Mitigating Curtailment and Carbon Emissions through Load Migration between Data Centers

Abstract

•Load migration can reduce renewable curtailment and GHG emissions•Existing data centers in the CAISO region can reduce up to 239 KtCO2e per year•Net abatement cost can largely stay negative Increasing the capacity of renewable electricity is crucial in achieving the urgently needed low-carbon transition. As the capacity of renewable electricity grows, however, so does the renewable electricity that is generated but cannot be used when supply exceeds demand, which is referred to as curtailment. Curtailment is a growing concern that deteriorates the economic viability of renewables. This paper evaluates the idea of migrating workloads between data centers to reduce curtailment. Our analysis shows that using the underutilized capacity of data centers during the hours of excessive electricity supply can reduce curtailment and greenhouse gas emissions at negative costs. Applying this idea on a global scale has the potential to address the growing challenges of renewable curtailment and climate change in a cost-effective manner. As the share of variable renewable energy (VRE) grows in the electric grid, so does the risk of curtailment. While energy storage and hydrogen production have been proposed as solutions to the curtailment problem, they often pose technological and economic challenges. Here, we analyze the potential of data center load migration for mitigating curtailment and greenhouse gas (GHG) emissions. Using historical hourly electricity generation, curtailment, and typical data center server utilization data, we simulate the effect of migrating data center workloads from the fossil-fuel-heavy PJM to the renewable-heavy CAISO. The results show that load migration within the existing data center capacity during the curtailment hours in CAISO has the potential to reduce 113–239 KtCO2e per year of GHG emissions and absorb up to 62% of the total curtailment with negative abatement costs in 2019. Our study demonstrates the overlooked role that data centers can play for VRE integration and GHG emissions mitigation. As the share of variable renewable energy (VRE) grows in the electric grid, so does the risk of curtailment. While energy storage and hydrogen production have been proposed as solutions to the curtailment problem, they often pose technological and economic challenges. Here, we analyze the potential of data center load migration for mitigating curtailment and greenhouse gas (GHG) emissions. Using historical hourly electricity generation, curtailment, and typical data center server utilization data, we simulate the effect of migrating data center workloads from the fossil-fuel-heavy PJM to the renewable-heavy CAISO. The results show that load migration within the existing data center capacity during the curtailment hours in CAISO has the potential to reduce 113–239 KtCO2e per year of GHG emissions and absorb up to 62% of the total curtailment with negative abatement costs in 2019. Our study demonstrates the overlooked role that data centers can play for VRE integration and GHG emissions mitigation. Driven by aggressive public policy and compelling economics, global capacity of variable renewable energy (VRE), such as solar photovoltaics (PV) and wind electricity, is growing rapidly. The European Union, for example, has a target of achieving at least 32% share of renewable energy by 2030,1Union European 2030 climate & energy framework. EU climate action.https://ec.europa.eu/clima/policies/strategies/2030_enDate: 2016Google Scholar and California aims at a 60% renewable portfolio standard by 2030.2State of California California renewables portfolio standard (RPS) program.https://www.cpuc.ca.gov/rps/Date: 2018Google Scholar As the penetration of VRE in the grid grows, so do the concerns of large-scale curtailment.3Jenkins J.D. Luke M. Thernstrom S. Getting to zero carbon emissions in the electric power sector.Joule. 2018; 2: 2498-2510Abstract Full Text Full Text PDF Scopus (47) Google Scholar, 4Frew B. Cole W. Denholm P. Frazier A.W. Vincent N. Margolis R. Sunny with a chance of curtailment: operating the U.S. grid with very high levels of solar photovoltaics.iScience. 