02 / The Physical Address

Compute has a physical address.

The infrastructure behind cloud, AI, and every shipped chip.

02 / The physical address

Where abstraction ends.

This page maps the systems that make compute a place: sites, utilities, tools, subfabs, movement, test, and packaging.

A software product can ship globally in seconds, but compute capacity is gated by physical throughput. Every transistor on every die passed through a building that sits on real land, draws real megawatts, pumps real gallons, and employs real people wearing real bunny suits. There is no abstraction layer for concrete.

"The cloud is someone else's fab, package substrate, board, rack, power contract, and yield curve."

Every API call, every model inference, every terabyte of training data traces back to a building with a street address, a utility meter, and a workforce. The cloud is not ephemeral. It is physical infrastructure with physical constraints, physical costs, and physical failure modes. The Physical Address is where abstraction ends and engineering begins.

SITE

What It Is

The site is the parcel of land, the foundation, the building envelope, the cleanroom shell, and the utility interconnections that must exist before a single tool is installed. A leading-edge fab is among the largest and most expensive buildings ever constructed. Intel's Ohio One campus spans 2.5 million square feet including 600,000 square feet of cleanroom production space. Samsung's Taylor complex sits on 1,200 acres of former farmland in Williamson County, Texas, with approximately 4 million square feet already constructed including a six-story office building. TSMC's Arizona campus is on track to become a "gigafab cluster" with $65 billion committed to three fabs and an additional $100 billion pledged for future expansion, making it the largest foreign direct investment in a greenfield project in U.S. history.

A single advanced fab (sub-4nm node) costs $10-20 billion or more. Intel's two Ohio fabs are budgeted at roughly $28 billion. Samsung's Taylor investment has expanded from an initial $17 billion to an estimated $44 billion with the addition of a second advanced fab and expanded R&D. TSMC's three Phoenix fabs total $65 billion. The CHIPS Act has driven over $540 billion in announced investments across 28+ U.S. states, transforming Arizona, Texas, and Ohio into the primary nodes of American semiconductor construction.

Building fabs in the U.S. is estimated to be 30-50% more expensive than in Asia. Arizona and Texas represent mature ecosystems with existing semiconductor supply chains; Ohio represents a "build it and they will come" approach. The construction timelines are sobering: TSMC CEO C.C. Wei noted that establishing the Arizona plant has taken "twice as long as similar facilities in Taiwan." Intel's Ohio project, originally slated for 2026 operation, has been pushed to 2030-2031. Samsung's Taylor fab, announced in 2021 for 2024 operation, is now targeting partial operation by late 2026 with full production by 2028.

Why It Matters

Site selection is the first irreversible decision. You cannot relocate a fab. The building must survive decades, accommodate technology transitions, and connect to utility infrastructure that takes 10-20 years to fully ramp. The electric, water, and wastewater requirements can rival that of a moderately-sized town, and the rise of more comprehensive ESG compliance standards is now a factor in siting decisions.

What Constrains It

  • Cost premium: 30-50% more expensive to build in the U.S. vs. Asia
  • Utility infrastructure: 10-20 year ramp-up for electric, water, wastewater, gas delivery
  • Talent availability: Each fab requires thousands of specialized workers; Intel has 162 Ohio employees as of early 2025
  • Construction timeline: U.S. projects consistently delayed 2-4 years vs. Asian equivalents
  • Ecosystem maturity: Ohio lacks the dense supplier network of Phoenix or Austin

Connects To

Utilities (power and water contracts are negotiated before ground breaks); Subfab & Vacuum (foundation must accommodate raised floors creating 1-3 meter plenum); Tools (building vibration specs determine lithography placement); Environmental (ESG requirements affect siting).

The Integrated View: Three American Megafabs

Key point: this table is evidence for the central claim. Compute capacity moves at the speed of sites, permits, utilities, tools, people, and qualification, not at the speed of software.

ParameterTSMC ArizonaIntel OhioSamsung Taylor
Total Investment$65B (3 fabs) + $100B pledged$28B (2 fabs)$17B initial → $44B expanded
Power (Phase 1)~200 MWTBDTBD
Power (Full Build)1 GW+TBDTBD
Site SizeGigafab cluster2.5M sq ft total; 600K sq ft cleanroom1,200 acres; ~4M sq ft built
Original Target2024 (Fab 1)20262024
Current TargetQ4 2024 (Fab 1 producing); Fab 2 H2 20272030-2031Partial: late 2026; Full: 2028
Process NodeN4 (Fab 1); N3 (Fab 2); N2/A16 (Fab 3)18A → future nodes4nm → upgraded to 2nm
CHIPS Act Funding$6.6B direct + $5B loans$7.865B$4.75B
Key CustomerApple, AMD, Nvidia, Broadcom, QualcommIntel internal + foundry customersTesla ($16.5B deal)

The pattern is unmistakable: every U.S. megafab is years behind schedule and billions over initial budget. The physical address of compute is not a line item, it is the line item. It absorbs capital on a scale that few other industrial projects in human history have matched. And it moves slower than the software it enables, creating the fundamental tension: a codebase can ship in a sprint cycle, but a fab takes a decade.

The site is the container. What it contains is one of the most complex machines humans have ever built.