For more than a decade, the default infrastructure play has been simple. Move more workloads to the cloud, retire on-premises gear and treat the data center as a cost to shrink. That logic held while public cloud offered near-limitless capacity at predictable prices, and AI demand is now testing both halves of that assumption.
The pressure is now showing up in hardware. In one recent case, a server originally quoted at $20,000 reached $39,000 by the time it shipped, driven almost entirely by memory and component costs that moved while the deal was still open. Government price data shows the same trend across the market. The US producer price index for computer storage devices rose nearly 11% in the two months from December 2025 to February 2026, and the index for semiconductors and electronic components climbed about 13% between October 2025 and February 2026.
Public cloud was supposed to insulate companies from exactly this kind of volatility, and increasingly it falls short of that promise.
Capacity Is No Longer Guaranteed
Underneath the price increases sits a harder physical constraint. Global electricity demand from data centers grew 17% in 2025, while consumption at AI-focused facilities surged 50% in the same year, and that AI-specific data center capacity more than tripled over 18 months. Power generation, land and grid connections now gate how fast new capacity comes online, and the IEA expects data center expansion to drive close to half of all US electricity demand growth through 2030.
Customers feel those limits as scarcity. The major public clouds are not “full” in any literal sense, yet organizations are running into more denied requests for storage, memory, CPUs, new subscriptions and specific instance types.
Elastic, on-demand capacity was the core promise of cloud, and in the busiest regions that promise now comes with conditions. In the field, the denials rarely arrive as a flat no. A request for a specific GPU-backed instance or a memory-heavy SKU sits in a quota queue for days, or clears only in a region far from where the data and users live, which quietly forces a redesign nobody budgeted for.
Together, rising prices and tightening capacity change the planning math. A migration that assumes a given region and instance type will be available on the planned timeline now carries real schedule risk. Lift-and-shift projects tied to end-of-support deadlines face the most exposure, because they often target the busiest instance types in the most congested regions.
Workload-First Beats Cloud-First
Cloud remains the right home for bursty, unpredictable and globally distributed workloads, and that has not changed. What has changed is the habit of sending every system there by default. A growing number of enterprises are revisiting hybrid models they had written off a few years ago, moving selected workloads back to private or on-premises environments where the economics and the control line up better.
A workload-first approach starts from the characteristics of each system rather than a blanket mandate. Three factors usually decide where a workload should run.
- Cost over the asset’s full life. Steady, high-utilization compute often runs cheaper on owned or collocated hardware after factoring in egress fees and sustained usage, while spiky demand favors the cloud’s pay-as-you-go model.
- Performance and data gravity. Latency-sensitive systems and large, frequently accessed datasets perform better close to where the data lives, whether that means on-premises, a private cloud or the edge.
- Control and compliance. Workloads with strict data residency or regulatory requirements often need a fixed location that public cloud regions cannot always guarantee on demand.
Plan for Tighter Supply
The infrastructure decisions that worked five years ago assumed compute was cheap, plentiful and a phone call away. For AI-adjacent workloads, those assumptions no longer hold cleanly in 2026. Teams that treat capacity as a coordinated program across platform, FinOps and security will weather this better than teams still optimizing for a single monthly cloud bill.
Waiting carries a concrete and compounding cost. A migration tied to an end-of-support deadline stalls the moment a busy region denies its preferred instance type, and a team that retired its on-premises footprint years ago has no fallback left. The $39,000 server shows where prices are heading, and a budget built on month-to-month cloud spend absorbs that kind of jump late, after the project is already committed.
Teams that plan for scarcity now choose where their workloads run and when, while teams that bet on the old abundance returning hand that choice to denied requests, stalled cutovers and hardware bills that keep climbing.