2019; 21: 436-447Abstract Full Text Full Text PDF PubMed Scopus (13) Google Scholar, 5Wiser R.H. Mills A. Seel J. Levin T. Botterud A. Impacts of variable renewable energy on bulk power system assets, pricing, and costs.https://emp.lbl.gov/sites/default/files/lbnl_anl_impacts_of_variable_renewable_energy_final.pdfDate: 2017Google Scholar Curtailment is the reduction of output of a VRE resource below what it could have otherwise produced. It has been repeatedly reported in different world regions across Europe, America, and Asia, significantly decreasing the market value of VRE.5Wiser R.H. Mills A. Seel J. Levin T. Botterud A. Impacts of variable renewable energy on bulk power system assets, pricing, and costs.https://emp.lbl.gov/sites/default/files/lbnl_anl_impacts_of_variable_renewable_energy_final.pdfDate: 2017Google Scholar,6Bird L. Lew D. Milligan M. Carlini E.M. Estanqueiro A. Flynn D. Gomez-Lazaro E. Holttinen H. Menemenlis N. Orths A. et al.Wind and solar energy curtailment: a review of international experience.Renew. Sustain. Energy Rev. 2016; 65: 577-586Crossref Scopus (196) Google Scholar Near-term reasons for VRE curtailment include minimum generation requirements for non-renewable energy sources and transmission constraints, but long-term, fundamental causes drive increasing pressure for curtailment.5Wiser R.H. Mills A. Seel J. Levin T. Botterud A. Impacts of variable renewable energy on bulk power system assets, pricing, and costs.https://emp.lbl.gov/sites/default/files/lbnl_anl_impacts_of_variable_renewable_energy_final.pdfDate: 2017Google Scholar,7Denholm P. O’Connell M. Brinkman G. Jorgenson J. Overgeneration from solar energy in California. A field guide to the duck chart. National Renewable Energy Laboratory, 2015https://www.nrel.gov/docs/fy16osti/65023.pdfCrossref Google Scholar Large-scale energy storage and electricity-transmission network expansion can mitigate VRE curtailment, but they are costly. With a system cost between $380 to $895 per kWh,8Fu R. Remo T. Magolis R. 2018 U.S. utility-scale photovoltaics-plus-energy storage system costs benchmark. National Renewable Energy Laboratory, 2018https://www.nrel.gov/docs/fy19osti/71714.pdfGoogle Scholar the battery storage capacity deployed globally (12 GWh in 2018)9Wood Mackenzie power & renewableGlobal energy storage outlook 2019.https://www.woodmac.com/reports/power-markets-global-energy-storage-outlook-2019-295618Date: 2019Google Scholar is infinitesimal compared with the amount of global electricity consumption (about 23,000 TWh per year).10Enerdata Global energy statistic yearbook 2019.https://yearbook.enerdata.net/electricity/electricity-domestic-consumption-data.htmlDate: 2019Google Scholar Long-distance transmission of VRE-generated electricity is possible, but the construction of transmission infrastructure and the associated transmission losses are often cost prohibitive.11Bird S. Achuthan A. Ait Maatallah O. Hu W. Janoyan K. Kwasinski A. Matthews J. Mayhew D. Owen J. Marzocca P. Distributed (green) data centers: a new concept for energy, computing, and telecommunications.Energy Sustain. Dev. 2014; 19: 83-91Crossref Scopus (19) Google Scholar Pumped hydro can be another storage solution, but it requires certain geographical features and may raise ecological concerns.12Rehman S. Al-Hadhrami L.M. Alam M.M. Pumped hydro energy storage system: a technological review.Renew. Sustain. Energy Rev. 2015; 44: 586-598Crossref Scopus (406) Google Scholar Another approach to reduce curtailment is to use excess VRE electricity to produce more easily storable materials or products, such as hydrogen through water electrolysis, which can be shipped upon demand.13Gahleitner G. Hydrogen from renewable electricity: an international review of power-to-gas pilot plants for stationary applications.International Journal of Hydrogen Energy. 2013; 38: 2039-2061Crossref Scopus (691) Google Scholar,14Barton J.P. Infield D.G. Energy storage and its use with intermittent renewable energy.IEEE Trans. On Energy Conversion. 