Data centers are a three decade American story

In a clinical trial reported in 2023​¹, sixteen people who had just had surgery for pancreatic cancer received a vaccine reverse engineered from the DNA of their own tumor. Researchers sequenced each tumor alongside the patient’s own healthy tissue, showing the mutations that exist only in their cancer. Software then ranked which mutations the patient’s immune system was most likely to recognize, and the ones that scored highest became the vaccine’s targets.

Eight of the sixteen patients developed a strong immune response, and when researchers followed up three years later​² those eight had gone far longer without cancer recurrence.

A randomized trial in melanoma has since gone further using the same approach, a personalized vaccine designed from each patient’s own tumor. Its five-year update in 2026​³ stated that patients who received a personalized vaccine alongside standard immunotherapy had roughly half the risk of recurrence or death, in a 157-person trial.

Reverse engineering a genetic sequence, whether for a human's entire genome or a tumor is no simple feat. It requires incredible amounts of compute to computationally process every piece of DNA. And this type of compute at scale is only made possible by data centers.

One of those data centers sits at the edge of Quincy, Washington, a farm town where 8,100 people drink from five deep wells. For years, it cooled its servers with the same water​⁴. In that time, the town ensured equitable systems for all, and the tax base tripled and paid for a new high school.

Data centers were part of American life far before AI

Data centers are not new. In fact, they're a 30 year American story. By 1995 the United States federal government had built so many, many of them small and duplicating one another, that running them had become a significant expense. That October the White House began ordering federal agencies to consolidate​⁵, closing redundant sites and consolidating into fewer, larger ones within two years, on the expectation of cutting operating costs by 30 to 50 percent. The primary concerns as this consolidation began focused on staffing, cooling, security, and the expense of moving. And the Federal government, as it is mandated, carefully tracked the results throughout this process.

In 2007 electricity usage by data centers came into stronger focus, and Congress asked the Environmental Protection Agency to measure it. The EPA found​⁶ that U.S. data centers had consumed about 61 billion kilowatt-hours in 2006, or 1.5% of national electricity. Demand had doubled in six years, and grid reliability was a chief concern. The EPA projected​⁷ that consumption could pass 100 billion kilowatt-hours by 2011.

But what happened next was a fascinating study in electrical efficiency driven by commercial innovation. Contrary to the earlier projections, virtualization and efficient hyperscale buildings held demand nearly flat for a decade​⁸ while the work they did multiplied. Eric Masanet and his colleagues​⁹ later measured the decoupling in Science magazine: between 2010 and 2018, global data center computing grew about 550% while energy use grew about 6%.

Of course the age of AI, particularly with the incredible advances in model intelligence in 2026, has upended previous growth rates. New AI campuses request hundreds of megawatts apiece, and the largest now cross a gigawatt​¹⁰. The entire fleet the EPA studied in 2006 drew about seven gigawatts on average.

A June 2026 update from Lawrence Berkeley National Laboratory projects data center electricity usage at 9.5% to 15.3% of total U.S. use by 2030. But what we're already seeing again are continued gains in innovation that may keep electrical usage on the lower end of the spectrum while computing use once again explodes, coupled with incredible advances in water cooling systems and more that keep data centers equitable and beneficial to all.

Data centers bring prosperity and investment into communities through taxation

Three numbers decide what a data center is worth to the place that hosts it: what the company spends building it, what the state gives up in tax breaks to attract it, and what the county collects in property taxes every year after it opens.

The first two are easy to see in Council Bluffs, Iowa, a city of 62,000 across the river from Omaha. Since 2007, Google has spent $5.5 billion there​¹¹ and has since announced another billion​¹² for the campus, alongside $1.3 million toward river restoration. The state added incentives of its own, including a $19.8 million sales and use tax refund​¹³ for the expansion. The site employs more than 900 people across technical, maintenance, security, and food service roles on an ongoing basis.

In Virginia, Loudoun County sits on the western edge of the Washington, DC suburbs, where subdivisions give way to horse farms and Dulles airport, and the town of Ashburn within it holds the largest concentration of data centers in the world.

