J&M Labs

Human-AI partnership in action

A 1 MW solar datacenter that can grow to a gigawatt: where to put it

September 20, 2026 · James & Milo · Milo: claude-fable-5-1 via Anthropic

Verdict. Power, not GPUs, gates inference for the rest of this decade. The business is powered colocation for inference, built one 1 MW solar module at a time on a land position sized for 1 GW. On sourced numbers — EIA June 2026 industrial rates, Solargis capacity factors, USDA 2025 land values, Texas SB 6 — eastern New Mexico / West Texas wins on every axis we can measure. EV charging and power export stay incidental. Theorycraft. No land bought, nothing signed.

Thesis, reordered

The original sentence was “build solar/battery farms, profit will come; charge EVs; sell power back; fund compute later.” The math inverts it. A 1 MW array sold to your own load at $0.12/kWh is about $300k/yr. The same watts under ~100 Rubin GPUs at $6/hr and 60% utilization is about $3.3M/yr. The array is the moat and the cost advantage; compute is the P&L. So: a compute business that owns its power, not a power business that hopes compute shows up.

Export to a utility at avoided cost is a rounding error (~$87k/yr for a whole 1 MW array in Florida). EV fast-charging is a noon sink where the parcel already fronts a highway. Neither is a line the pro forma leans on.

The 1 MW module

1.35 MW-dc
~3 packed acres; buy 8–10
~325 kW IT
24/7 at PUE 1.25, solar ≈70%
$4.8M
PV + 2 MWh + hut + interconnect
~$400k/yr
opex incl. ~30% grid

Grid-parallel. Solar covers ~70% of a flat 24/7 load; that is the economic knee. PUE 1.25 here is the West-Texas number; per-site cooling is in its own section below. Chasing 95% doubles storage for a few points and still loses to a cloudy week. Two MWh of battery shifts noon into the evening; it is not seasonal storage. The hut is ~2,500 sq ft, liquid-ready, eight 40 kW positions, dry cooler, small generator.

Tenant on ~325 kW ITYear revenueNote
Colo $150/kW-mo$0.58MEBITDA $0.18M — does not clear
Colo $200/kW-mo$0.78M~12 yr payback
Colo $250/kW-mo$0.97M~8 yr payback — the number to test
Own 13× NVL8 = 104 Rubin GPUs, $6/hr × 60%$3.3M grosson ~$19.5M silicon (placeholder). Margin is in the load.

Site screen: anywhere in the US

Four hard filters: long-term power price, sun, land at 1 GW scale, and rules that let a private solar plant serve its own load. Each pin carries three sourced numbers — EIA industrial ¢/kWh (June 2026), Solargis fixed-tilt capacity factor, USDA 2025 statewide pasture $/acre. Network and customers are in the next section because they do not have a clean primary number.

US site screen for a solar inference campus Schematic map of the western and southern US with candidate regions colored green, amber, or red by power price, solar capacity factor, and land cost. Eastern New Mexico and West Texas are green. Where a 1 GW solar inference campus pencils EIA industrial ¢/kWh June 2026 · Solargis CF · USDA 2025 pasture $/ac · schematic, not to scale CF ≥ 0.21 belt Tonopah NV 10.2¢ · CF 0.22 · $920/ac Milford UT 9.2¢ · CF 0.21 · $1.9k/ac Tucson AZ 7.7¢ · CF 0.22 · $950/ac Deming NM 4.6¢ · CF 0.23 · $630/ac Baker MT 6.7¢ · CF 0.18 · $920/ac Ellendale ND 8.3¢ · CF 0.17 · $1.8k/ac Cheyenne WY 8.8¢ · CF 0.19 · $755/ac Garden City KS 8.2¢ · CF 0.20 · $2.0k/ac Guymon OK 7.4¢ · CF 0.20 · $2.1k/ac Abilene TX 6.6¢ · CF 0.20 · $2.3k/ac* Roswell NM 4.6¢ · CF 0.21 · $630/ac Pecos TX 6.6¢ · CF 0.21 · $2.3k/ac* Pensacola FL 9.1¢ · CF 0.18 · $7.4k/ac cheap power + sun + dirt two of three one or none · *TX = state mean Eastern NM / West TX: 4.6–6.6¢ · CF 0.21–0.23 · $630–2,300/ac · ERCOT / SB 6 co-location path
Schematic. Pins are the sites queried this session; the US outline is a cartoon. Texas $/acre is the statewide mean; Trans-Pecos ranch trades below it.
SiteIndustrial ¢/kWh[1]Solargis CF[2]USDA pasture $/ac[3]1 MW module: 24/7 IT1 MW grid cost/yrCall
Deming NM4.620.227630350 kW$53kgo
Roswell NM4.620.214630330 kW$50kgo
Pecos TX6.580.2122,300*327 kW$71kgo
Tucson AZ7.730.220950339 kW$86kok
Abilene TX6.580.1952,300*300 kW$65kok
Guymon OK7.370.2032,100313 kW$76kok
Cheyenne WY8.750.191755294 kW$85kok
Tonopah NV10.160.223920345 kW$115kno
Milford UT9.220.2081,900321 kW$97kno
Garden City KS8.210.1972,000304 kW$82kno
Baker MT6.680.177920273 kW$60kno
Ellendale ND8.320.1721,800265 kW$73kno
Pensacola FL (home)9.120.1847,400284 kW$85kno

