The Datacenter Water Floor

TL;DR

Two of the most capital-intensive industries of the next decade — AI compute and advanced chip manufacturing — are competing for the same physical input in the same stressed watersheds at the same time. A hyperscale fab consumes 2–10 million gallons of ultrapure water per day. A 100 MW datacenter rejects 70–80 MW of heat continuously, most of it through evaporative cooling that does not return the water. The CHIPS Act capital is in motion. The water permits are not. The binding constraint on the next decade of compute is not transistor density. It is the thermodynamics of heat rejection.

The market is pricing this as a siting problem. It is a materials problem. Photonic interconnects cut datacenter energy by 30–50% on GPU-to-GPU communication.¹ Superconducting interconnects take it further. Both depend on materials that aren’t in volume production at the scale the industry needs.

§1 — Where the water goes

A modern semiconductor fab needs ultrapure water — water stripped of every ion, particle, and biological trace, thousands of times cleaner than drinking water. Wafers are cleaned, etched, rinsed, and cleaned again, dozens of times per fabrication cycle. Producing the UPW itself loses 20–25% of incoming raw water.² The fab draws a million gallons; 200,000 are gone before a wafer is touched.

A datacenter’s water problem is different in kind. Cooling towers evaporate water to reject server heat. Once it leaves the tower, it is gone. Texas datacenters were projected to consume 49 billion gallons in 2025, with 2030 projections approaching 399 billion.³ Communities in Hays County have blocked proposed datacenter developments by council vote. San Marcos rejected one outright. Water permitting — not capital, not bandwidth, not power — has become the hard constraint on where new compute can land.

Two industries, same watershed, same decade.

§2 — The interconnect wall

Inside the datacenter, a second physics problem is closing in on the first. As AI models scale to trillions of parameters distributed across thousands of accelerators, the bottleneck has shifted from compute to communication. Electronic interconnects — copper traces, advanced silicon packaging, even chiplet-to-chiplet bridges — are hitting bandwidth and energy limits.⁴ Moving data between GPUs is now more expensive than the matrix math itself.

NVIDIA has publicly committed roughly $2B to photonics R&D — optical interconnects and photonic computing.⁵ This is not a moonshot allocation; it is the company that owns the AI compute stack hedging against the thermodynamic ceiling of its own architecture. Photons move data faster, cooler, and with dramatically less energy than electrons at scale. Photonic interconnects could cut datacenter energy 30–50% on GPU-to-GPU traffic alone.¹ Less energy dissipated means less heat to reject. Less heat to reject means less water evaporated. The interconnect problem and the water problem are the same problem.

The materials underneath are what nobody outside the supply chain sees. A photonic interconnect needs a low-loss waveguide platform — typically silicon nitride or thin-film lithium niobate. It needs modulators, photodetectors, packaging that survives reflow at scale. The silicon-photonics fab line at GlobalFoundries (GF Fotonix) is one of two in the world running at production volume.⁶ That is the entire ecosystem. The capacity to build the next architecture is roughly half a fab.

§3 — The next floor down

Photonics moves the floor. Superconductors take it out from under.

Superconducting interconnects carry signal with zero resistive loss. Single-flux-quantum logic operates at femtojoule-per-operation energies — three to four orders of magnitude below CMOS.⁷ The thermal load on a superconducting digital system is dominated not by the logic itself but by the cryogenic cooling overhead. At sufficient scale (tens of MW datacenters and up), the overhead amortizes. At sub-megawatt scale it doesn’t.

This is why superconducting compute has been a “twenty years away” story for forty years. The physics works. The cryogenic plant economics only work at hyperscale. The materials supply chain to deliver superconducting wire and tape at the volume the industry would need does not exist yet.

REBCO tape — the high-temperature superconductor that powers the current fusion-magnet supply chain — is produced by a small number of companies globally.⁸ Annual global REBCO output is measured in kilometers, not megameters. A single hyperscale superconducting datacenter would absorb several years of current global production. The physics works. The factory doesn’t.

§4 — Why Central Texas is the collision point

The national picture compresses sharply when you zoom in on Texas. CHIPS Act capital, hyperscale buildouts, and ERCOT’s relatively permissive interconnection regime have made Central Texas the densest collision between fab water demand and datacenter cooling demand in North America.

