Hyperscalers’ Infrastructure Spend Accelerating

Generative AI is undoubtedly transformational. We are at an inflection point where human knowledge, which is mostly written in words, can be ingested, read, learned, and transformed into actionable intelligence by machines. This process allows problems to be solved in minutes and hours, not decades. Generative AI should not be judged by hallucinating ChatBots that are mushrooming everywhere; Large Language Models (LLM) offer much more than that!

We estimate that a Generative AI application requires at least 20 times more computing power than a regular enterprise application. Generative AI applications, in their current form, are much more compute-intensive. As a result, cloud computing was the first birthplace of Generative AI and it is most likely where it will grow exponentially. With that, we believe hyperscalers¹ could emerge as big winners, especially those that remain agnostic to LLM types. Hyperscalers infrastructure buildup is so behemoth that estimating demand a few years out is simply impossible. Hyperscalers are scrambling to expand Generative AI processing capabilities fast, reigniting a multi-year acceleration in their capital expenditure (CAPEX).

For the full calendar year of 2024, our machines project a year-over-year growth rate in the high teens for hyper-scalers’ CAPEX, primarily driven by the Generative AI infrastructure buildup, yet partially offset by reductions in enterprise and telco spending. Enterprise spending on data center hardware and software, such as servers, networking, storage, and security software, as well as cloud infrastructure services, has fallen. Telecom spending on data centers followed a similar trend, influenced by the advanced progress of the virtualization of telecom switch centers to the cloud, already developed 5G infrastructure, and the nearing completion of 5G wireless network rollouts.
 

New AI infrastructure spending spans far beyond just Graphics Processing Units (GPU). We are witnessing unprecedented demand emerging for general Central Processing Units (CPU), memory chips, network storage, accelerators of all types, Field Programmable Gate Arrays (FPGAs), hardware systems, and high-speed networking equipment. Such acceleration is currently so steep that we could potentially observe inflation in electricity prices too.  

Like all other phases of infrastructure, cloud capital expenditure will go through the build-up, absorption, and normalized expansion phases. Currently, we are in the build-up phase and, it is highly probable that within the next 12 months, we may transition into the absorption phase. While we anticipate algorithmic enhancements to reduce training and inference costs of LLM, such savings will be countered by a surge in usage demand as AI spreads its wings across every sector, industry, and society.

Embrace AI, buckle up, and enjoy the ride!