As far as the market implications are concerned, this
phenomenon is a prime example of the Jevons Paradox.
The cost of the cheapest Large Language Model
(LLM) with a minimum Massive Multitask Language
Understanding (MMLU) score of 83 has decreased by
approximately 10x every year from 2022 (GPT-3, 60b)
to 2024 (Llama 3.2, 3B). This reduction in the cost of AI
training and inference does not, however, lead to reduced
consumption. Instead, like improvements in automobile
fuel efficiency that led to more car sales and larger cars,
the declining price of intelligence is expected to fuel an
exponential increase in the demand curve for AI.
The scale of investment is accelerating this cost-per-unit
benefit. Hyperscalers are making enormous investments
in GPUs and infrastructure because, in the race to deliver
services and achieve Artificial General Intelligence
(AGI) they understand that the risk of under-investing is
dramatically greater than the risk of over-investing.
In the same way, the market is diversifying. LLMs are
enabling the growth of a new generation of specialist
Small Language Models (SML) which use just a few
billion parameters to train, are more adaptable to mobile
devices, and deliver niche services using exclusive data
sets. This in turn drives up broader AI revenues, expands
the market, and contributes to lower unit costs
The resulting explosion in demand will yield substantial
returns, with increasing efficiency driving rather than
limiting overall utilization and innovation.
Data center demand: through the roof
Roughly half of AI investment is going on data centers.
AI colocation revenue for training and inference is
forecast to grow at an extraordinary 77% CAGR from
2025-2030, reaching $134 billion by 2030. By 2030, AI
is expected to account for approximately 44% of total
global data center colocation market revenue, all from a
near standing start a few years ago.
The steepness of this curve has huge implications for the
data center industry. Demand for data center capacity is
projected to outstrip supply from 2027-2030, with annual
global demand reaching nearly 90 GW by 2030, exceeding
supply by as much as 500%.
