The biggest problem with a One-Year Cadence of AI Infrastructure system refreshes is the compression of useful life. One of the biggest factors in the reduction of useful life is obsolescence.

AI infrastructure even at the data center level is subject to accelerated obsolescence as the entire infrastructure from the accelerators to the power and cooling system architectures are changing rapidly.

This means data center designs are not durable. Massive AI data centers coming online now are based on two-year old assumptions.

Hyperscalers and cloud service providers have extended the useful life of their servers and network equipment from 3 to 4 years to 6 years. This is for traditional data center equipment which are generally on 3 to 4 year obsolescence cycles.

If NVIDIA and AMD have their way, you are looking at a 1-year cadence, you are not looking at 6-year useful life for your neocloud infrastructure.

Warranties and extended service contracts don’t remedy this issue or the relative power-inefficiency that make their comparative OPEX versus newer systems dramatically unfavorable.

These assets are also facing unfavorable tokenomics that make the ROI window much shorter than for traditional data center equipment.