logo
How Business Insider investigated the true cost of data centers

How Business Insider investigated the true cost of data centers

Tech companies are spending hundreds of billions of dollars building data centers, racing to power an AI revolution they say will bring broad benefits and enormous profits.
The actual price, though, has never been fully measured. So a team of Business Insider reporters and editors set out to find the true cost of the US data center boom — in water, power, pollution, and tax incentives.
We pulled obscure state records, scoured corporate disclosures, and sought analysis and guidance from government researchers in Virginia and elsewhere. We spoke to engineering specialists who study data centers, including Shaolei Ren at the University of California, Riverside, and Vijay Gadepally of the Massachusetts Institute of Technology. We also spoke to researchers who study incentive packages, such as Kasia Tarczynska of Good Jobs First, a research group that advocates for transparency and equality in public subsidies. And we sought input from companies and industry groups.
The result is the most complete accounting yet of data centers' impact on our environment, economy, and lives.
Here's how we did it.
We started by identifying how many data centers there are, and who owns them
Data centers can never be without power. Developers build backup generators to provide emergency electricity. Since these generators emit air pollution, they are regulated by state environmental agencies and need permits.
We requested these air permits and applications from 50 states and Washington, DC, asking agencies to search permits using federal industry codes associated with data centers. To confirm state searches were complete, Business Insider cross-referenced permits against company disclosures, news articles, and data from datacentermap.com, Baxtel, and other trade websites.
Some of the permits we obtained were actually for other types of facilities. If public or corporate records confirmed the facility was not a data center, we removed it from our analysis.
We compiled permits for 1,240 existing or planned data centers across the country, the most definitive tally to date. Tracking the year each permit was first issued reveals a construction boom. Until 2010, companies had filed permits for 311 data centers. By 2024, that number had nearly quadrupled.
Often the permit holders are subsidiaries of Big Tech companies, with unfamiliar names like Nova Mango Farms. Business Insider determined the owners using emails, addresses, and other information on permit applications, secretary-of-state disclosures, county databases, and SEC filings.
Then we mapped those centers' locations and estimated the resources they consume
Business Insider mapped location information from the permits onto the Aqueduct Water Risk Atlas built by the World Resources Institute, a Washington, DC, group that advocates sustainability. The atlas measures a location's relative water stress — the ratio between demand and available, renewable supplies. Business Insider found that 40% of data centers were sited in areas of high or extremely high water stress.
Data centers can use staggering amounts of water. Business Insider requested metered water use records and water supply agreements from dozens of public utility agencies for every data center in areas with extremely high water stress.
Data center water use is a closely guarded secret. Many agencies denied our requests for facility-specific water use or released the records only in aggregate for all data centers in a water utility's service area. In Denver and Colorado Springs, Colorado, utility agencies sued Business Insider to prevent the release of metered water use records. Business Insider prevailed in both lawsuits.
Business Insider found that some of the largest data center facilities were permitted to use more water a day than you might expect nearly 49,000 Americans to use.
Companies also disclose little information on how much power their US data centers consume. Data center power use varies based on factors including the types and frequency of use of servers, cooling systems, and other infrastructure. To account for this, Business Insider calculated a range of estimated annual data center electricity use.
To calculate this range, Business Insider used air permit data to identify the total power-generation capacity of every backup generator at each facility. Companies rarely install generator capacity that is more than twice what a data center consumes — and often they install significantly less than that. So Business Insider, in consultation with experts, conservatively estimated that a data center's maximum electricity use is between half and 80% of its generators' capacity.
A 2024 federal report estimates that across the industry, data centers operate at just 50% of that maximum electricity use level, to avoid straining their equipment. So Business Insider multiplied the low and high ends of its initial range by 50%.
Business Insider's data includes permits for data centers that are not yet built and fully online, and therefore estimates expected electricity use in the near future. If all permitted data centers go online, we estimate, they will use between 149.6 terawatt-hours and 239.3 terawatt-hours a year. Business Insider's low-end estimate is roughly equivalent to the state of Ohio's electricity needs in 2023, and on the high end, is nearly as much power as the entire state of Florida used that same year.
Our estimates are conservative. The low end of our estimate is below the federal report's estimate for data center power use in 2023. The federal report estimates data center power use could reach the upper end of our estimate by 2026.
