Perhaps depreciation is not typically a key business model issue, but that seems not to be the case for hyperscalers who have extended the useful lives of their servers and networking equipment.
Historically, hyperscalers depreciated servers over three years. These days, server depreciation occurs over as much as five years to six years and networking gear is depreciated over as much as five to six years.
Some observers may not like the practice, as longer depreciation periods extend the period when revenue is recorded against the capital investment. That essentially lowers the hurdle rate for making the investments.
The concern, in some quarters, is that the treatment of useful economic lives of the GPUs is being extended so the firms buying the GPUs have more time to record revenue. And that might potentially obscure the actual business cases for deploying the GPUs.
The concern is that longer depreciation cycles mean capitalized GPUs can be deployed to generate training and inference revenues before significant depreciation expenses have to be recorded, distorting the payback.
Others disagree, arguing that the useful business life of any generation of GPU is longer than most assume. Some would argue that the rapid functional depreciation of GPUs for cutting-edge AI model training (typically after about three years) does not impair GPU value, and that the longer cycles are justified. In fact, some argue that useful GPU lives can stretch out nearly a decade.
Even GPUs no longer on the cutting edge have value in a tiered ecosystem for compute usage, where hardware shifts from high-value training to sustained value in inference operations.
So there is a "value cascade" model. The newest GPUs handle demanding training workloads for frontier models, while depreciated ones (two to three years old) are repurposed for inference, fine-tuning, or less intensive tasks like batch processing.
Consider the business case for an Nvidia H100 with a cost of $250,000. Assume
depreciation over six years, straight-line ($41,667/year for six years).
Hourly Rate: from $6.15 initially, dropping 70 percent to $1.85 by year 10
Utilization: from 100 percent initially, decreasing to 50 percent by year 10
Overhead: power at $0.10/kWh and networking costs of $6,000/year
For example, Azure announced the retirement of its original NC, NCv2, and ND-series VMs (powered by Nvidia K80, P100, and P40 GPUs) for August/September 2023. Given these GPUs were launched between 2014 and 2016, this implies a useful service life of seven to nine years.
More recently, the retirement of the NCv3-series (powered by Nvidia V100 GPUs) was announced for September 2025, approximately 7.5 years after the V100’s launch.
To be sure, chip wear and tear (thermal and electrical stress) does happen. But useful lifespans can be manipulated by controlling the utilization rates over time. That is a business decision: run at lower utilization rates to prolong useful life, or run at higher rates to maximize efficiency of an expensive asset.
But it seems clear enough that the value cascade is part of the reason for rapidly-declining cost of inference operations.
So there might be suspicion in some quarters about accounting decisions that obscure the payback on big investments in GPUs. But others argue the value cascade means the useful business life of any generation of GPU is much longer than some imagine.
No comments:
Post a Comment