Managing asset depreciation is a core component of most modern businesses, but where you can spread the cost of some assets over a decade, or less, modern GPUs threaten an altogether more aggressive lifecycle. The gains being made generation upon generation, particularly in AI performance and chip efficiency, threaten to accelerate asset depreciation beyond what even some of the largest companies can handle. Now, analysts worry that the rapid pace of AI processing power advances from new generations of GPUs could overwhelm companies riding the AI train.
Most corporations operate with an understanding that their servers will remain relevant for between three and five years, but in the world of "AI factories," where the speed and efficiency of your data center may equate to how much you can earn, even falling one generation behind could be terminal. What happens when the pace of innovation accelerates beyond the potential profitability of the hardware you spent so much time and money investing in?
Keeping up with the times
(Image credit: Microsoft / Nvidia)
For many familiar with the industry, the cycle of new CPU and GPU upgrades is nothing new. But beyond the FOMO, there's nothing problematic about it. However, the optimization equation is slightly different in the context of hyperscalers.
The business world has never had to contend with this kind of upgrade cycle before. Sure, some industries could always benefit from new workstations for their staff, but with AI, there's the potential for your competitors to upgrade, making your hardware less profitable, and therefore your services far less desirable.
A next-gen GPU that can offer, for example, 50% greater performance and/or 30% efficiency savings (some generational upgrades can be greater than that), could make a data center running last-generation hardware decidedly unprofitable. Suddenly, your competitors' services are faster and cheaper to run than your own.
You can upgrade too, but you're competing for a limited pool of hardware, and if everyone else is trying to sell off their old GPUs at the same time, who's buying?
Maybe this wouldn't be so bad if the upgrade cycles were once every half-decade, but Nvidia is pushing for annual GPU releases, and with "Ultra" versions of those architectures often following later. For companies investing tens of billions of dollars in hardware, that could prove entirely unsustainable.
Throw in spiking electricity costs, increasing pressure for eco-conscious data center design, and an ongoing lack of clear profitability for AI in the first place, and it's a disaster waiting to happen.
... continue reading