The public conversation about data centers treats them as if they appeared in Arizona by accident. A power grid attracted them. A tax break attracted them. A water table that nobody bothered to check attracted them. None of those framings explains why the world's largest semiconductor manufacturer chose Phoenix for its biggest American investment, or why Intel has been quietly expanding Ocotillo for 30 years, or why those two facts are connected to the data center buildout in ways the casual observer never sees.

Arizona is the only place in the United States where both ends of the AI supply chain sit inside a single metropolitan area. The chips that train large language models and run AI inference are getting manufactured at TSMC's North Phoenix campus and Intel's Chandler campus. The data centers that put those chips to work are 30 miles south and west, in the East Valley and West Valley. That co-location is rare, it is deliberate, and it is the part of the story that almost nobody tells.

What TSMC Actually Built

TSMC, the Taiwan Semiconductor Manufacturing Company, is the largest contract chipmaker in the world. By most estimates Taiwan produces around 90 percent of the world's most advanced semiconductors, the bulk of them at TSMC. Apple's chips, Nvidia's chips, AMD's chips, Qualcomm's chips, all of them are fabricated at TSMC fabs. The economic and strategic importance of that one company is hard to overstate. If TSMC went offline tomorrow, the global technology industry would stop within weeks.

In 2020, TSMC announced a Phoenix fab, called Fab 21, with an initial $12 billion price tag. By March 2025 the announced investment had grown to $165 billion, covering six fabs plus advanced packaging and a research and development center. The first fab is producing 4-nanometer chips. The next are targeted for 3-nanometer and 2-nanometer with future expansion. When fully built, the Phoenix campus will be one of the largest and most advanced semiconductor manufacturing complexes in the world, and the only one of its scale outside Asia.

The decision to put it in Phoenix was not made by accident. TSMC evaluated dozens of sites across multiple states and ultimately picked North Phoenix because of three things: existing semiconductor industry presence, water rights and infrastructure that could support fab-scale demand, and proximity to the talent pool that Intel had been developing for 40 years.

The talent point matters. A modern semiconductor fab employs thousands of process engineers, equipment technicians, materials scientists, and ultra-clean construction specialists. That workforce doesn't exist in most places. It exists in Phoenix because Intel has been training it since 1979.

Intel's Quiet 40-Year Build

Intel opened its first Arizona facility in 1979. Fab 12 came online in 1996. Fab 22 followed in 2007. The larger Ocotillo expansion built out through the 2010s. By 2022 and 2023, Intel was operating a multi-billion-dollar campus on the south side of Chandler with active fab capacity, an extreme ultraviolet (EUV) lithography line, and a workforce of roughly 12,000 people locally, a figure that came down with the company's 2024 restructuring.

Intel's recent decade has been turbulent. The company missed several manufacturing process transitions, lost market share in mobile and AI, and announced significant restructuring through 2024 and 2025. None of that erased the Arizona footprint. The fabs are still running. The workforce is still there. The supply chain that grew up around Ocotillo, the gas vendors, chemical suppliers, photoresist distributors, equipment service contractors, is still there. That ecosystem is what makes a fab city a fab city, and Phoenix has been one for two generations.

When TSMC came looking, the supply chain was waiting. When the next fab came looking, the supply chain expanded. The compound infrastructure is what economists call agglomeration. Once it forms, it self-reinforces.

Why Co-Location Matters for AI

Most of the world's AI workloads run on a small set of accelerator chips: Nvidia's H100, H200, Blackwell, and the upcoming Rubin generation; AMD's MI300 and MI400; custom silicon from Google (TPU), AWS (Trainium and Inferentia), Meta, and Microsoft. Each of these is designed by a company in California or Washington and manufactured by TSMC.

In the conventional supply chain, those chips would be made in Taiwan, packaged in Malaysia or the Philippines, tested in another country, and shipped to a data center somewhere else. That works fine in normal times. It does not work in geopolitical crisis, and the entire AI buildout is happening in an era where the geopolitical assumptions of the last 40 years are visibly fraying.

