When people think about data centers and the resources that afford their development, they usually think in terms of four material resources: energy, land, water, and network infrastructure. In this research, we introduce an often overlooked but increasingly deterministic constraint: public sentiment.
We characterize public support as a true resource that affords - or constrains - data center construction in the same way that energy or water availability can. In study 1, the goal is to quantify the cost of public sentiment.
The goals are to address the following research questions: To what extent can public support be secured in the same way that a tech company would secure a water rights agreement? In other words, is public support for sale? And if so, how much does it cost?
Participants were recruited and paid through Verasight. Verasight gauranteed a nationally representative sample. Participants who were interested in taking the study were directed to our Qualtrics link, where they chose to consent or not before proceeding to the survey.
Sentiment
Overall, would you favor or oppose the construction of a new data center in your local area to support artificial intelligence, or AI, technology? [-2 = Strongly oppose, -1 = Somewhat oppose, 0 = No opinion, 1 = Somewhat favor, 2 = Strongly favor]
Data Center Density
How many data centers are currently within 5 miles of your home? [0 = None, 1 = One to two, 2 = More than two, 3 = I don’t know]
Free Response
You indicated that you ${q://QID70/ChoiceGroup/SelectedChoices} a tech company building a data center in your community. Write down some thoughts you had as you arrived at your judgment.
Reasons for support
You indicated that you ${q://QID70/ChoiceGroup/SelectedChoices} the construction of a data center in your community. To what extent did each of the following factors influence your judgment? [1 = Not at all influential; 5 = Extremely influential]
Quality of life for local residents
Local economic effects
The development of AI, or factors related to AI
Environmental effects
Effects on local resources
Price on Support
Displayed to participants who reported opposing or having no opinion.
Imagine a tech company is interested in building a data center in your community. They are willing to negotiate terms with the local residents. They are in talks with your local government to agree on a deal that would give all local residents a one-time tax rebate.
You indicated that you ${q://QID70/ChoiceGroup/SelectedChoices} ${e://Field/about}the construction of a data center in your local area to support AI. Under what conditions - in terms of a tax rebate - would you be willing to say that you support the construction of this data center in your community?
Indicate the dollar amount that the tech company should offer to local residents to gain your support.
Demographics
Demographic items were collected by Verasight and linked to the participant via verasight id:
gender
age
education
income
US census region
metropolitan area or not
race
hispanic
Political ID
political ID lean
voted in presidential election in 2024?
vote choice in 2024
show code
dcOpinion_study1 =merge(read.csv('study 1/dataCenterOpinion_01.csv'), readRDS('study 1/verasightDemogs.rds'), by='vsid') %>%filter(dataCenterSentiment !="") %>%mutate(dataCenterSentiment =factor(dataCenterSentiment, labels=c('Strongly oppose', 'Somewhat oppose', 'Have no opinion', 'Somewhat favor', 'Strongly favor')),dataCenterDensity =factor(dataCenterDensity, labels=c('None','1-2','More than 2', "I'm not sure"))) %>%mutate(dataCenterDensity_collapsed =factor(ifelse(dataCenterDensity =="I'm not sure", 1, dataCenterDensity), labels=c('None','1-2', 'More than 2')),priceForSupport_withoutStrongOpposerZeroes =ifelse(dataCenterSentiment =="Strongly oppose"& priceForSupport <1, NA, priceForSupport)) %>%mutate(priceForSupport_withSupporters =ifelse(dataCenterSentiment =="Somewhat favor"| dataCenterSentiment =="Strongly favor", 0, priceForSupport),priceForSupport_withSupporters =ifelse(dataCenterSentiment =="Strongly oppose"& priceForSupport <1, NA, priceForSupport_withSupporters),priceForSupport_binned =factor(ifelse(priceForSupport_withoutStrongOpposerZeroes <1, "$0", ifelse(priceForSupport_withoutStrongOpposerZeroes >0& priceForSupport_withoutStrongOpposerZeroes <1001, "$1K or less",ifelse(priceForSupport_withoutStrongOpposerZeroes >1000& priceForSupport_withoutStrongOpposerZeroes <10001, "More than $1K",ifelse(priceForSupport_withoutStrongOpposerZeroes >10000& priceForSupport_withoutStrongOpposerZeroes <50001, "More than $10K",ifelse(priceForSupport_withoutStrongOpposerZeroes >50000& priceForSupport_withoutStrongOpposerZeroes <100001, "More than $50K", ifelse(priceForSupport_withoutStrongOpposerZeroes >100000& priceForSupport_withoutStrongOpposerZeroes <1000001, 'More than $100K',ifelse(priceForSupport_withoutStrongOpposerZeroes >1000000& priceForSupport_withoutStrongOpposerZeroes <1000000001, 'More than $1M', "More than $1B"))))))), levels=c(NA, '$0','$1K or less','More than $1K', 'More than $10K','More than $50K','More than $100K','More than $1M','More than $1B')),income_buckets =ifelse(as.numeric(income) <4, "lower income",ifelse(as.numeric(income) ==4|as.numeric(income) ==5, "middle income",ifelse(as.numeric(income) >5, "upper income", NA))),reasonsTally =rowSums(pick("support_qualityOfLife", "support_localEconomy", "support_AIdev", "support_environment", "support_naturalResources") >1, na.rm =TRUE)) %>%mutate(reasonsTally_fct =factor(ifelse(reasonsTally <2, "1 or less", reasonsTally)))
We collected a total of 3234 responses. Note that responding to the “price for support” option was encouraged but not required, so some were left NA.
Participants who did not support a local data center build were asked to register their required price to gain their support. In the following analyses, participants who supported a local data center build are logged as “$0” for this variable.
In addition, there were a decent amount of participants who reported strong opposition to a local data center build, but registered “$0” as the price for their support. I interpreted these to mean “there is no amount of money that would buy my support.” Accordingly, I set those data points to NA
Raw results
Raw results can’t be plotted bc the distribution is too big:
mean = 3.4662045^{304}
min = 0
max = 10^{308}
median = 5000
Somebody registered “$-1” and they strongly opposed a local data center build. Probably should be interpreted as “nothing will buy my support”? Set to NA
Cutoff: billion
174 participants registered price tags higher than $1 billion. Excluding them yields 2711 data points.
show code
nice_table(dcOpinion_study1 %>%filter(priceForSupport_withSupporters<1000000001) %>%summarise(mean =mean(priceForSupport_withSupporters/1000), sd =sd(priceForSupport_withSupporters)/1000, median=median(priceForSupport_withSupporters)/1000, min =min(priceForSupport_withSupporters), max =max(priceForSupport_withSupporters)/1000), title ="Payment required for support, in thousands. (minimum reported in raw)")ggplot(subset(dcOpinion_study1,priceForSupport_withSupporters<1000000001&!is.na(priceForSupport_withSupporters)), aes(x=priceForSupport_withSupporters/1000))+geom_density(fill="#00bfc4", alpha=0.4)+geom_vline(aes(xintercept=mean(priceForSupport_withSupporters)/1000),color="#00bfc4", linetype="dashed", linewidth=0.5)+labs(x="Payment required for support, in thousands.", title="Cutoff: $1 billion")+theme_classic()
Payment required for support, in thousands. (minimum reported in raw)
mean
sd
median
min
max
23,966.79
144,119.32
5.00
0.00
1,000,000.00
Cutoff: million
401 participants registered price tags higher than $1 million. Excluding them yields 2484 data points.