2004; 19: 441-448Crossref Scopus (985) Google Scholar However, the logistics and handling of these materials and associated costs can pose additional challenges.15Klerke A. Christensen C.H. Nørskov J.K. Vegge T. Ammonia for hydrogen storage: challenges and opportunities.J. Mater. Chem. 2008; 18: 2304-2310Crossref Scopus (562) Google Scholar A potential solution to this logistics and handling problem is to use over-generated electricity to produce something that can be transported at minimal cost and energy: information. Data centers can provide battery-like demand-side management service by powering data processing with excess VRE. The technical and economic potential of zero-carbon cloud (ZCC) data centers that run solely on stranded renewable power has been explored.16Yang F. Chien A.A. Large-scale and extreme-scale computing with stranded green power: opportunities and costs.IEEE Trans. Parallel Distrib. Syst. 2018; 29: 1103-1116Crossref Scopus (6) Google Scholar,17Chien A.A. Wolski R. Yang F. The zero-carbon cloud: high-value, dispatchable demand for renewable power generators.The Electricity Journal. 2015; 28: 110-118Crossref Scopus (10) Google Scholar Geographical load balancing has been widely studied to maximize renewable energy use by distributing the workloads among data centers in different locations.18Liu Z. Lin M. Wierman A. Low S.H. Andrew L.L.H. Geographical load balancing with renewables.SIGMETRICS Perform. Eval. Rev. 2011; 39: 62-66Crossref Scopus (111) Google Scholar,19Kong F. Liu X. A survey on green-energy-aware power management for datacenters.ACM Comput. Surv. 2015; 47: 1-38Crossref Scopus (94) Google Scholar Data centers are highly automated and monitored with little human intervention. Importantly, they have considerable flexible workloads which can be distributed geographically.17Chien A.A. Wolski R. Yang F. The zero-carbon cloud: high-value, dispatchable demand for renewable power generators.The Electricity Journal. 2015; 28: 110-118Crossref Scopus (10) Google Scholar,20Goiri Í. Haque M.E. Le K. Beauchea R. Nguyen T.D. Guitart J. Torres J. Bianchini R. Matching renewable energy supply and demand in green datacenters.Ad Hoc Netw. 2015; 25: 520-534Crossref Scopus (48) Google Scholar Provided the requisite data are available, those data centers with access to renewable energy can process the requests routed from other regions and return the results to users while meeting service level agreements (SLAs). Moreover, most data centers operate well below 100% capacity most of the time—overprovisioning for peaks leaves servers and network resources underutilized.21Jin X. Zhang F. Vasilakos A.V. Liu Z. Green Data Centers: A Survey, Perspectives, and Future Directions.ArXiv. 2016; (160800687 Cs)Google Scholar Furthermore, the peak loads for data processing often does not coincide with the peak time of VRE overgeneration, providing room for data centers to use their excess capacity to process additional workloads with excess VRE. Compared with building large-scale transmission infrastructure, building fiber-optic networks and transmitting data are much cheaper and take significantly less time.11Bird S. Achuthan A. Ait Maatallah O. Hu W. Janoyan K. Kwasinski A. Matthews J. Mayhew D. Owen J. Marzocca P. Distributed (green) data centers: a new concept for energy, computing, and telecommunications.Energy Sustain. Dev. 2014; 19: 83-91Crossref Scopus (19) Google Scholar Therefore, the transmission of data are more economically favorable than the transmission of electricity, i.e., “moving bits, not watts.”22Liu Z. Wierman A. Chen Y. Razon B. Chen N. Data center demand response: avoiding the coincident peak via workload shifting and local generation.Perform. Eval. 2013; 70: 770-791Crossref Scopus (120) Google Scholar Furthermore, the society’s need for data processing is growing rapidly. Global data centers used 205 TWh electricity in 2018 or 1% of global electricity consumption.23Masanet E. Shehabi A. Lei N. Smith S. Koomey J. Recalibrating global data center energy-use estimates.Science. 