if demand softens even slightly.
Then there’s the infrastructure challenge. Deploying tens of thousands of high-powered GPUs requires enormous amounts of electricity and cooling capacity. Securing power in the right locations at the right price has become one of the most significant bottlenecks in AI infrastructure buildout, and neoclouds are competing for those resources against the very hyperscalers they’re partnering with.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·