While most counties tax the land and the building where commercial entities sit, Loudoun also taxes the computer equipment inside and its fiscal 2026 budget​¹⁴ projected $794.8 million from that tax alone. And that revenue is outsized: the county reports​¹⁵ that the facilities occupy about 4% of its commercial parcels while producing 38% of General Fund revenue.

In December 2024, Virginia's nonpartisan legislative auditors found data center revenue ranging from less than 1% to 31% of total local revenue across five mature markets in 2023, with officials reporting schools built in Mecklenburg, an affordable-housing trust in Henrico, reserve funds, and lower real-estate tax rates as a result of this revenue. The same audit also found the effect on jobs: a typical 250,000-square-foot facility may settle at about 50 permanent workers, and a county collects franchise fees on its electricity on top of the property tax.

In Loudoun's case, while the assessed value of the equipment inside may fall slightly each year, the revenue is durable and long term and is not the result of a construction-boom spike. In fact, reinvestment in the facilities in Loudoun will translate into an increase for the fiscal 2027 budget​¹⁶ to $879 million, and close to $1.3 billion once real property is counted, about 45% of all county tax revenue fueling schools, medical centers, healthcare, affordable-housing and more.

The report also notes a projected further $35 billion of data center investment in new Virginia locations by 2040.

The newest data center cooling systems use almost no water

Data centers consume power that fuels computer chips that generate heat and require cooling. There are a number of methods by which these chips can be cooled: cooling towers that evaporate water, dry coolers that shed heat to the air, and closed direct-to-chip loop systems that reuse the same water after each cycle.

Different cooling approaches carry vastly different efficiencies. A 2025 Lawrence Berkeley study​¹⁷ found that the same computing job can consume ten thousand times more water in one facility than in another, depending on the servers, the grid, the cooling, and the climate.

And yet nationally, the relative totals for water usage are small. Lawrence Berkeley put the direct water consumption​¹⁸ of every U.S. data center at 17.4 billion gallons in 2023, with roughly another 211 billion used by the power plants supplying them. That is about 228 billion gallons for the year. The United States withdraws around 322 billion gallons​¹⁹ of water every day, so a full year of data center use is less than a single day of national withdrawal.

But water use is also a granular, local issue, not a purely national one.

Quincy, Washington shows us that local, small-town story. Its 8,100 residents share five deep wells, and the EPA estimates that its data centers once accounted for about 10% of the water pumped from them. What the city and Microsoft built in response is the Quincy Water Reuse Utility, which since 2021 has treated the cooling water and sent it back to the data centers, returning an estimated 138 million gallons of groundwater a year. Getting there took more than ten years and $31 million, because the original design cooled servers with drinking water and sent the mineral-rich remainder into a treatment plant that was much more costly than the current system.

Even while that more efficient system was being built, Quincy’s property tax revenue tripled​²⁰, resident rates fell, and the same buildings drawing the wells paid for a new public high school, a medical center, a fire station, and an aquatic center. Microsoft counts $123.1 million in property taxes since 2019. The town ended up both solving water efficiency obstacles and the financial prosperity to build things it could not have built otherwise. What Quincy spent a decade inventing, a town today can ask for in an initial agreement.

The original water cooling design in Quincy is now being retired industry-wide. Microsoft now reports using about 90% less water per unit of computing than its first generation of data centers in the early 2000s, with average water use effectiveness down from 2.3 liters per kilowatt-hour to roughly 0.3.

At its Fairwater AI data center in Wisconsin, Microsoft says​²¹ more than 90% of capacity runs on a closed water loop that is filled once at construction and recirculates without evaporation, a design intended to consume zero water for cooling​²². A site built on a closed water loop draws almost nothing from the local supply and sends nothing back out.

Every Microsoft data center designed since August 2024 uses that approach, with the first of them coming online in late 2027, and Oracle announced the same design for its AI sites in February 2026.

Where these deals go wrong: water, grid costs, noise, and budgets built on one taxpayer

Electricity is where investment and costs can go wrong at scale. When a utility builds new power lines and generating capacity for a data center that arrives late, or smaller than promised, or never at all, households on that grid are left paying for equipment built for one customer.