2020; 367: 984-986Crossref PubMed Scopus (86) Google Scholar In 2014, U.S. data centers consumed 70 TWh electricity, which was 1.8% of the total annual U.S. electricity consumption.24Shehabi A. Smith S. Sartor D. Brown R. Herrlin M. Koomey J. Masanet E. Horner N. Azevedo I. Lintner W. United States data center energy usage report.https://www.osti.gov/biblio/1372902Date: 2016Google Scholar It is estimated that the global datasphere will grow from 33 zettabytes (ZB) in 2018 to 175 ZB in 2025 at an annual growth rate of 27%, implying growing needs for data center infrastructures.25Reinsel D. Gantz J. Rydning J. The digitization of the world from edge to core.https://www.seagate.com/files/www-content/our-story/trends/files/idc-seagate-dataage-whitepaper.pdfDate: 2018Google Scholar The decarbonization of data centers is imperative and requires combined efforts including maximizing IT-device efficiency, adoption of low-carbon electricity, and improving infrastructure efficiency.26Masanet E. Shehabi A. Koomey J. Characteristics of low-carbon data centres.Nat. Clim. Change. 2013; 3: 627-630Crossref Scopus (25) Google Scholar Load migration between data centers can collectively improve IT efficiency and utilization of renewable energy. Nevertheless, the potential for load migration between data centers to utilize excess VRE generation and reduce greenhouse gas (GHG) emissions has not been quantified. In this study, we use two independent system operators (ISOs) in the U.S., California ISO (CAISO) and Pennsylvania-New Jersey-Maryland interconnection (PJM), as a case study to explore the potential of workload migration between data centers to mitigate curtailment and GHG emissions. PJM is the largest ISO in the U.S., which predominantly relies on thermal energy sources, such as coal and natural gas, with solar and wind accounting for only 3.2% in 2019.27PJMPJM at a glance.https://www.pjm.com/∼/media/about-pjm/newsroom/fact-sheets/pjm-at-a-glance.ashxDate: 2019Google Scholar,28PJMPJM Data Miner 2 Generation by fuel type.https://dataminer2.pjm.com/feed/gen_by_fuelGoogle Scholar The states covered by PJM host a large amount of data centers, with the most noteworthy area being North Virginia. In the second half of 2018, North Virginia absorbed over a third of the world’s new data center capacity, with an addition of 270 megawatts (MW) of data center power.29Sverdlik Y. The cloud data center construction boom, in two charts Data Center Knowledge, February 28, 2019.https://www.datacenterknowledge.com/cloud/cloud-data-center-construction-boom-two-chartsDate: 2019Google Scholar As the hub of technology and media companies, California also has many data centers, mostly located in the Bay Area and Southern California.30Ghatikar G. Ganti V. Matson N. Piette M.A. Demand response opportunities and enabling technologies for data centers: findings from field studies. Lawrence Berkeley National Laboratory, 2012https://www.osti.gov/servlets/purl/1174175Google Scholar Of all the data center colocation establishments in the U.S., 15% are located in California and about 16% are in the PJM region.31Data Center Map. Colocation USA - Data Centers.https://www.datacentermap.com/usa/Google Scholar,32Moses J. IBISWorld Industry Report 51821 - Data Processing & Hosting Services in the US. IBISWorld Services, 2020https://clients1.ibisworld.com/reports/us/industry/default.aspx?entid=1281Google Scholar Based on the historical hourly curtailment data of CAISO and a typical data center energy consumption profile, we evaluate the potential of the existing and additional data center capacities to absorb excess VRE and reduce GHG emissions by migrating data center workloads from the PJM region. In this analysis, we use counterfactual scenarios as an illustration of the potential rather than as a record of historical accounts. We collected and analyzed the historical curtailment data of CAISO during 2015–2019. The total annual curtailment of CAISO grew from 188 to 965 GWh from 2015 to 2019, at an average annual growth rate of 51%. Curtailment data at CAISO shows wide daily and seasonal variations, with an upward trend over time (Figure 1A). Solar PV curtailment accounted for 90% of the total cumulative curtailment during this period and wind accounted for 10%. The majority of curtailment occurred in the first and second quarters, which when combined accounted for 69% of the total curtailment in the period. Monthly curtailment peaked in April or May. This results from growing solar radiation strength and extended daytime length during spring, combined with mandatory runoffs from northwest hydro generation imports and cool weather. Both solar and wind curtailment occurred the least in the third quarter with July or August seeing the minimum, which