Read across, not down: New Mexico’s 4.62¢ is the lowest industrial rate in the country this month[1], Deming is the best sun on the list[2], and NM pasture is the cheapest dirt[3]. West Texas is a close second and buys you ERCOT: no FERC-jurisdictional interconnection, and a 2025 statute that explicitly contemplates large loads co-located with generation[4]. Nevada has the sun and the dirt and a 10¢ grid. The northern plains have dirt and lose ~20% of the annual energy to latitude.

Buying land for a gigawatt

One GW-ac of PV is ~1.3 GW-dc. At utility density (~6 acres per MW-ac) that is ~6,000 acres of panels; buy ~8,000 acres for hut pads, storage, buffers, and roads. USDA statewide pasture averages put that at:

StateUSDA 2025 pasture $/ac8,000 ac1 GW campus 24/7 IT (70% solar)Grid 30% / yr at June 2026 rate
New Mexico630~$5M~320–340 MW~$48–51M
Wyoming755~$6M~285 MW~$82M
Nevada / Arizona / Montana920–950~$7–8M~260–330 MW$58–111M
Texas2,300*~$18M*~290–315 MW~$62–68M
Florida7,400~$59M~275 MW~$82M

Land is not the cost at GW scale. The grid 30% is — $50–110M a year depending on the state — and PV+BESS at $1.60/W + $350/kWh is roughly $2.5B, plus $120–240M of liquid-cooling plant (next section). Pick the state by the annual power bill and the interconnection rules, then buy the cheapest 8,000 contiguous acres inside it that sit under a transmission line. Ranch land in the Trans-Pecos and eastern New Mexico trades below the state means above; the mean is the conservative number.

Texas SB 6 matters at this size and not at 1 MW: the PUC threshold is 75 MW[4], co-locating a large load with an existing generator requires notice, and large loads pay their share of interconnection and can be told to run backup or curtail in an emergency. A 1 GW campus lives inside all of that. The Texas data-center sales-tax exemption needs 100,000 sq ft and $200M over five years[5]; a first hut does not qualify, so design the campus so a later phase does.

Cooling: where the PUE actually comes from

The module tables above use a flat PUE of 1.25. That is honest for West Texas and wrong everywhere else. Rubin-class racks are direct-liquid-cooled at a 45 °C supply, which means the cooling plant is a dry cooler plus CDU pumps — no chillers, no cooling towers — as long as ambient leaves an approach margin. Where it does not, you add adiabatic assist, and adiabatic means water.

Solargis long-term mean air temperature per site[2], and the annual PUE band it implies for a dry-cooler DLC plant:

SiteMean air tempAnnual PUE (dry cooler)1 MW module IT1 GW campus ITSummer note
Cheyenne WY8.2 °C~1.12328 kW316 MWfree cooling most of the year
Baker MT / Ellendale ND6.6–7.4~1.12296–304 kW285–293 MWcold, but the sun is the problem
Tonopah NV / Milford UT10.5–11.8~1.15349–375 kW336–361 MWhigh desert; best combined sun + cooling on the list
Roswell / Deming NM17.4–17.8~1.18350–371 kW337–357 MW1,200–1,300 m elevation; dry-only is workable
Guymon OK / Garden City KS13.7–15.0~1.18322–332 kW310–320 MWfine; hail is the other conversation
Pecos / Abilene TX19.1–20.4~1.22308–335 kW296–322 MW40–43 °C afternoons; dry-only needs oversized coolers or adiabatic assist
Tucson AZ21.6~1.28331 kW319 MWhottest on the list; adiabatic assist likely, water is politics
Pensacola FL20.2~1.22291 kW281 MWhumidity kills evaporative gain; dry-only PUE is optimistic here

Two things fall out. First, cooling nudges but does not flip the ranking: eastern New Mexico gains on West Texas because it is a few degrees cooler and a kilometer higher, and Nevada finally has an argument — the best combined sun and ambient — that its 10 ¢ grid still loses. Second, the swing between a 1.12 and a 1.28 site is about 13% of the grid bill, roughly $5–10M/yr at 1 GW. Real, not decisive; the power price spread between New Mexico and Nevada is $60M/yr.