Business Insider used the range average to identify the 322 largest facilities in our dataset that, according to our estimate, consume 40 megawatt-hours of electricity or more an hour.
An Amazon spokesperson said the methodology "oversimplifies complex data center operations and is based on assumptions that do not account for important differences in how companies build and operate data centers."
Google and Meta did not respond to Business Insider's queries about data center power use estimates, and QTS declined to comment. A Microsoft spokesperson acknowledged that its data centers "do not always run at 100% of their installed capacity."
Nineteen permits across 10 states — including 12 held by Google — have their generator capacity information fully or partially redacted under public disclosure exemptions for trade secrets. Business Insider was unable to estimate the electricity use of these facilities, so our estimate likely undercounts the true total.
Next, we calculated data centers' likely public health and environmental impact
The US uses highly polluting fossil-fuel-fired sources for at least 60% of its electricity. Data centers draw power from a wide regional grid, so Business Insider needed to estimate the public health impact attributable to data centers' use of electricity generated from far-away power plants.
The US electricity grid is complex and does not map closely to state boundaries. To model the complexity of the US electricity grid, Business Insider used an Environmental Protection Agency tool called AVERT that splits the contiguous 48 states into 14 independent electricity regions that approximate grid regions.
Using our estimate for data center power consumption, along with US Energy Information Administration data and the AVERT tool, Business Insider calculated the likely share of total power use attributable to data centers in each of these 14 regions. Then, using another EPA tool with the acronym COBRA, Business Insider estimated the public health cost engendered by that share of total electricity-generation-related pollution in each region. Even a careful calculation like this is inherently imperfect, but in the absence of industry transparency about such emissions it provides a rare and well-founded estimate of the health costs of the data center boom.
Business Insider estimated that annual public health costs from electricity generation for data centers could reach between $5.7 billion and $9.2 billion.
Power plants are not the only source of data center air pollution.
The diesel backup generators installed at data centers emit volatile organic compounds, nitrous oxides, sulfur dioxides, and particulate matter — all regulated air pollutants the EPA labels as hazardous to human health. Air permits issued to these facilities set total annual pollutant emission limits that determine how frequently their generators can be used.
Data centers generally run their generators infrequently so emit only a fraction of the pollutants their permits allow. A 2024 report to the state legislature in Virginia — which is home to nearly a third of US data centers with permits — calculated that data center backup generators on average emit 7% of their permitted pollution limits each year. Business Insider used that percentage to estimate each county's air pollutant emissions from data center backup generators.
Business Insider then input these county-by-county totals into COBRA, which models how air pollution emitted in a county affects public health nationwide. The tool assigns a dollar figure to what it would cost to avoid negative health outcomes like premature deaths, asthma attacks, heart attacks, and missed school or work days.
The COBRA tool estimated that air pollutants emitted by data centers across the country could trigger nearly 20,000 asthma symptom cases a year and create a $385 million annual burden on public health.
Generator pollution is more harmful to communities already overburdened by environmental pollutants. A community that already has high levels of air pollution will experience greater negative health impacts from additional pollutants, for example, than a community that isn't already saturated with harmful pollutants.
To better understand their impact on these communities, Business Insider mapped all 1,240 data center locations onto a third EPA tool called EJ Screen that combines census and pollutant data within a mile radius of a specified location. We found that more than 230 data center locations were in communities highly overburdened by environmental pollutants.
Finally, we reported on data centers' impact on the economy
Business Insider requested and obtained the most detailed data-center-related economic development deals and annual tax incentive disclosures for facilities in and around Columbus, Ohio. Some of the oldest data center deals in the state are in Franklin and Licking counties, and companies are required to disclose how much they've saved from these tax exemption agreements each year, which made the state a useful case study for judging data centers' actual economic impact. Business Insider tallied these numbers, then used company-reported property investment totals to calculate total state and local sales and use tax savings (7.5% for Franklin County and 5.75% for Licking County).
We did not include personal property tax in our calculations, as Ohio eliminated that tax. Because of fluctuations in the number of jobs employed at data centers, Business Insider also did not include state or local job creation tax credits in the analysis.
We nevertheless found that these tax incentive packages given to tech companies were highly lucrative, amounting to as much as $2 million in tax savings per full time job at a data center.
Orange background