can be explained by higher cooling demands in summer’s warmer weather. The surge of solar curtailment during 2015–2019 mirrors the fact that the share of solar power in total CAISO generation had increased from 6.7% to 13.0% in this period. In comparison, the share of wind power in the generation mix increased from 5.3% to 7.2%, representing a milder growth than solar. When disaggregated at hourly resolution (Figure 1B), curtailment took place rather randomly throughout the 24 h of a day in 2015 and 2016, but as solar PV capacity grew and nighttime wind curtailment decreased during 2017–2019, total curtailment became increasingly more conspicuous in the daytime. Instead of curtailing, the excess VRE generation in CAISO could be used to process data center workloads migrated from carbon-intensive grid regions. Many data centers operate at less than 50% average server utilization rate (UR),24Shehabi A. Smith S. Sartor D. Brown R. Herrlin M. Koomey J. Masanet E. Horner N. Azevedo I. Lintner W. United States data center energy usage report.https://www.osti.gov/biblio/1372902Date: 2016Google Scholar,33Cortez E. Bonde A. Muzio A. Russinovich M. Fontoura M. Bianchini R. Resource central: understanding and predicting workloads for improved resource management in large cloud platforms.in: Proceedings of the 26th Symposium on Operating Systems Principles - SOSP ’17. ACM Press, 2017: 153-167Crossref Scopus (143) Google Scholar,34Barroso L.A. Hölzle U. Ranganathan P. The Datacenter as a Computer: Designing Warehouse-Scale Machines. Morgan & Claypool, 2019Google Scholar and the time-zone difference between PJM and CAISO helps avoid peaking load at the same time, allowing the data centers served by CAISO to take on additional data-processing jobs migrated from PJM-served data centers during the off-peak hours. We use the historical curtailment data, which is referred to as “excess VRE” hereafter, to evaluate the potential of migrating workloads between data centers. We collected and treated the electricity generation by energy resource data for CAISO and PJM during 2016–2019.28PJMPJM Data Miner 2 Generation by fuel type.https://dataminer2.pjm.com/feed/gen_by_fuelGoogle Scholar,35CAISO Renewables and emissions reports.http://www.caiso.com/market/Pages/ReportsBulletins/RenewablesReporting.aspxDate: 2020Google Scholar The life-cycle GHG intensities of the two ISOs during 2016–2019 (Figure 2) were calculated on an hourly basis based on the historical generation data and U.S.-specific GHG emission factors, which include both combustion emissions and life-cycle emissions embodied in the inputs to power generation (Table S1). Imported electricity is not included in the calculation. The annual average GHG intensity of PJM decreased from 499 to 452 kg CO2e/MWh during 2016–2019. The monthly average intensity of PJM peaked in summer (July or August) and reached its lowest around April and October, with a range between 417–557 kg CO2e/MWh. For CAISO, the annual average GHG intensity changed from 262 to 231 kg CO2e/MWh during the same period. The monthly intensity of CAISO hit the lowest in April or May due to prominent solar and hydro-power production. During the hours when curtailment occurred in CAISO, only the intensity values of the curtailment (i.e., excess VRE) are shown (Figure 2, CAISO), which is assumed to be proportionally contributed by curtailed solar and wind power. In other words, the GHG emissions intensity of the excess generation is calculated as the average life-cycle GHG emissions intensity of solar and wind weighted by their shares in the total curtailment during that hour. While the average GHG intensity of CAISO excess generation during 2016–2019 was 41 kgCO2e/MWh, the average GHG intensity of PJM during CAISO’s excess generation time was 476 kgCO2e/MWh. The significant differences of the GHG intensities between the two grids during CAISO excess generation time present a great opportunity for GHG emissions mitigation by migrating the data center workloads geographically. We first estimated the existing data center capacity in the CAISO region. We collected available data on data center location and power consumption.36Cloud and ColocationCalifornia data center location map. Cloud Coloca, 