Thunderclap Newman’s 1969 UK #1 hit ‘Something in The Air’ referred to a very different revolution, but 55 years later the AI-revolution is upon us - that is abundantly clear.
The rapid adoption rate of generative (gen) AI tools is not only reshaping business processes but driving unprecedented demand for computing power, with far-reaching implications for data centres, energy consumption, grid infrastructure, utilities and upstream/midstream operators.
Data centre growth is accelerating, driven by gen AI and ‘traditional’ cloud demand
Power demand from data centres is straining existing grid infrastructure
Short timelines and 24/7 reliability make gas a natural energy partner for data centres
Such AI-led growth in power demand clearly presents challenges for incumbent utilities and data centres but offers fresh and direct investment opportunities for upstream & midstream gas operators, new market entrants as well as innovative technologies and services.
The adoption rate of generative (gen) AI-led tools and content reportedly far exceeds that of previous disruptive technologies such as the internet, smartphones and PCs. Indeed, whether or not people actively seek out ChatGPT or any other gen AI tool, gen AI is fast becoming ubiquitous as it permeates the very fabric of Big Tech’s products and services - such as Apple’s iPhone and Apple Intelligence, Meta’s social media apps, even its Rayban glasses, Adobe’s PhotoShop, Google Docs and even the humble Google Search!
Gen AI is playing an ever-growing role across all manner of businesses - from media, energy and healthcare to pharmaceuticals, financial services and manufacturing.
Why this explosive demand for gen AI? Gen AI has effectively ‘democratized’ AI, with a wide range of user-friendly tools that provide a broad array of creative content - written, visual, audio, code etc - from unstructured data. By contrast, traditional AI relies on structured data to deliver analysis and/or automation for specific tasks, thus limiting its broad appeal.
A May 2024 McKinsey global survey reported that 55% of respondents regularly use gen AI tools at business and/or home, up from 38% a year ago. A more recent EY survey reports that Gen AI adoption in the workplace has skyrocketed from 22% in 2023 to 75% in 2024!
The creative industries are no doubt early adopters and regular users, but 3 in 4 people across all workplaces, up 3-fold in a year, seems a somewhat extravagant claim for most business roles - but then perhaps I’m the dinosaur!
Nevertheless, it is all too easy to forget that our seemingly insatiable appetite for data - whether it be streamed music or videos, social media, a Google search or a ChatGPT query - is satisfied by a fast-growing global network of datacentres, almost 11,000 at year-end 2023 according to industry estimates - all of which require reliable 24/7 electricity to power and cool their numerous racks of CPUs1, GPUs1, storage media and networking equipment.
‘FOMO’ - The Race Is On As Hyperscalers Spend Huge Sums To Dominate Gen AI
Global investment in data centres has underpinned the scale and ubiquity of cloud-based data processing, streaming and storage services for businesses and consumers alike. But the race is now on - particularly amongst the ‘hyperscalers’2 - to build massive AI-ready data centres to support the rapid adoption of AI, gen AI in particular, following the release of OpenAI’s ChatGPT just two years ago.
Almost half of all data centres worldwide are in the United States, where annual investment is up almost four-fold since 2019 and has doubled in the last two years alone. Indeed, in 2023, overall investment by the Top Three ‘hyperscalers’ - Google, Microsoft and Amazon - global leaders in AI adoption and data centre installation, exceeded that of the entire US oil and gas industry – totalling around 0.5% of US GDP. Other major economies, such as China and the EU, are also witnessing major growth in data centre investment and construction.
Indexed US$ Investment in US-based Datacentres - Up Almost Four-Fold Since 2019
This huge global scale-up in data centres will continue, likely doubling in number over the next 4 - 5 years as industry players - led by the ‘hyperscalers’ - not only seek to meet surging consumer and business-led demand for gen AI solutions but establish market dominance.
AI's Voracious Appetite For Power Is Straining The Electricity Grid
But, unlike prior (pre-2020) capacity and workload expansion, this rapid buildout of data centre capacity is now accompanied by an outsize near-term surge in power demand.
This issue was recently highlighted by Dr. Sultan Al Jaber, UAE Minister of Industry & Advanced Technology in his opening remarks at this year’s ADIPEC conference: “… the exponential growth of AI is creating a power surge that no one anticipated 18 months ago, when ChatGPT took off."
Until 2020, a host of efficiency improvements ‘capped’ overall data centre energy usage despite ever-growing workloads. But since 2020 diminishing returns on such measures have seen the power usage efficiency (PUE)3 of data centres largely stall at 1.5-1.6.
Gen AI-led workloads are energy-intensive: a typical ChatGPT query requires 10-fold the power of a Google search, according to the EPRI4.
