Data Center - The "Refinery" of the AI Era

Authors: Bu Shuqing, Li Xiaoyin

Source: Hard AI

Data centers are becoming the "refineries" of the AI era!

It is well known that the rapid advancement of AI has led to a massive demand for data processing. In order to meet the booming computational power requirements, a new wave of data center construction is taking place globally, and the data center business is expanding at an exponential rate.

This scenario is reminiscent of the phenomenon of refinery construction that occurred after the second industrial revolution, where the data center in the AI era plays the role of the "refinery" in the era of oil prosperity. It serves as the core infrastructure for data storage, management, and distribution, providing necessary computational power support for AI and machine learning algorithms.

Bank of America Merrill Lynch predicts that by 2025, the total capacity of data centers will increase from 64 zettabytes (ZB) in 2020 to 181 zettabytes, with a high compound annual growth rate (CAGR) of 23%.

In fact, in a report released early last year, Bank of America Merrill Lynch pointed out that global data production is expected to double every 2-3 years. Huang Renxun, the founder and CEO of NVIDIA, also boldly stated that approximately $1 trillion will be spent on AI data center upgrades globally in the next four years.

From the perspective of electricity demand, this trend is particularly evident. Citigroup's latest research report provides an outlook for the global data center market for the first time by evaluating IT equipment loads and utility energy consumption.

In a report released on Wednesday local time, Citigroup's analyst team led by Michael Rollins stated that by 2030, the IT equipment load of global data centers is expected to increase from 33 gigawatts (GW) in 2023 to 100 GW, with a CAGR of 17%.

AI workloads will account for over half of data center capacity

Generative AI will gradually become the dominant force driving the growth of IT equipment loads in data centers.

Citigroup predicts in the report that within data centers, AI workloads will grow at a CAGR of 43%, with the share of AI workloads in the total IT equipment load of data centers increasing from 13% in 2023 to over half by 2030.

However, the report points out that as workloads gradually shift from CPUs to GPUs, cloud computing workloads may be replaced by AI workloads, with an expected CAGR of 8% for cloud computing workloads from 2023 to 2030.

From the perspective of smaller AI infrastructure units - chips, the report predicts that the demand for AI chips will support an annual increase of 4-9 gigawatts of data center IT load.

The report indicates that the total number of AI chips will grow at a CAGR of 22%, from about 31 million pieces in 2023 to 2030.

Given the importance of NVIDIA chips in data center construction, Citi predicts that NVIDIA's AI chip sales will grow at a CAGR of 18% from 2023 to 2030 (calendar year). With the increase in chip power density, the power demand of NVIDIA AI chips will grow at a faster rate of 32% CAGR.

However, Citi also states that with the development of AI chip R&D technology, more and more large data centers may adopt chips with lower power density or self-developed chips for AI inference training. At that time, NVIDIA's market share of chips will decline.

The report predicts that NVIDIA's market share will decrease from around 81% in 2025 to around 63% in 2030.

With a CAGR of up to 43%, AI becomes the dominant force driving data center power consumption growth

With the prosperity of AI chip demand, Citi predicts that from 2023 to 2030, the IT equipment load of the global data center market will increase from 33 gigawatts to 100 gigawatts, with a CAGR of 17%.

Generative AI will become the dominant force driving the growth of IT equipment load in data centers, with its workload growing at a CAGR of 43% until 2030. In contrast, the CAGR of traditional sectors such as core IT and cloud computing in data centers is only 8%, showing slower growth.

Between the traditional and emerging AI sectors, Citi predicts that although AI power demand accounted for only 13% of the total data center demand in 2023, by 2030, it will account for over 50%.

In addition, from 2024 to 2030, AI chips will bring new demand of 4-9 gigawatts per year to the IT equipment load of data centers, accounting for 70% of all new IT equipment load added in data centers.

In conclusion, the global data center market size will grow rapidly in the future, although improvements in data center efficiency will result in energy growth lower than the growth of IT equipment load. Citi writes: Due to the expected improvement in data center efficiency (especially power usage effectiveness, i.e. PUE), we estimate that the peak energy demand required by utilities during the same period will only need to expand at a 15% CAGR.

North America Holds the Largest Market Share, EMEA Shows the Fastest Growth

From a regional perspective, Citigroup predicts that in the coming years, North America will hold the largest market share, while Europe, the Middle East, and Africa (EMEA) will show the fastest growth:

From 2023 to 2030, the IT equipment load of data centers in North America is expected to grow from 12.2 GW to 38.4 GW, with a CAGR of 18%. By 2030, the market share is expected to reach around 38%, making it the largest globally;

The IT equipment load in the EMEA region is projected to grow from 5 GW to 21.6 GW, with a CAGR of 23%, showing the fastest growth, with a market share of 33%;

The IT equipment load in the Asia-Pacific region is expected to grow at a CAGR of 14%, from 2.2 GW to 5.6 GW, with a market share of 9%.

It is worth noting that 93% of the IT equipment load in North American data centers comes from the U.S. market. Citigroup predicts that by 2030, the U.S. data center market will require 5 to 8 GW of peak power annually, as well as 3 to 5 GW of average power.

To meet these power demands of data centers, Citigroup forecasts that the size of the U.S. utility market will grow at an annual average rate of 6%. Utilities will need to add 3-5 GW of average supply capacity each year, equivalent to about 0.6%-1% of last year's total U.S. electricity generation.

Due to the expected strong demand for data centers in the future, Citigroup believes that in the coming years, the U.S. power supply will remain at a relatively tight level.

Amid the Wave of Data Center Construction, Who Will Benefit the Most?

According to Citigroup's research report, by 2030, global data center IT equipment load net growth will reach 9-11 GW, with approximately 70% coming from the third-party hosting market, and the remaining 30% from tech giants such as Microsoft, Amazon, Meta, and Google building their own hyperscale data centers.

Citigroup points out that the current prices of data center-related stocks may not fully reflect the market's demand for data center capacity growth, especially for the demand for hyperscale leasing, thus maintaining an optimistic outlook on data center stocks The institution also believes that data center real estate investment trust company Digital Realty and data operation giant Equinix will continue to benefit from the growth of the data center market. The former achieved a record high in leasing revenue in the first quarter, with about half of the orders in the quarter related to AI, while the latter holds a 5% share in the global data center colocation market