Try Our AI Features

Explore what Daily8 AI can do for you:

Comments

No comments yet...

Related Articles

Cedars-Sinai's AI tool delivered 24/7 care to 42,000 patients. Now, doctors can focus on treatment, not paperwork.
Cedars-Sinai's AI tool delivered 24/7 care to 42,000 patients. Now, doctors can focus on treatment, not paperwork.

Business Insider

time13 minutes ago

  • Business Insider

Cedars-Sinai's AI tool delivered 24/7 care to 42,000 patients. Now, doctors can focus on treatment, not paperwork.

Cedars-Sinai launched CS Connect, an AI-powered platform for 24/7 healthcare support. The platform aims to reduce wait times and administrative tasks, improving care delivery. This article is part of " CXO AI Playbook" — straight talk from business leaders on how they're testing and using AI. Cedars-Sinai is a nonprofit healthcare organization based in Los Angeles. It's comprised of hospitals, clinics, and research facilities. Founded in 1902, Cedars-Sinai has more than 40 locations, employs over 4,500 physicians and nurses, and serves more than 1 million patients every year. Situation analysis Caroline Goldzweig, chief medical officer of Cedars-Sinai Medical Network, told Business Insider that the organization wanted to expand patients' access to primary care in a more efficient and convenient way. For instance, sometimes there are lengthy wait times for in-person doctor's appointments. At the same time, Goldzweig said the organization was looking to reduce the administrative burden on physicians, including patient intake and data entry, so that they could focus on providing care. To address these challenges, the organization launched Cedars-Sinai Connect in 2023. CS Connect is an artificial intelligence -powered virtual platform where patients can access healthcare support 24/7. Goldzweig said the AI technology allows healthcare providers to better support their patients while improving the speed of care delivery. "The other exciting thing is the ability to offer patients care in ways that meet their needs," she said. Key staff and partners Goldzweig told Business Insider that CS Connect was built using K Health, a digital healthcare company providing AI-powered primary and urgent care services. Cedars-Sinai's information technology teams worked with K Health to sync patients' electronic health records into the platform, she said. Cedars-Sinai's enterprise data intelligence and digital strategy teams were also involved in creating CS Connect. They continue to help manage it. AI in action The goal of the AI tool is to enable providers to spend more time counseling patients and making treatment decisions, rather than doing patient intake. To access the AI features, patients log into CS Connect via a mobile app or its website. Goldzweig said a chat feature then starts asking the user questions about their symptoms. The AI algorithm compares the patient's responses to their existing medical records and the records of other patients in the system who had similar symptoms. It then asks more detailed questions about a patient's specific health problems. The chatbot, for example, may prompt the patient to submit photos of their sore throat or rash. Goldzweig said the chat function is similar to how a physician might identify a patient's ailment. The AI then summarizes the patient's information — a task that usually demands manual effort from physicians — and recommends a treatment. The data collected by CS Connect is "part of the efficiency process," Goldzweig said. "It allows the physicians to review it, to understand what's going on, and to validate things with the patient — ask a couple of clarifying questions and maybe get a little bit more history." Physicians can choose to disagree with the AI's recommendations. Goldzweig said either way, they're required to sign off on each patient's treatment plan. Did it work, and how did leaders know? Goldzweig told BI that about 42,000 individual patients have used CS Connect. In April 2025, Cedars-Sinai published a study in the journal Annals of Internal Medicine comparing AI treatment recommendations in CS Connect with final recommendations from doctors. The study reviewed 461 physician-managed visits with AI recommendations for respiratory, urinary, vaginal, eye, and dental symptoms. Results showed that when AI and physician recommendations differed, the AI suggestions were often rated as higher quality. The study found that 77% of AI recommendations were rated as optimal, while 67% of physicians' decisions were rated optimal. For example, patients with recurring urinary tract infections sometimes encounter antibiotic resistance. Goldzweig said the AI was successful at identifying these patients and recommended a bacterial culture before prescribing antibiotics. In contrast, doctors sometimes prescribe medication without testing, which she said could result in the infection coming back. Goldzweig added that the study suggests that the AI tool tends to be more guideline-focused, while physicians have the ability to adapt medical guidelines based on the nuance of a patient's case. Goldzweig said the study has limitations, though. It examined only a few medical conditions and didn't factor in the nuance of individual patient cases. Another limitation is that the medical chart reviewers could see whether the recommendation was made by the AI or a physician. Cedars-Sinai is working to expand CS Connect. The organization is piloting remote patient monitoring technology using AI to help people manage chronic diseases, like high blood pressure. It's also planning to use the AI technology to connect in-person urgent care visits with virtual care.