High adoption rates for gen AI workloads are displacing other simpler (less process-intensive) workloads. Even the humble Google search has become more energy-intensive, according to MIT, with gen AI tools now embedded in the search function.
Finally, the ‘Jevens Paradox’5: further advances in the computational and/or PUE of data centres will inevitably drive down the unit cost of gen AI - but the resulting growth of gen AI demand may well increase rather than reduce overall energy usage, as seen with efficiency improvements in many other areas of commerce and resource use.
These factors, combined with the ongoing ‘arms race’ amongst the global ‘hyperscalers’, have created a ‘perfect storm’ for data centre developers, utilities and grid operators alike.
Land and power constraints - limited access to reliable power, grid infrastructure and equipment - are hampering the scale-up of existing data centres and limiting the choice of location and pace of construction and commissioning of greenfield facilities.
By way of context, a decade ago, a data centre that required 25 - 30 MW of power would be considered large. These days, the largest data centres now require 100 MW - 250 MW of power – more than an average US EAF6 steel mill, nearing that of an aluminium smelter!
Hardly surprising, therefore, that this surge in energy demand is already placing strains on the US electricity grid, parts of which are now struggling to provide enough new supply.
In Santa Clara, CA and Northern Virginia, the latter reputedly the data capital of the world, utilities cannot build grid infrastructure fast enough and there are growing concerns over the future availability of sufficient power to meet additional demand in these regions.
And Microsoft and Oracle have both announced plans to build data centres – effectively vast data ‘campuses’ – with 10-fold the power requirements, well above 1 GW!
Outside the US, Amsterdam and Singapore have already imposed moratoriums on new data centres; Ireland, with its grid already under strain, has similarly banned any new grid connections for data centres near Dublin until 2028. Yet by 2031, Ireland’s grid operator estimates that data centres will consume 28% of the country’s power demand!
Data Centre Energy Consumption Projected To Grow Four-Fold By 2030*
Data centres currently consume ca. 1.5% of global electricity generation output (2023: 29295 TWh7) but 3.7% of US electricity generation output (3,973 TWh7), more than twice the global share due to their high level of concentration within the US.
With gen AI a novel and fast-growing technology, any forecasts of data centre energy usage will hinge on many yet unknown factors, such as the long-term pace of gen AI adoption; GPU chip supply chain constraints; improved GPU chip designs; migration toward smaller, more efficient gen AI models; cloud-based vs. edge-based (ie low latency) computing et al.
Of the myriad forecasts of future AI-led power demand, we have chosen, for better or worse, to stick with McKinsey’s forecasts – both global and US-centric.
McKinsey’s low-case, mid-case* and high-case forecasts imply that global data centre energy usage will increase three-fold, four-fold, potentially five-fold respectively by 2030.
Global Data Centre Energy Usage, 2023 – 2030E
US Data Centre Energy Usage – Mid-Case Forecast, 2023 - 2030E
Within the US, applying a similar 22% ‘mid-case’ CAGR, data centres will, at over 600 TWh per annum, consume almost 12% of overall US electricity generation output by 2030 - three-fold their current share.
US power generation capacity will need to grow at a CAGR of near 4% to 2030, ten-fold the near-stagnant CAGR of 0.4% observed over the last decade – to meet forecast demand from industrial reshoring and electrification, transport electrification and data centres – requiring ca. 140 GW of additional power generation capacity by 2030.
Data centres will consume ca. 40% (52 GW) of such additional power generation output.
Given the recent US election result, the pace of electrification of both industry and transportation may well slow as renewables and EV subsidies are abolished; conversely, the pace of industrial reshoring may well accelerate as non-US companies seek shelter from increased tariffs.
It’s far too early to assess the net impact of the new administration on such factors.
However, we do not expect the voracious appetite for gen AI-led workloads to be affected.
Natural Gas Will Play An Important Role In Powering Data Centres …
The ultimate determinant of data centre capacity growth will of course be power – reliable 24/7 power either generated in close proximity or amply available via grid infrastructure. Even McKinsey’s ‘low-case’ forecast – a three-fold uplift in power demand - could result in near-term shortfalls in power generation capacity or grid infrastructure in certain geographic locations, as highlighted earlier.
Seeking to sidestep such ‘grid-lock’ (excuse the pun), ‘hyperscalers’ and other data centre developers would ideally power their data centres from local renewable energy sources, thus bolstering their ‘green’ credentials.
However, reality requires that they prioritize reliability over emissions.
Even when renewables are twinned with battery storage, the absolute requirement for 24/7 reliability mandates an alternate source of baseload power. Furthermore, the pace of this AI-led ‘race’ demands a swift, expedient source of power to avoid the risk of falling behind.
Natural gas provides a swift, mature, cost-effective (given current North American gas prices) and flexible source of power generation via CCGT (combined-cycle gas turbine) power plants.