2020https://cloudandcolocation.com/state/california/Google Scholar By examining the profiles of all the listed data centers in California, we calculated that the average annual total power consumption per colocation site is 9.92 MW, based on 26 data points that provided the information. We also identified that there were 288 data centers in the CAISO region by the end of 2019.37Cloudscene Cloudscene data centers market in United States of America.https://cloudscene.com/market/data-centers-in-united-states/allDate: 2020Google Scholar We use a typical data center energy profile with an IT peak power (or critical power) of 10 MW as a standardized unit38Rahmani R. Moser I. Seyedmahmoudian M. A complete model for modular simulation of data centre power load.arXiv. 2018; http://arxiv.org/abs/1804.00703Google Scholar in this study to estimate the excess VRE absorption capacity, GHG emissions reduction potential, and abatement costs. We then simulate the dynamic range (DR) and power usage effectiveness (PUE) of the data centers served by CAISO. DR is the ratio of a server’s idling power to its maximum power,24Shehabi A. Smith S. Sartor D. Brown R. Herrlin M. Koomey J. Masanet E. Horner N. Azevedo I. Lintner W. United States data center energy usage report.https://www.osti.gov/biblio/1372902Date: 2016Google Scholar based on which we calculate the energy consumption of servers given the rated power and UR. PUE is defined as the ratio of the total energy consumption of a data center to the energy used by its IT equipment, calculated, measured, or assessed across the same period.39The International Organization for Standardization and The International Electrotechnical CommissionISO/IEC 30134-2:2016(en) Information technology — Data centres — Key performance indicators — Part 2: Power usage effectiveness (PUE). International Organization for Standardization, 2016Google Scholar PUE values vary depending on data center type and geographical location. Here, we model the average PUE of colocation data centers in California. Detailed assumptions of the two parameters can be found in Experimental Procedures and Table S3. We also developed a linear model between the hourly server UR and the energy use of non-server components,38Rahmani R. Moser I. Seyedmahmoudian M. A complete model for modular simulation of data centre power load.arXiv. 2018; http://arxiv.org/abs/1804.00703Google Scholar through which we can calculate the non-server energy consumption given a certain PUE value in a year. We compare two scenarios for evaluating the excess VRE absorption potential of migrating workload between data centers: baseline scenario and migration scenario. In baseline scenario, workloads are processed by typical data centers served by PJM without any migration. In migration scenario, workloads are first migrated to and processed by the existing typical data centers served by CAISO. We assume that the migration occurs between data centers of similar scale with typical energy-use characteristics in our model. Once the existing capacity is exhausted, we assume that additional data center capacity, which runs solely on the remaining excess VRE, can be built. We assume that the data centers all have advanced algorithms and automation mechanisms in place to enable the load migration. Figure 3 illustrates the excess VRE absorption potential of a typical data center in a week. The remaining capacity of an existing data center in an underutilized hour is calculated by subtracting the existing load of the data center in that hour from its maximum allowed load. Load migration is only enabled during the hours when the servers in data centers served by CAISO are underutilized. We test different scenarios by varying the assumption of the maximum allowed server UR between 65% and 90% during underutilized time, representing an improved management and a maximized utilization scenario, respectively. Average UR of large-scale cloud providers is estimated as 65%.40Barr J. Cloud computing, server utilization, & the environment. AWS news blog. 5 June, 2015.https://aws.amazon.com/blogs/aws/cloud-computing-server-utilization-the-environment/Date: 2015Google Scholar Once the remaining capacities of all existing data centers are exhausted, we calculate the respective additional data center