Unpacking the biggest bear on Wall Street's call for a 13% drop in the stock market
Unpacking the biggest bear on Wall Street's call for a 13% drop in the stock market

Business Insider

timean hour ago

  • Business Insider

Unpacking the biggest bear on Wall Street's call for a 13% drop in the stock market

Don't get too comfortable at all-time highs, Stifel's Barry Bannister says. Despite the market's resurgence since plunging in April, Bannister thinks stocks are probably set for a rocky second half of 2025. The chief US equity strategist at Stifel has a year-end price target of 5,500 on the S&P 500, representing a 12.8% downside from current levels. While it may seem like an unlikely scenario as the market continues to punch fresh highs, investors may want to take heed. Bannister came into the year with one of the lowest price targets among strategists at major Wall Street institutions. His caution turned out to be warranted, as the S&P 500 plunged 19% from mid-February to mid-April. Bannister has a couple of arguments in support of his outlook: valuations are high, and economic growth is likely to slow. "Just look at the core of the economy. How is the consumer and capital spending going," Bannister told Business Insider on Monday. He was referring to core GDP, or real final consumer sales and fixed business investment. Standard GDP takes into account the volatile inventory building that firms did earlier in the year to prepare for tariffs. "It looks to us like there's going to be a slowdown in the second half," Bannister added. In a note to clients this month, Bannister shared the chart below, showing the relationship between core GDP and S&P 500 returns. Current trends imply that year-over-year S&P 500 returns are set to dip negative in the months ahead. On valuations, Bannister said stocks are "expensive" and sitting at a "very high level." Another chart he shared in the note is one way of measuring how rich the market is broadly. It considers the S&P 500's price-to-earnings ratio, with the earnings part of the equation being real operating earnings and normalized on a trailing 10-year basis, similar to how the Shiller cyclically-adjusted PE ratio is. The measure is currently in "valuation extreme mania" territory, rivaling levels reached in 2021, 2000, and 1929 as it's 1.5 standard deviations away from the 10-year trend. It shows investors may be too optimistic about AI by historical standards. A macroeconomic downturn is what will likely eventually disrupt the euphoria, Bannister said. But a drop in the stock market isn't guaranteed, Bannister said. If Powell cuts rates a couple of times in the near future because inflation stays down, he said, expect stocks to jump. Others on Wall Street predict roughly flat returns the rest of the year, with many targets sitting around 6,000. But some join Bannister's greater skepticism about the health of the rally: Julian Emmanuel of Evercore ISI said a 15% decline could be coming in the months ahead. In the short-term, Bannister said he's most bullish on defensive value stocks. Over a 10-year horizon, he said he likes value stocks, small-caps, and international stocks.

An AI startup that wants to be the Canva of video has raised $9.2 million. Read its pitch deck.
An AI startup that wants to be the Canva of video has raised $9.2 million. Read its pitch deck.

Business Insider

time2 hours ago

  • Business Insider

An AI startup that wants to be the Canva of video has raised $9.2 million. Read its pitch deck.