The CHIPS and Science Act of 2022 was Congress's response to that fragility. The law provided roughly $52.7 billion for semiconductor research and manufacturing, plus a separate 25 percent investment tax credit, to bring chip production back to the United States. TSMC's Arizona expansion, Intel's Arizona expansion, Samsung's Texas fab, and several smaller facilities are all consequences of that policy. The legislation passed with bipartisan support because national security planners, military leadership, and the executive branch all reached the same conclusion: depending on Taiwan for the world's most strategically important commodity is a vulnerability the United States cannot afford.

The data center buildout in Arizona is part of the same calculation, even though it doesn't usually get framed that way. If you're going to manufacture AI chips in Phoenix, the natural buyer of those chips is a hyperscale data center within driving distance of the fab. The chips don't have to drive across the Pacific Ocean to get used. They go to a building 30 miles away, get installed in a server rack, and start running training jobs the next week.

That tight coupling is unusual. It exists in only a few places in the world: Hsinchu, Taiwan, where the original TSMC fabs sit near the Asian hyperscale buildout; Beijing, where China is racing to build domestic equivalents; and now Phoenix.

The Power Density Problem That Connects the Two

There's a second technical reason chips and data centers belong near each other, and it has nothing to do with shipping. The chips coming out of TSMC's 4-nanometer and 3-nanometer nodes are extraordinarily power-dense. An Nvidia H200 GPU draws roughly 700 watts continuous. A Blackwell B200 draws over 1,000 watts. The next generation will draw more.

When you stack 72 of those into a single rack, as Nvidia's GB200 NVL72 system does, you get a rack that draws on the order of 120 kilowatts. A standard data center rack 10 years ago drew 5 kilowatts. The thermal density of a modern AI rack is closer to a steel mill furnace than to a traditional server room. The data center has to be designed around that thermal load, with liquid cooling plumbed to every rack, with high-amperage electrical distribution, with structural reinforcement for the weight, with redundant cooling and power that match the failure tolerance of the chips themselves.

Designing a building like that is not a generic data center skill. It's a specialty. The construction crews, mechanical contractors, electrical contractors, and controls integrators who can do it well are concentrated where the work is. Phoenix has them because Phoenix has been building this kind of facility at scale for several years. The next campus benefits from the experience of the previous campus. The construction timeline shrinks. The cost per megawatt drops. The reliability improves.

The chips drove the rack design. The rack design drove the building design. The building design drove the contractor specialization. The contractor specialization drove the regional concentration. Phoenix is where this stack came together.

The Strategic Picture

What's happening in Phoenix is, viewed from a national security perspective, the construction of a domestic AI supply chain end-to-end. Chip manufacturing at TSMC and Intel. Chip packaging at the under-construction packaging facilities (Amkor's roughly $7 billion Peoria expansion, ASE Technology Holding's investments). Substrate manufacturing at the suppliers TSMC pulled in. Server assembly at the contract manufacturers. Hyperscale deployment at the data center campuses. Power generation from the regional grid mix including Palo Verde nuclear. Workforce training through community colleges, ASU, the apprenticeship programs.

Every one of those links is in or adjacent to the Phoenix metro. That is not true anywhere else in the United States. Texas has the data centers but doesn't have leading-edge fabs. Oregon has fabs but doesn't have the data center concentration. New York and Ohio are building toward it but are years behind. The full stack, in one metro, is here.

That's the part of the story the water-and-power conversation buries. Arizona didn't get picked for tax breaks. Arizona is becoming the central node of the American AI manufacturing strategy, and the data centers are one piece of an infrastructure investment that is among the largest industrial buildouts in modern American history. Whether the public framing catches up to that reality or not, the buildout is happening, and the strategic logic behind it is solid.

Sources