The defensive arguments about data centers get tedious because they are stuck on the wrong side of the equation. Every conversation about water, every conversation about power, every conversation about land, every conversation about noise, treats the data center as a cost that has to be justified. The justification, when offered, tends to come back to abstract economic activity: jobs, taxes, technology leadership. The numbers are real. They also feel hollow because they don't connect to anything the listener cares about personally.

Let's try the other direction. What does a hyperscale data center actually do? Not in the abstract. In the concrete. Take a single 100-megawatt building somewhere in the West Valley, the kind of facility that an Arizona resident might drive past without noticing, and trace what runs through it on an ordinary Tuesday afternoon.

The Radiologist in Tucson

Imagine a 47-year-old woman in a community hospital in Tucson getting a chest CT scan. The radiologist on duty has 30 scans in the queue and an average of 8 minutes per case to read each one. That ratio is not a sustainable workload, and it is the ratio that radiology departments across rural and mid-size American hospitals have been operating under for years.

The CT image gets routed to an AI screening system before it reaches the radiologist. The AI runs on GPUs in a data center somewhere in the Phoenix metro. In the time it takes the radiologist to walk from one workstation to the next, the AI has flagged a small lesion in the lower lobe of the right lung, drawn a measurement line across it, and pulled up the patient's prior imaging from two years ago for comparison. The lesion was not visible on the prior scan. The AI doesn't make the diagnosis. It surfaces the finding so the radiologist sees it within the 8-minute window instead of missing it under workload pressure.

Multiply that workflow across thousands of community hospitals, dozens of medical AI vendors, and millions of imaging studies a year. The compute that does the work is not in the hospital. The hospital does not own a data center. The hospital pays a per-study fee to an AI vendor whose servers sit in a building near Mesa or Goodyear or Casa Grande. The patient never sees the building. The radiologist never sees the building. The lung cancer that gets caught at stage I instead of stage III gets caught because the building is there.

This is not speculative. Aidoc, Annalise, RapidAI, Viz.ai, dozens of other clinical AI vendors are deployed in U.S. hospitals today. The FDA has now authorized more than 1,000 AI- and machine-learning-enabled medical devices as of its December 2024 list. The infrastructure they run on is hyperscale public cloud. The cloud is data centers. The data centers, increasingly, are in Arizona.

The Student Who Couldn't Hear

A college sophomore who is deaf attends a 90-minute lecture on Tuesday morning at a university in the Midwest. The lecture is on organic chemistry, taught by a professor with a mild speech pattern that human captioners struggled to keep up with last semester. The university paid for live human captioning anyway because the alternative was failing to comply with the ADA and losing federal funding.

This semester, the student is using an AI captioning service. The audio from the classroom microphone streams to a cloud service, gets transcribed in real time by a speech recognition model, gets correction-passed for technical vocabulary, and arrives on the student's laptop with under a second of delay. The captions are more accurate than last semester's human captioning, especially on the chemistry-specific terms. The cost to the university is a small fraction of what live captioning cost. The student is no longer at the mercy of the captioner's familiarity with the subject material.

The speech recognition model that does the transcription is hosted on hyperscale infrastructure. The latency budget for real-time captioning, from microphone to screen, is roughly 800 milliseconds. That budget can only be met if the data center hosting the model is geographically close enough that the round-trip network time is short. A data center in Arizona serves a student in Iowa with about 30 milliseconds of network latency, well inside the budget. A data center in another country could not meet the budget. The student's accommodation only works because the infrastructure is here.

Multiply across thousands of deaf students nationally, millions of meeting participants who use captions for any reason, every Zoom call with cloud auto-captions enabled, every YouTube video with AI-generated subtitles, every accessibility feature that didn't exist 5 years ago and is now considered standard. The cloud-hosted versions all run on data centers, and they require the latency that only nearby infrastructure provides.

The Farmer Who Saved Half His Water

A 1,200-acre farm operation outside Yuma installed precision irrigation sensors three years ago. The system uses soil moisture probes, weather data, and a model trained on the specific crop, soil type, and microclimate of each field. The model runs in a data center somewhere in the Southwest. Every 15 minutes, it integrates new sensor readings, updates its predictions, and sends irrigation commands to the field-level controllers. The system has reduced the farm's water consumption by roughly 40 percent compared to the prior schedule-based irrigation, while maintaining or improving yield.

The farmer is not a software engineer. He doesn't know what cloud the AI runs in. He pays a per-acre subscription to an ag-tech vendor (FarmWise, Climate FieldView, Cropwise, others) and he gets the water savings in his utility bill. The water that doesn't get pumped doesn't deplete the aquifer. The aquifer that doesn't deplete has more headroom for the next 30 years of agricultural production. The water savings, multiplied across the precision agriculture deployments that are scaling across the Southwest, are larger than the entire data center industry's water footprint by orders of magnitude.

