If 2025 was about securing the shell—land, power, and GPUs—then 2026 is about making the machine work.
The AI infrastructure roadmap is shifting from aggressive capital allocation to operational execution. Three forces make that transition especially important.
The figures below are the projections cited when this article was originally published in December 2025.
Hardware cycles are accelerating
The gap between major hardware generations is shrinking. At the time of publication, Blackwell Ultra was expected to give way to NVIDIA Rubin in late 2026, with a claimed performance increase of roughly 3.3×.
That speed creates a depreciation problem. Infrastructure installed during 2024 and 2025 may face a new generation before it has produced the expected return on investment.
Infrastructure is changing underneath the hardware
Global capital expenditure was projected to rise from roughly $443 billion to more than $600 billion, but the destination of that investment is changing.
Liquid cooling was expected to move from about 20% to more than 45% market share. That is not a component swap. It changes the architecture of the facility and can demand far more maintenance than traditional systems.
Skilled labour is becoming the constraint
Capital can be raised and GPUs can be purchased. Skilled labour cannot be produced on the same schedule.
A single 5 GW build can require around nine million electrician-hours. At the same time, the United States faces a projected shortage of more than 140,000 specialists in high-voltage power and advanced cooling.
The next advantage will not come only from building the largest data centre. It will come from operating one profitably when the people required to run it are scarce.