capacity needed to absorb different portions of the total excess VRE. During excess generation hours, the servers in the additional data centers would be activated to process the workloads migrated from the PJM region and operate at the maximum allowed UR assumed. The servers are assumed to be shut down at times when there is no available excess VRE. We calculate the achieved total GHG emissions reduction by summing up the products of the hourly GHG intensity difference between the two scenarios and the amount of excess VRE absorbed for each year between 2016 and 2019. We then estimate the total abatement cost of the plan by comparing the cost difference between the two scenarios. For the excess VRE that falls within the remaining capacity of existing data centers, workload migration causes only a change in electricity bills between the two scenarios. When additional data centers are built to absorb extra excess VRE, changes in electricity cost, amortized facility cost, and additional cost are all captured. We use the cost estimates developed specifically for ZCC data centers that run on stranded renewable power16Yang F. Chien A.A. Large-scale and extreme-scale computing with stranded green power: opportunities and costs.IEEE Trans. Parallel Distrib. Syst. 2018; 29: 1103-1116Crossref Scopus (6) Google Scholar for the additional data center capacity in the migration scenario. These intermittent data centers have lower facility costs because they use containers and can be located near renewable generation sites with lower power distribution costs.16Yang F. Chien A.A. Large-scale and extreme-scale computing with stranded green power: opportunities and costs.IEEE Trans. Parallel Distrib. Syst. 2018; 29: 1103-1116Crossref Scopus (6) Google Scholar The electricity cost is also significantly lower for ZCC data centers than the traditional ones as the otherwise-curtailed VRE electricity is assumed to have zero cost. Additional cost for installing data and energy-storage devices will incur due to the intermittent characteristic of the power supply for ZCC data centers. The cost of applications, including software licenses and system and database administration, are not considered as they vary greatly and do not constitute part of the infrastructure-related capital or operational cost. The total abatement cost sums up the difference of facility, electricity, and additional costs between the two scenarios on an annual basis. The net abatement cost (in $/metric ton CO2e) is then calculated by dividing the total abatement cost by the total net GHG emissions reduction achieved. Figure 4 summarizes the results of total GHG emissions reduction and net GHG abatement cost using 2019 data. The existing data center capacity alone (i.e., when additional data center capacity is zero) can absorb 29%–62% of the total excess VRE in CAISO in 2019, assuming that the maximum server UR ranges between 65% and 90%. As we increase the maximum server UR and additional data center capacity, the excess VRE absorption level grows. At a given absorption level, a higher maximum server UR means a reduced need for additional data center capacity. The resulting GHG emissions reduction is the net reduction after accounting for the embodied GHG emissions of the additional data centers, which are incurred due to the manufacturing of IT equipment and infrastructure materials. The embodied emissions are proportional to the total number of new data centers built and would offset a fair amount of the operational GHG emissions reduction achieved by data center workload migration. A total net GHG emissions of 113–239 KtCO2e could have been reduced in 2019 given the maximum server UR range assumed, and additional data center capacity can bring the total reduction further up to 247 KtCO2e (Figure 4A). The net GHG emissions reduction peaks at an absorption level between 68% and 75% when the maximum server UR exceeds 85%. Absorbing 80% of the total excess VRE or above does not bring more GHG emissions reduction benefits because the embodied emissions from additional data centers would outweigh the reduction from operational phase. Figure 4B shows the estimated net abatement cost of the migration scenario. A negative abatement cost means that data centers can g