show code
nice_table(dcOpinion_study1 %>%filter(priceForSupport_withSupporters<1000001) %>%summarise(mean =mean(priceForSupport_withSupporters/1000), sd =sd(priceForSupport_withSupporters)/1000, median=median(priceForSupport_withSupporters)/1000, min =min(priceForSupport_withSupporters), max =max(priceForSupport_withSupporters)/1000), title ="Payment required for support, in thousands. (minimum reported in raw)")ggplot(subset(dcOpinion_study1,priceForSupport_withSupporters<1000001&!is.na(priceForSupport_withSupporters)), aes(x=priceForSupport_withSupporters/1000))+geom_density(fill="#00bfc4", alpha=0.4)+geom_vline(aes(xintercept=mean(priceForSupport_withSupporters)/1000),color="#00bfc4", linetype="dashed", linewidth=0.5)+labs(x="Payment required for support, in thousands.", title="Cutoff: $1 million")+theme_classic()
Payment required for support, in thousands. (minimum reported in raw)
mean
sd
median
min
max
109.96
281.67
3.00
0.00
1,000.00
further visualizations
Among non-supporters (opposers and no opinions):
show code
nice_table(dcOpinion_study1 %>%filter(dataCenterSentiment !="Somewhat favor"& dataCenterSentiment !="Strongly favor"&!is.na(priceForSupport_binned)) %>%group_by(priceForSupport_binned) %>%tally() %>%mutate(percent = n/sum(n)*100))horizontalDotPlot =ggplot(dcOpinion_study1, aes(x =1, y=priceForSupport_withoutStrongOpposerZeroes))+geom_count(position=position_jitter(), alpha=0.65, color='#E57200')+scale_y_continuous(breaks =c(10000, 50000, 100000, 250000, 500000, 1000000), limits=c(0,1000001))+coord_flip()+theme_classic(base_size=13)+theme(axis.title.y=element_blank(),axis.text.y=element_blank(),axis.ticks.y=element_blank(),axis.line.y =element_blank())#ggsave(horizontalDotPlot, filename='study 1/horizontalDotPlot.svg', dpi = 300, width=8, height=2)horizontalDotPlotpercentBinsTable =data.frame("price"=c('$1K or less', '$10K or less', '$50K or less', '$100K or less', '$1M or less'),"percentage"=c(22.72, 50.63, 62.14, 69.28, 82.41)) %>%mutate(price =factor(price, levels=c('$1M or less','$100K or less', '$50K or less', '$10K or less', '$1K or less')))percentBins_plot =ggplot(percentBinsTable, aes(x = price, y=percentage, fill = price))+geom_bar(stat ="identity", position=position_dodge(), color='white') +scale_y_continuous(breaks=c(0,50,100), limits=c(0,100))+theme_classic(base_size=15)+labs(fill="", x='') +theme(legend.position='none')+scale_fill_manual(values=c('#E57200','#ec9049','#f1ac79','#f7c7a7','#F9DCBF'))+coord_flip()# ggsave(percentBins_plot, filename='study 1/percentBinsPlot.png', dpi = 300, width = 8, height = 6)percentBins_plot
Claude simulated prior approval/opposition to a data center being built in the local area.
show code
simulatedData %>%group_by(prior_attitude_numeric) %>%tally() %>%mutate(percent = n/sum(n)*100) %>%ggplot(aes(x = prior_attitude_numeric, y=percent, fill =as.factor(prior_attitude_numeric)))+geom_bar(stat ="identity", position ="dodge") +theme_minimal()+labs(fill="")
The simulated data very much missed the mark here, likely because public sentiment about data centers is so quickly changing.
And finally, the simulated requested tax rebate amounts:
show code
gghistogram(subset(simulatedData, rebate_condition=='one-time'), x ='wta_amount')+labs(title='One Time Payment')gghistogram(subset(simulatedData, rebate_condition=='yearly'), x ='wta_amount')+labs(title='Yearly Payments')
Note that Claude implemented a randomly assigned condition to either report a yearly or a one-time payment requirement.
Claude noted that the following likely issues:
Respondents will report $0 or refuse to answer if there is no amount of money that is sufficient
Respondents will anchor to the $1200 stimulus payment checks