Indeed, the AI-led surge in energy demand, largely due to the ‘hyperscalers’, is already boosting the US natural gas demand by up to 4 bcf/d, according to Kinder Morgan research.
Seldom missing an opportunity to talk their own book:
BP’s CEO commented on the Q3 24 investor call: ‘… the hyperscalers are driving crazy demand into natural gas right now. We could see something like another 10 bcf/d of demand by the end of the decade. Hard to predict right now, but crazy, crazy demand coming out for baseload.’;
Chevron’s CEO also recently said: ‘Natural gas will help power the rapid growth of artificial intelligence with its insatiable demand for reliable electricity … AI's advance will depend not only on the design labs of Silicon Valley, but also on the gas fields of the Permian Basin.’; and
EQT’s CEO has also suggested that data centre-driven power demand could prompt up to 10 bcf/d of incremental gas demand by 2030.
Despite such boosterish comments, nobody yet has a good handle on future natural gas demand from US data centres: 2030 forecasts range from 1.3 bcf/d to as high as 18 bcf/d.
That said, as we’ll see in a moment, McKinsey’s estimates lie in the ‘middle of the fairway’.
… But Will Data Centres Play An Important Role In The US Gas Market?
To set the context for any such gas demand forecasts, US natural gas set several new record highs in 2023: dry gas production of 103 bcf/d, natural gas consumption of 89.1 bcf/d, of which power generation accounted for 35.4 bcf/d, and LNG exports of 11.9 bcf/d.
There are six LNG projects already under construction that should increase LNG export capacity by 80% from the current 14 bcf/d to over 25 bcf/d by 2028. In addition, there are further pre-FID LNG projects which, if approved, will increase post-2028 export capacity.
McKinsey’s ‘mid-case’ growth in US data centre power demand (22% CAGR) will require an incremental 52 GW of available power by 2030, equivalent to 6.8 bcf/d of incremental gas demand if gas were to meet 100% of the energy shortfall, 4 bcf/d at a more likely 60:40 split with an alternate ‘solar plus battery’ renewable source.
Likewise, the ‘high-case’ scenario would generate additional gas demand of almost 6 bcf/d or 10 bcf/d if gas were to meet 60% or 100% of the energy shortfall respectively.
Against current overall US gas production of over 100 bcf/d, such forecasts for incremental
data centre-led gas demand appear modest, particularly when compared with prospective LNG export capacity of 25 bcf/d by 2028.
With all the column-inches and CEO comments related to AI, data centres and gas demand, perhaps a degree of AI-led ‘hype’ has entered the US gas market!
Bringing The Data Centres To Energy, Rather Than The Energy To Data Centres …
However, that is to miss the point – the incremental gas demand may prove to be relatively modest, but such fresh gas demand can still offer opportunities to nimble, geographically-advantaged upstream gas operators, midstream pipeline companies and entrepreneurs.
With data centre clusters already straining local grid infrastructure, ‘hyperscalers’ and others are necessarily commissioning new data centres in new locations across the US.
Unlike other major sources of gas demand such as cities, energy-intensive industry, even LNG export facilities, new data centres can be located wherever sufficient reliable power resources exist or can be readily brought online.
So, rather than bringing the energy to data centres, bring data centres to the energy!
Of course, choosing where to construct a new data centre is never that simple - the location must also provide access to sufficient water for cooling and fibreoptic networks for internet connectivity. Internet latency (response time) may also play a role - e.g. low internet latency is far less critical for AI training - the energy-intensive process of building an AI model - than AI inference - the real-life, real-time operation of the AI model.
… Offers Fresh Opportunities For Upstream, Midstream Operators & ‘CoLos’
Grid congestion and the ensuing and direct convergence of digital infrastructure and power is creating investment opportunities for upstream, midstream and 3rd-party CoLo8 developers.
Upstream operators with ‘stranded’ assets - significant gas resources that currently offer little to no economic value due to physical distance from viable ‘traditional’ markets - can vertically integrate or partner with IPPs (independent power producers) to power greenfield data centres constructed in proximity to such resources.
By way of example, assuming CCGT energy conversion efficiency of 60%, a 250MW data centre would, if solely powered by gas, consume ca. 33 mmscf/d of natural gas.
In the Marcellus-Utica region of the US, regulatory hurdles and permitting challenges have ‘bottlenecked’ gas developments for many years, one example being the troubled Mountain Valley gas pipeline project which finally came onstream in June 2024, ten years from proposal.
Gas and pipeline operators and construction companies in the region are all pitching for business or fielding requests from ‘hyperscalers’ and others seeking fresh data centre locations.
In Canada, recent months have seen a growing number of gas operators shut in gas production as burgeoning storage surpluses have driven AECO gas pricing down to as little as C$0.05 per mmbtu. While the long-awaited start-up of the LNG Canada project is expected to pull up AECO pricing in 2025, some gas operators may view fast-growing data centre power demand as an alternate attractive, stable long-term PPA-led market for their gas.