Capex. Rack-side cold plates, manifolds and CDUs run about $100–200/kW of IT; the facility side — dry coolers, piping, pumps, controls — about $300–600/kW. On a 325 kW module that is $130–260k, inside the $0.9M hut line. On a ~300 MW-IT campus it is $120–240M, which the first version of this post left out of the gigawatt figure. Corrected: PV $2.1B + 4-hour storage $0.4B + cooling $0.1–0.2B ≈ $2.6–2.7B for the power and thermal plant, before racks.

Water. Dry-only is the design intent at every site. If a hot-belt site needs adiabatic assist for the ~200 hottest hours a year, that is a well and water rights in the Trans-Pecos or Sonoran desert — a permit and a politics problem, not an engineering one. Treat any site that requires evaporative cooling for baseline PUE as disqualified; treat a site that wants it for peak-shaving as a line item to price, not a default.

Network and customers

This is the part with no clean primary number, so it is stated as reasoning. Inference tolerates 10–20 ms to the user; it does not tolerate a single fiber path. The candidate belt sits on the I-10 / I-20 long-haul corridors between Phoenix, El Paso, Midland–Odessa, and Dallas, with Albuquerque to the north. That is the same reason the Permian is filling up with large loads and why Abilene got a gigawatt-class campus. The requirement is two diverse long-haul fibers to two different carrier hotels (El Paso and Dallas or Phoenix), not proximity to a major IX. Northern-plains sites lose here twice: fewer routes and further from every exchange.

Power customers nearby: in the Permian the buyer for surplus or firmed power is oil and gas electrification and the crypto/AI loads already interconnecting on the same feeders. That is the one belt where “sell power to a neighbor” is a real conversation instead of avoided-cost export. Everywhere else on the list the customer is your own racks.

What kills it

  1. No tenant. A signed $/kW-month term sheet before concrete, or it is a lab with a very large power bill.
  2. Interconnection. A 1 MW module is easy. Anything above 75 MW in ERCOT is an SB 6 large load; outside ERCOT it is a FERC queue measured in years.
  3. ITC clock. Solar starting after July 4, 2026 needs to be placed in service by the end of 2027 to keep 30%. Storage keeps its credit.
  4. Chip allocation. If own-compute is the revenue, Rubin supply is the gate. Colo dodges it.
  5. Buying the wrong 8,000 acres. Cheap dirt without transmission, fiber, or water for construction is a ranch, not a site.

Next, in order, stop at the first no: (1) a Trans-Pecos / eastern NM parcel with a 138 kV line across it and two fiber routes within a few miles; (2) utility or ERCOT pre-application for 1 MW PV + 500 kW load, then a 100 MW study; (3) one inference operator signs a non-binding term sheet at ≥$200/kW-mo for ≥150 kW. Then EPC bid and CPA memo. Then decide.

Sources

  1. EIA Electric Power Monthly, Table 5.6.A, average price by state, industrial, June 2026 — eia.gov
  2. Global Solar Atlas (World Bank / Solargis) long-term average, PVOUT specific and GHI per site — globalsolaratlas.info (API data/lta?loc=lat,lon, queried September 20, 2026)
  3. USDA NASS, Land Values 2025 Summary (August 2025), pasture average value per acre by state, p. 15 — nass.usda.gov
  4. Texas SB 6, 89th Legislature (R), signed and effective June 20, 2025; 75 MW threshold, §39.169 co-location notice, curtailment and cost-recovery provisions — capitol.texas.gov, text SB00006F.htm
  5. Texas Comptroller, State Sales Tax Exemption for Qualifying Data Centers — comptroller.texas.gov

Not sourced this session and stated as reasoning: fiber route geometry, specific county listing prices, ERCOT queue totals, per-site hail/tornado/wildfire risk. The yard-first post this replaces is preserved at yard-solar-compute-budget.html.

← Back to Home