Turning the ideas in your head into visual content has never been easier thanks to a flurry of new AI tools. And Rebecca Kossnick's goal is to simplify content creation even further for the everyday person with Hypernatural, a generative AI platform that creates both image and video content. "We think about what we're building as kind of the Canva for video," Kossnick, Hypernatural's CEO, told Business Insider. If Canva proved that you don't need a degree for graphic design to be your passion, Hypernatural wants to prove that you don't need to know Adobe After Effects to turn ideas into videos. Founded in 2023 by Kossnick and Taylor Hughes (CTO), who first met while working at Facebook, Hypernatural's video generation tech uses an average of 15 different AI models, such as Google 's Veo 3, Lightricks' LTX, and Black Forest Labs ' Flux. Kossnick said that the platform's speed is a key selling point. "We're not building a foundation model, we're building the technology to understand and deploy the foundation models as quickly as possible," Kossnick said. The startup told Business Insider exclusively that it has raised a total of $9.2 million in funding across two rounds from investors AIX Ventures, Underscore VC, Adverb Ventures, Character Capital, and 43 VC. Part of Hypernatural's total funding includes a $6.8 million seed round led by AIX that closed in December 2024. Currently a team of five, Hypernatural's funding will go toward hiring (Kossnick emphasized its focus on recruiting engineering talent) and building out the platform's technology. "We're throwing money on what's working," Kossnick said. AI video generation startups have been raising capital left and right. Moonvalley, which has its own AI video model, told BI earlier this month that it raised $84 million. Hedra, another AI video creation platform, announced a $32 million Series A round in May. Building for both creators and everyday consumers "People are excited to try and buy products that support their creativity," Kossnick said. "That's sort of where we see consumer and the 'prosumer' creator economy merge. Sometimes you're creating for other people … sometimes you're creating for you because you wanted to see the rest of the story." It's quite easy to navigate Hypernatural's tools. For instance, I made a 54-second video (with 20 AI-generated images Hypernatural created) about a journalist writing a story in roughly a minute. Here's how the platform works: Users input a text prompt and choose a style (such as anime), length, and size of the video. It then generates a full script (users can also use an existing script, and paid subscribers can upload audio files). At this stage, users can choose from a selection of narrator voices and characters (custom-made characters require a paid subscription). Hypernatural's platform then spits out a video and brings the user to an editing page where they can rearrange scenes, edit the script, and either reanimate scenes or pull from its stock library of AI-generated video. "Our users really care about the rights of all of the images and videos that we create," Kossnick said. "We're only working with models that are providing us with commercial rights, which would pass on to users as well." Its "creator" tier annual subscription costs $12 a month, and higher tiers that offer more content credits and custom characters go up to $48 a month. It also has a free tier with limited access to the platform's tools. Hypernatural said it has more than 1 million users. While Kossnick said Hypernatural's early users include "niche creators and storytellers," such as podcasters, writers, and faceless influencers, it wants to expand beyond professionals. "We're building the platform for everyone," Kossnick said. Note: The company updated and redacted some details in the deck. The deck begins with a video example. Hypernatural introduces its 5-person team. "We're a 5-person team with deep experience in video and AI," the slide reads. "Collectively, we've launched 4 companies and held product, design and eng leadership roles at Descript, Figma, Meta, YouTube, Clubhouse, and Twitter." Hypernatural emphasizes that its tool is fun. Here's what the slide says: Even bigger picture: Fun IS the point By inventing a new form of video that's fun, accessible, and personalized, we dramatically grow the pie of who make video — and how often they can do it. Right now, people export their content to post it elsewhere, but they also watch the content in Hypernatural. We envision this behavior growing and becoming dominant over time. Then it maps out the current market. It then describes its customers as storytellers. Hypernatural Here's what the slide says: Our customers today are storytellers: authors, entrepreneurs, and creative people. Authors Author 1 [608 videos] A sci-fi novelist who has serialized his books across hundreds of TikTok videos Author 2 [129 videos] A fantasy novelist who uses his videos to drive Amazon sales of his books. He recently announced that one of his books reached #1 in Amazon's Military Fantasy category Entrepreneurs Simpli Faceless/James Ward [522 videos]: A consultant for content creators who runs a creator community and multiple successful channels built on Hypernatural Witness Mi. [590 videos] A memory-preservation service built entirely on Hypernatural video creation technology Niche Content Creators Some creators' names were redacted in this slide. Then it shares customer testimonials of the platform. Courtesy of Hypernatural Here's what the slide says: "Hypernatural AI has been a game-changer for us. It's the only AI video generation platform we've found that actually produces usable, high-quality videos fit for real marketing and storytelling needs. We tested several other platforms that took hours just to generate short clips—Hypernatural gives us full-length, emotionally compelling videos in under two minutes. The speed, quality, and reliability make it clear: we've found our long-term video partner." Courtesy of Hypernatural Here's what the slide says: "I just wanted to take a moment to express how absolutely blown away I am by your video generator. I'm very new to the AI world, especially video creation, and after using your software, I'm completely speechless. I just generated an animation, and the entire experience was fun, easy, and incredibly fast. I've never used anything like this before, and I can honestly say it's the perfect software for someone like me—simple yet powerful. It made me feel like a creator in a way I never have before.... Courtesy of Hypernatural Here's what the slide says: "I can't tell you how thrilled we are with the Hypernatural software. If there is one tool we can't live without it would definitely be Hypernatural. I have a community called Simpli Faceless, and almost 100% of our members are using Hypernatural. We have been able to help people get hundreds of millions of views at this point, and there is no way we could do it without you." — James Ward, Founder of Simpli Faceless Hypernatural ends the deck by laying out its fundraising ask. "We're raising $7M, led by AIX Ventures, to pour gas on the fire," the slide reads. Here's what else the slide says: Quality Double down on our differentiating features Characters Custom Styles Expand data moat and invest more in eval tooling (we are really good at this), fine tuning, custom models We have unique, scaled data signal for what good means for generated video (good = done = exported) Marketing Scale up influencer marketing Specifically long-tail and international influencers Invest in community advocates Continue to grow organic posting/advocacy loop Scale up affiliate program Product Build the obvious features (on the endless list of features) that people are begging for. E.g. Music tracks High quality exports The ability to undo changes Right-to-left editing Bring more inference in house Drive down costs, more cutting edge models

DOWNLOAD THE APP

Get Started Now: Download the App

Ready to dive into a world of global content with local flavor? Download Daily8 app today from your preferred app store and start exploring.
app-storeplay-store