This is the irony nobody mentions in the water debate. The same compute infrastructure that critics worry uses too much water is also the infrastructure that powers the AI models that are saving farmers more water than the data centers consume. The net effect on regional water budgets, properly accounted, is favorable. Nobody has run the comprehensive number, but the order of magnitude is clear.

The Veteran in a Rural Clinic

A Vietnam-era veteran lives 90 miles from the nearest VA medical center. He has chronic conditions that require monthly check-ins. The drive used to be a 3-hour round trip plus the appointment time, and he missed appointments routinely because he couldn't make the trip.

His current care is mostly telehealth. The video consultation runs through the VA's secure cloud platform. The clinical decision support that the nurse uses to update his medications runs on AI models in the same cloud. The prescription that gets sent to his rural pharmacy gets transmitted, verified, and dispensed in minutes through interconnected systems that didn't exist 10 years ago. The whole care episode happens at his kitchen table.

The VA's cloud infrastructure is hosted across multiple data center regions. Several of those regions are in Arizona. The latency from his home to the nearest data center is short enough that the video call is high-quality and the prescription transaction completes faster than a manual workflow ever could. He goes to the VA in person twice a year now instead of monthly. His outcomes are measurably better. The cost to the system is lower.

The reason any of this works is that the compute infrastructure is geographically distributed and close to where he lives. A telehealth model that depended on infrastructure in Northern Virginia or another country would have higher latency, more dropouts, and worse user experience. The model that works depends on the infrastructure being where the patient is. That includes Arizona.

The Three-Person Company in Flagstaff

A small business owner in Flagstaff runs a specialty retail operation with three employees. She uses Shopify for her storefront, QuickBooks Online for her books, Slack for team communication, Microsoft 365 for documents, Salesforce for customer management, Mailchimp for marketing, and Stripe for payments. Her total monthly software bill is roughly $800. The infrastructure that runs all of these tools is hyperscale public cloud.

In 1995, the equivalent capability would have required an IT staff, a server closet, software licenses with five-figure upfront costs, and integration work that her three-person business could not have afforded. The democratization of enterprise software, the fact that a small business in Flagstaff can run on the same Salesforce stack as Procter and Gamble, only exists because hyperscale data centers made the marginal cost of serving an additional small business approximately zero. The infrastructure exists. The software runs on it. The small business pays a tiny fraction of what enterprise software used to cost.

This is not a marketing claim. It is the basic structural fact of how the cloud changed business economics. Every small business in America benefits from this, whether or not they are paying attention to where the servers actually are. The servers are in places like Phoenix. The benefits propagate everywhere.

The Volume of It All

The radiologist, the student, the farmer, the veteran, the small business owner. Five stories that all run through the same data center infrastructure. Now multiply.

Start with the input the critics fixate on. A 100-megawatt facility running flat around the clock draws about 876,000 megawatt-hours a year, which is roughly the annual electricity use of 80,000 U.S. homes (the average home buys about 10,800 kilowatt-hours a year). That is the cost side, stated plainly. Now the output side. A single 100-megawatt hyperscale building can host roughly 50,000 to 100,000 GPU-equivalent workloads simultaneously, depending on architecture. It can serve millions of users in real-time interactive applications. It can run training workloads that update billions of model parameters in days. The throughput of one building is on the order of magnitude of the entire compute capacity of the world in 1995.

What that capacity gets used for, at scale, is the actual answer to the question of whether data centers are worth it. The radiology AI that catches cancers earlier. The captioning that makes lectures accessible. The precision agriculture that saves aquifers. The telehealth that serves rural patients. The democratized software that lets small businesses compete. The drug discovery pipelines that compress research timelines. The climate models that improve weather forecasting. The translation services that break down language barriers. The accessibility tools that read screens to blind users. The fraud detection that protects bank accounts. The mapping and routing that gets ambulances to crash sites faster.

All of this is the output side of the equation. None of it is hypothetical. All of it depends on the infrastructure being there.

The Frame That Was Missing

The defensive argument keeps getting reshuffled because it never quite lands. The reason it never lands is that the defensive argument is operating in the wrong frame. It is justifying inputs without describing outputs. It is explaining what a data center costs without explaining what it does.

The actual answer, when written out, is straightforward. A 100-megawatt data center is the physical infrastructure for several million people's daily access to capabilities that, 15 years ago, were either impossible, unaffordable, or available only to large institutions. The capabilities are concrete. They are improving lives in measurable ways. They are getting better every year. They are not abstract economic activity. They are radiology results, captions on a lecture, water savings on a farm, prescription refills, lower software costs, better fraud protection, faster ambulance routes.

That is the offensive frame. It is also, conveniently, the truth. The buildings exist because they are necessary infrastructure for things people actually want. The conversation can keep arguing about whether to build them or not. The conversation should at least be clear about what the buildings do.

Sources