Midstream operators can exploit the advantaged economics of siting greenfield data centres near their major gas pipelines by partnering with IPPs and/or data centre operators.
Large US midstream players such as Kinder Morgan, Williams and Energy Transfer all speak of growing project backlogs to serve data centre power demand as it migrates across the US.
Canadian midstream player TC Energy, which operates 58,000 miles of gas pipelines across the US, Canada and Mexico, also views the growth of data centres in North America as a significant business opportunity. On a recent investor call, TC Energy’s COO said ‘… over 60 percent of the more than 300 data centres currently under construction or proposed in the US are within 80 km of TC Energy’s existing natural gas pipeline system.’ Off-grid power generation, a growing preference for data centre operators, can be fed via short laterals to its gas pipeline network.
CoLo developers build large data centres to provide turnkey solutions for ‘hyperscalers’ and other end users, traditionally within or nearly existing US data clusters. They are increasingly scouring the entire US - no doubt alongside ‘hyperscalers’ - to secure advantaged locations to meet growing demand for data centre capacity.
Power being the single largest hurdle - CoLo developers are signing PPAs directly with renewables, natural gas and even geothermal energy providers rather than utilities or constructing facilities adjacent to power plants to avoid grid issues e.g. Talen Energy’s CoLo data centre powered by a nuclear power plant and recently sold to Amazon for US$650 million!
The ‘hyperscalers’ are also looking at nuclear power - Microsoft and Constellation Energy plan to reopen the Three Mile Island nuclear plant - the site of the worst nuclear accident in US history. If approved, Microsoft would purchase all the power for the next 20 years to feed its voracious AI energy needs. Likewise, Oracle plans a gigawatt-scale data centre powered by three small nuclear reactors (SMRs), according to founder Larry Ellison.
Conclusion
Wherever traditional markets encounter disruption, investment opportunities abound: new market entrants displace incumbents, new technologies, new services etc. While direct upstream or midstream investments lie outside our mandate, PillarFour monitors all energy-related markets to identify promising trends and investment opportunities in related technologies and services.
1 CPU: central processing unit; GPU: graphic processing unit
2 Large technology companies that operate massive, global networks of data centres with substantial cloud storage & AI computing capabilities
3 PUE: the ratio of total facility energy usage (incl. heating/cooling etc) to IT-related usage (computational, storage, networking etc)
4 Electric Power Research Institute
5 'The Jevons Paradox', was first expressed in 1865 by William Stanley Jevons in relation to the use of coal for steam engine
6 EAF - electric arc furnace: typical power per tonne of steel: 450 kWh, average capacity of US EAF mill: 540,000 tonnes/year
7 2024 Statistical Review of World Energy, Energy Institute; McKinsey analysis
8 ‘CoLo’ - a large data centre facility that rents out rack space to third parties for their servers or other network equipment.
Download the PDF
This material is intended for information purposes only. This material is based on current public information that we consider reliable, but we do not represent it as accurate or complete, and it should not be relied upon as such. We seek to update our research as appropriate, but various regulations may prevent us from doing so.
Estimates, opinions and recommendations expressed herein constitute judgments as of the date of this research report and are subject to change without notice. PillarFour Capital Partners Inc. does not accept any obligation to update, modify or amend its research or to otherwise notify a recipient of this research in the event that any estimates, opinions and recommendations contained herein change or subsequently become inaccurate or if this research report is subsequently withdrawn.
No part of this material or any research report may be (i) copied, photocopied or duplicated in any form by any means or (ii) redistributed without the prior written consent of PillarFour Capital Partners Inc.
Website links or e-mail communications may contain viruses or other defects, and PillarFour Capital Partners Inc. does not accept liability for any such virus or defect, nor does PillarFour Capital Partners Inc. warrant that e-mail communications are virus or defect free.
This document has been approved under section 21(1) of the FMSA 2000 by PillarFour Securities LLP (“PillarFour”) for communication only to eligible counterparties and professional clients as those terms are defined by the rules of the Financial Conduct Authority. Its contents are not directed at UK retail clients. PillarFour does not provide investment services to retail clients. PillarFour publishes this document as a marketing communication and NOT Independent Research. It has not been prepared in accordance with the regulatory rules relating to independent research, nor is it subject to the prohibition on dealing ahead of the dissemination of investment research. It does not constitute a personal recommendation and does not constitute an offer or a solicitation to buy or sell any security. PillarFour consider this note to be an acceptable minor non-monetary benefit as defined by the FCA which may be received without charge.
This note has been approved by PillarFour Securities LLP (FRN 722816) which is authorised and regulated by the Financial Services Authority.