In June 2026 the Federal Energy Regulatory Commission ordered grid operators to stop that happening in a huge step forward for sensible data center investment and roll out.​²³

Noise is another major concern. Virginia’s auditors found that the low hum of a data center runs around the clock and can fall outside rules written for ordinary industrial noise. Loudoun now requires a separate review for each project, plus soundproofing, setbacks from homes, and limits on when generators can be tested that lead in new investment in data centers nationally.​²⁴

Economic insurance is also another risk accounted for in Loudoun. The county draws close to half its tax revenue from a single industry and so it keeps a reserve fund in case that ever reverses.

Rising electricity bills are also often blamed on data centers, but the picture is mixed. California’s increases are mostly the cost of wildfires​²⁵. New England’s follow the price of imported gas, and two of the sharpest recent jumps were in Maine and Montana​²⁶, neither of them data center states. Sometimes data centers really are the cause, though how much is contested. In the mid-Atlantic, the price utilities pay to keep spare power on hand went from about $29 per megawatt-day in 2024 to $329 in 2026​²⁷. The grid’s independent monitor blamed data centers for $6.3 billion of the latest $16.4 billion auction​²⁸. PJM itself points to supply as much as demand, with more than 30 gigawatts of coal and gas plants due to retire by 2030 and retirements running ahead of new generation​²⁹. Virginia also hosts more data centers than anywhere on earth and saw only modest increases.

Each of these scenarios would have benefited from strong, sensible oversight that places innovation, economic benefit and potential costs equally at the forefront and grounded in facts.

Power comes first: how much the site can draw at maximum, and who pays for the new lines and substations. Then water: which cooling system it will use, where the water comes from, how much it takes and how much it never returns, and what happens in a drought. Then the local impact: limits on noise, generator testing, traffic and emissions. Then the financials: how many jobs during construction and how many after it opens, and what the county will actually collect each year once the tax breaks run out, set alongside the schools, roads or other major public benefits that money is meant to fund.

Compute at scale benefits humanity

The future of AI and the data centers that power them is vast and complex and requires discussion grounded in facts and honest debate.

The reality is that humanity stands to gain far more than we can comprehend from AI today. That greater picture must remain in focus in any good faith discussion.

Most consumer understanding of AI as of August 2026 is in chatbots and image generation. But the vast majority of future compute continues to be dedicated to far more existential endeavors that stand to benefit humanity as a whole.

The same buildings many have thought are overrun with image generation actually run the protein database and compute behind cutting edge cancer trials and other disease research, processes that require far more machine effort than generative AI for images. They're also dominated by the models used by cutting edge software developers and scientists today. And many more highly experimental or commercial activity that stands to benefit us all through innovation.

The pace at which these systems are becoming more intelligent continues to astound even the most seasoned scientists and researchers, too.

Paul Erdős spent a lifetime posing mathematics problems he could not solve himself, and hundreds of them have stayed open for decades since his death in 1996. In May 2026 an AI system was set loose on 353 of them and solved nine, with every proof checked by machine, at a compute cost of a few hundred dollars per problem​³⁰. That same month an OpenAI model disproved a question Erdős asked in 1946, about how many pairs among a set of points in a plane can sit exactly one unit apart. Mathematicians had assumed for eighty years that a square grid was about as good as it gets. The disproof was checked and written up by a group of mathematicians including the Fields medalist Tim Gowers​³¹. Stanford's 2026 AI Index further notes that the tests it has created to track AI skill for years have been beaten in months.

In biology, protein structure was the field's central bottleneck for fifty years, and the AlphaFold database now holds predicted structures for more than 214 million proteins​³², freely available and folded into the reference resources researchers already use.

As of 2026, advances in AI backed science and technology vastly outpace any year prior, and that pace is only quickening.

AI innovation is the precise reason why we can now create personalized cancer vaccines on the fly that will continue to save lives in a way not previously possible without the compute data centers provide, and it's also the reason why a new world of opportunity and scientific advancement will benefit humanity in ways we can't yet comprehend.