AI, and generative AI in particular, is expected to greatly enhance productivity within work processes. Some studies estimate that generative AI could contribute between $2.6 trillion and $4.4 trillion annually to the economy.
However, AI infrastructure is costly because the underlying algorithmic problems are extremely computationally intensive and this means there’s a potential gap between demand and the capacity needed to meet it.
A new report from OpenNebula Systems identifies key challenges, including data availability, scalability of centralized systems, power constraints, and the intricacies of accelerator manufacturing.
The report estimates that this year a single AI training run will demand one exaflop of performance (an exaflop is not a sandal you go running in, it’s 1018 or one quintillion floating point operations per second), with projections reaching between 100 and 10,000 exaflop/s by 2030.
Analysis the top 10 LLM models (i.e. those released by OpenAI, Google DeepMind and
Meta) by compute, and have estimated that their training processing needs have increased by a factor of four of five times a year since 2010. If this continues at the same rate then by 2030 the computing demand for AI training will be 10,000 times higher.
There’s an effect on power consumption too, various studies indicate that data centers’ electricity consumption is projected to grow by five percent annually until 2030, resulting in an increase of 1.5 to two times current levels by that time.
The report’s author’s note, “From a strategic perspective, the most effective way to meet future AI processing needs is through the development of new distributed and decentralized systems. Leveraging a continuum of HPC, cloud, and edge resources will be crucial for addressing the intensive processing demands of AI training and the low-latency requirements of AI inference.”
You can find out more on the OpenNebula site and there’s an infographic summary below.
Image credit: Kittipong Jirasukhanont/Dreamstime.com
