Grid Challenges & Opportunities in AI Growth

Article published in the July 2024 Newsletter of the Global Smart Energy Federation.

In earlier articles, I have written about rising HVAC-R cooling loads and policy limitations of providing green-energy to data centers. This article examines AI growth as a strategic national priority and its huge power demand.

?Many nations are encouraging Artificial Intelligence (AI) and Data Science as a competitive edge. These have the potential to enhance (or disrupt) many aspects of our industrial/commercial ecosystem. The United States AI policy announcement of January 29, 2024, highlights the importance of AI leadership for the country. Other nations have followed suit and have their own doctrines. Many view AI as the new Industry 4.0 in terms of technology revolution. Expect this policy area to undergo further revisions in this decade.

The AI policy focus is leading to a huge growth in new AI data centers (AIDC). This in turn is reshaping national and state energy policies to meet such large loads. The race to accommodate AI (or risk losing to overseas competition) is the key driver. The current global list of 15 biggest (pre-AI) data centers are listed below to show the investment/size of such facilities and their large power requirements:

?1.?????? USA:

a)?????? Switch Super NAP, Las Vegas, NV: 1000 acres, 7.3 million sq. ft; $5B; 500 miles fiber

b)????? Quincy Data Center, Quincy, Washington State: Agricultural location; 5 large investors

c)?????? Phoenix One Data Center, Phoenix, NV: 630,000 sq. ft; 24-hour backup local fuel storage

d)????? NAP of Americas, Miami, FL: 750,000 sq. ft; connectivity 150 countries; 15 subsea cables

e)????? Utah D/Center, Bluffdale, Utah: $2.5b; 1.5 mil. sq. ft; Cyber HQ; cooling 1.7 million gpd

f)??????? Lakeside Tech Center, Chicago, Ill.: 1.1 mil. sq. ft; 100MW; 8.5-million-gallon brine cooling

g)?????? Apple’s Mesa Data Center, Mesa, Arizona: $2B; 1.3 mil. sq. ft; 300-acre PV generation

h)????? Iron Mountain Underground, Butler, Pennsylvania: 1000 acre abandoned limestone mine

i)??????? Switch Citadel Campus, Tahoe-Reno, NV: 2,000 acres; 7.2 mil. sq. ft; 130 MW (to 650 MW)

?2.?????? EU/UK:

a)?????? CLW1 NGD New Port Campus, Marshfield, Cardiff, UK: Largest in UK/EU; 180MW sourced directly from nearby hydroelectric power; dedicated 400KV connection

b)????? Bahnhof Center, Sweden: 100 ft underground mountain (former nuclear shelter); Environmentally cooled

?3.?????? China:

a)?????? China-Telecom Inner Mongolia Information Park, Mongolia; 3b$ build; 10 million sq. ft; 100MW+?

b)????? Range Group, Luongfong, China: IBM Collaboration; 6 million sq. ft.; Next-gen cloud computing; 150 MW

?4.?????? India:

a)?????? YottHNM1 Data Center Park, Panvel, India: 2nd. largest in Asia; 600 acres; 820,000 sq. ft; 150MW

b)????? Tulip Data Center, Bengaluru, India: IBM collaboration; 1 million sq. ft; High efficiency rack cooling

??AI data centers in comparison to the above are much larger and consume even more power. To put AI in context of traditional search engines, a single chat GPT query consumes 9 times the energy as a typical internet search. It is estimated that AI processing and computing needs will be 7x – 10x typical data processing and search engines. These new AI data centers will typically need about 1-3 GW in electricity demand per site.

?Per IEA’s mid-2024 report, energy demand from global AI data centers could increase three-fold to 3% by 2026 from the current 1.3% (2022) levels. This would be in addition to EV charging power demand of 2% (2026). Individual country situations amplify this with Ireland currently at 18% demand from data centers, while Singapore is about 7%. Grid constraints to supplying new data centers are emerging in many countries requiring new planning and delivery methods.

?In the United States, AI demand is centered in a few specific states such as Texas, Virginia, Ohio, California and Nevada. These areas will likely experience a 5-year growth of 15% in electricity demand. AEP in July 2024 disclosed plans to supply 15 GW by 2030 to AI data centers in Texas and Ohio (about 42% of its own current peak load). PG&E expects 3.5 GW of new AI data center load by 2029. To complicate matters the typical 2-year build-timetable of such GW-scale AI facilities is not compatible with utility lead times of 5-7 years.

?Despite projections of AI data centers utilizing efficient chipsets, smaller form-factors, and more computing power, impacted utilities are seeing a tripling in annual growth-rate from 2% (last few decades) to 7-10%. At a district level this could translate to 15-20% growth and at a small city/town level to as high as 50%. To meet this huge power demand, several key questions need to be examined:

a)?????? Can utilities drastically shorten their planning and execution cycles

b)????? Will interconnections address supply issues

c)?????? What is the Fossil/Nuclear to RE generation mix for decarbonization and grid stability

d)????? Can such 24x7x365 baseloads be made “flexible” to meet grid needs or even islanded

?Discussions are emerging as to how to supply these large loads. Power Engineering article “New study favors co-locating data centers with nuclear plants”, July 17, 2024, by Kevin Clark, notes benefits such as (a) no build-out of additional Tx capacity; and (b) data center owner accepts behind the meter backup costs as well as risk of not being served by the grid. This is a good option as it eliminates Tx costs and risk from the business case. However other questions arise such as:

(a)??? will resultant reduced net output of nuclear plant pose a supply risk to the grid

(b)??? will AIDC tariff be rate-regulated or bilateral at generator’s marginal cost (a subsidy)

(c)???? is there a stranded asset risk (due to relocation/shutdown)

(d)??? what will be optimum risk sharing (connection charges, owner’s local generation, grid supply, tariff)

?In my view, notwithstanding the above challenges, supplying large AIDC loads offers an opportunity for utilities. It provides new benefits as well, including a long-awaited opportunity to improve processes and standards. These include:

(a)??? leveraging steady “higher yield” revenues to potentially lower/offset other rate payer segments

(b)??? exploring if cloud-connected AI/ML technology can support load flexibility

(c)???? meeting short-term/seasonal AIDC supply shortfall with VPP tools and better load curtailment incentives

(d)??? revamping planning, engineering and construction processes to meet these short timelines

(e)??? leveraging new technologies for a faster, cheaper, and better build

?After almost 50 years there is now an opportunity for utilities and generators to design and build new GW-scale power infrastructure. It will partially replace and contribute to the mainstay of our current (old) power infrastructure for the next 50 years. This power industry rose to the challenge of building heavy power infrastructure in the 1960s and 1970s and again in early 2000s towards Smart Grid, RE and digitalization. It can do it again. The upcoming AI may indeed be its long-awaited impetus.