Samsung Austin Semiconductor anchors the fab side. The Taylor expansion (~$17B announced commitment, ~30 miles north of Austin) raised Samsung’s local water draw substantially when it came online, and the build-out continues through 2027. On the compute side, the corridor from Round Rock through Pflugerville and Taylor has become a major hyperscaler footprint, with additional capacity east toward Lockhart and west into the Hill Country. Texas datacenter capital expenditure ran up roughly 3,000% over the five years from 2019 to 2024, exceeding $10B annually by the end of that window.⁹

Both industries draw from the same constrained sources: the Edwards Aquifer, the Colorado River basin, and the Trinity Aquifer. None of those were engineered for this scale. The downstream consequences are now visible in permitting:

• Texas datacenter water consumption was projected at ~49 billion gallons in 2025, with Texas Water Development Board projections reaching 399 billion gallons annually by 2030.³
• Hays County and San Marcos have actively blocked or rejected proposed datacenter developments by council vote, citing water draw — a meaningful precedent because pre-2023 the assumption was that Texas’s permitting environment would say yes by default.
• EY’s 2026 Geostrategic Outlook stated explicitly that “access to water rights and regulatory approval — not investment appetite or technological capability — is becoming the decisive factor in where fabs can be built or expanded.” That sentence applies with equal force to datacenters.

For the ERCOT side of the picture, the load consequence matters as much as the water consequence. Datacenters are flat 24/7 consumers — they fill in what used to be the overnight load trough. Combined with bitcoin mining (a flexible price-following load that lifts the shoulders), the Texas grid load curve has flattened into a wide plateau with an evening ramp instead of a sharp afternoon peak. Storage and gas peakers fight for that ramp.

Updated 2026-04-28 — covered separately in The ERCOT Curve Has Broad Shoulders Now.

The collision is not theoretical and it is not five years away. It is happening now in the permits being denied in Hays County, in the Samsung Taylor water budget, and in the ERCOT interconnection queue. Central Texas is the place the broader infrastructure story is going to break first — for better or worse.

§5 — Where a materials company sits

The standard framing of the datacenter water problem is geographic: build fabs near water, build datacenters near water, repermit when you can. This is a real problem and it has a real answer, but it is a finite answer. There are only so many water-secure watersheds and the CHIPS Act capital is overrunning them faster than permits can keep up.

The structural answer is a different stack. Photonic interconnects to cut the cooling load. Superconducting interconnects and storage where the scale justifies cryogenic overhead. The transition is bottlenecked on materials that don’t exist at production volume:

1. Low-loss photonic substrates with the right CTE match for high-density 3D packaging
2. High-temperature superconducting tape at 10–100× current global production capacity
3. Cryogenic-compatible packaging that survives the thermal cycling between assembly and operating temperature

Each of these is a materials problem with a clean spec. None of them are software problems or capital-allocation problems. They are scale-up problems, and scale-up problems are won by the company that pairs the right material with the right substrate with the right qualification package.

Armadillo Labs is a materials company. We screen and qualify the materials that sit underneath the next compute architecture — superconductors first, but the same screening machinery extends to photonic substrates and cryogenic-compatible composites. The infrastructure investment thesis everyone else is writing assumes the materials are ready. They aren’t. That gap is the work.

---

Footnotes

1. Wade, M. et al., “TeraPHY: A Chiplet Technology for Low-Power, High-Bandwidth In-Package Optical I/O” IEEE Micro 40(2), 63–71 (2020). Also: NVIDIA GTC 2024 photonics keynote materials. ↩ ↩²

2. Cope, P. & Edmund, S., “Water Conservation in Semiconductor Manufacturing” Solid State Technology, 2021 industry briefing. Industry-standard UPW recovery loss figures from SEMI S23-0813 guidelines. ↩

3. Texas Comptroller, Water Use Survey, 2024 data center annex. Forecast: Texas Water Development Board 2027 State Water Plan, datacenter demand chapter. ↩ ↩²

4. Khan, S. & Mann, A. (Stanford / Center for Security and Emerging Technology), “AI Chips: What They Are and Why They Matter” CSET issue brief, 2020. Updated bandwidth-energy ceiling analysis: IEEE Solid-State Circuits Magazine Spring 2024. ↩

5. NVIDIA GTC 2025 keynote, photonics roadmap segment; subsequent investor commentary (Q4 FY2025, Q1 FY2026 calls). ↩

6. GlobalFoundries GF Fotonix platform announcement, 2022; subsequent capacity disclosures, 2024–2025. The other production-volume Si-photonics line is operated by Intel. ↩

7. Holmes, D. S., Ripple, A. L., Manheimer, M. A., “Energy-Efficient Superconducting Computing — Power Budgets and Requirements” IEEE Trans. Appl. Supercond. 23(3), 1701610 (2013). ↩

8. U.S. DOE / ARPA-E REBCO supply chain workshop reports, 2022–2024; commercial production figures aggregated from public disclosures by SuperPower, Faraday Factory, SuNAM, and Shanghai Superconductor Technology. ↩

9. U.S. Bureau of Economic Analysis, Texas private fixed investment in information processing and software, datacenter subcategory, 2019–2024. Cross-checked against EY 2026 Geostrategic Outlook, Texas chapter. ↩