01
Convert capacity into compute
AI growth is limited by power delivered to servers and heat removed — not by nameplate generation alone.
More Compute. Naturally.
Komorebi · (ko-moh-ray-bee)
More AI compute. On the PV array you already have.
Most of the field keeps making power. A slice of it takes the heat. Those freed megawatts go to GPUs, not chillers.
A cooling overlay on a plant that already exports — not a new power plant, and not a new interconnect. Cooling tables on a slice of the field. The rest stays PV. Dry coolers remain as backup.
- Existing solar plant. Cooling tables on a slice of the field. The rest keeps making power.
- 50 MW of firm computers on the connection you already have.
- Cooling uses 8.5 MW less than chillers. Dry coolers stay as backup.
The worked case is 50 MW of firm IT on an existing 100 MWdc PV farm. Cooling uses 8.5 MW less than chillers, so the meter reads 54 MW, not 62.5.
A cooling overlay on the fenceline you already have. Closing 50 MW of firm IT is $36–39 million — $0.72–0.78M per MW — against $150 million per year of planning offtake. Two comparisons: an already-connected solar deed becomes a 50 MW IT host this year; 8.5 MW less parasitic load versus PUE 1.25 at the same 50 MW. Original dry coolers stay as fail-safe.
Phase 1
Existing large PV farms in leading economies
~91 GW
AI compute these farms’ dual-use can host
Hybrid radiator tables and a heat-pump cold bank. 50 MW of design-target 99.9% firm AI per 100 MWdc.
Phase 2 / 3
Same land
~116 GW
Optional on the same land
TES/ORC and recovered-row CSP after Phase 1. Engines last. Not in the Phase 1 offtake. Peak / flexible ~149 GW. Qualifying plants, not signed contracts.
Pecos planning case
The worked case is 50 MW of firm IT on a 100 MWdc farm
100 MWdc legacy farm / 50 MW firm IT / 8.5 MW parasitic cut / 500-acre 2015 fence. Phase 1 books 50 MW at the interconnect you already have. TES/ORC and extra halls are later, on the same fence.
50 MW firm IT. 8.5 MW parasitic cut (12.5 → 4 MW). Meter 54, not 62.5.
Added this phase8.5 MW parasitic recoveryCumulative additionalY-axis: MW
Pecos, Texas planning weather
Comparative parasitic load across the year
Monthly mean MW. Lower is better. Cooling parasitics at each architecture on the same 50 MW hall. Shape from 2024 NSRDB for Pecos; means scaled to 12.5 MW at PUE 1.25 and 4 MW at Komorebi Phase 1.
- Komorebi Phase 14 MW · 8% · PUE 1.08
- Dry cooler + adiabatic9 MW · 18% · PUE 1.18
- Dry cooler + chiller12.5 MW · 25% · PUE 1.25
Not gigawatts. The source files labeled monthly-mean MW as GWe. Scaled so the chiller line is the 12.5 MW PUE 1.25 block on a 50 MW hall. Komorebi is Phase 1 only — 50 acres of radiator tables — not later recovery.
100%
More AI compute per 100 MWdc — inclusive of the 8.5 MW parasitic cut. Weather and soils set the acres.
Existing
At the same interconnect
50 MW
Pecos planning case baseline
474 sites
U.S. existing-PV screen
Named case: 50 MW firm IT on 100 MWdc, inclusive of 8.5 MW parasitic recovery. Weather and soils set the acres. Pecos planning case — 100 MWdc legacy farm / 50 MW firm IT / 8.5 MW parasitic cut / 500-acre 2015 fence.
Three insights
Three insights. Four constraints. One platform.
The bottleneck in hyperscale AI is no longer compute demand. It is time-to-energy.
02
Integrate the resources
Cooling, recovered heat, CSP, thermal storage, and ORC reinforce one another as one system.
03
Stage the investment
Add each capability only when its incremental compute value exceeds its incremental cost.
Four constraints. Infrastructure has not caught up.
Growth is constrained by the infrastructure required to support it — not by demand for servers.
Electrical
Limited interconnect capacity and substation availability delay or cap growth.
Fuel
Additional natural-gas supply, infrastructure, and permitting are slow and uncertain.
Groundwater
Cooling towers withdraw and evaporate water, creating permitting and environmental limits.
Time
New infrastructure takes years and billions — stranding racks and delaying revenue.
Why capacity is left on the table
Conventional architecture rejects the energy that could run more GPUs.
Conventional architecture
Nearly all IT electrical energy ultimately becomes heat.
Fuel and generation
Scarce interconnect and turbines are spent on cooling and support loads as well as GPUs.
Rejected heat
In high-temperature liquid-cooled configurations, a large fraction of that heat can be captured in the liquid loop. Komorebi models a 70°C source-loop condition — a design case, not an industry universal.
Groundwater loss
Towers withdraw and evaporate water, creating permitting and environmental limits year after year.
Komorebi architecture
Phase 1 removes heat. Recovery comes later — after uptime is proven.
Remove heat first
Radiator tables on extra A-frames under cooler rows only. Remaining PV stays generation. Phase 1 is simpler and faster to implement.
Recover heat later
After uptime isolation, capture AI-server heat, then heat from qualified generation — each asset only when it pays.
Enable compute
Return electrical headroom to IT load — without requiring an equivalent new grid interconnection or material groundwater withdrawal. Later thermal recovery can use heat from qualified existing generation.
The product
More compute from the power, heat, and land already there.
Dual-purpose an existing PV farm as 24/7 cooling, then drop in modular AI. The hyperscaler is buying additional compute on an existing generation interconnect — not an equivalent new transmission line. Heat recovery is later. Phase 1 is the hall.
Existing PV farm
Existing land, interconnection, and 100 MWdc+ PV. Dual-purpose the array as a 24/7 cooling asset. Modular compute plus the same thermal plant. Optional CSP, TES, and ORC. Then qualified heat recovery. No field expansion.
Constrained AI campus
Existing compute and power envelope. Needs adjacent land — ideally with PV already installed. Ambient-air, buried-loop, and cold-storage cooling first. Additional compute on the existing generation interconnect next. Optional AI heat recovery after uptime isolation. Qualified generator recovery last.
Architecture
Remove heat first. Recover heat later.
Each layer is independently valuable. Phase 1 is the product — cooling, simpler and faster. Phase 2 is the filing and the plumbing. Phase 3 is the residual plant. No later stage may degrade rack cooling. Thermal recovery may be unavailable. Rack cooling may not.
01
Remove
Radiator tables and a cold bank cut parasitics. Phase 1 — simpler and faster.
02
Recover
After uptime isolation, capture AI-server heat, then heat from qualified generation when each asset pays.
03
Upgrade
Lift temperature with heat pumps and CSP only when incremental compute value exceeds cost.
04
Enable
Free electrical headroom for additional AI compute inside the same envelope.
Reliability principle
Thermal recovery may be unavailable. Rack cooling may not.
Opportunity set
Additional compute. Existing generation interconnect.
- Candidate U.S. PV sites
- 474
- Phase 1 AI this set can host
- ~21.8 GW
- 99.9% firm AI
- $3M / MW-yr
Sequence
Cooling first. Heat later. Same fence.
Start the hall. Put cooling on a slice of the tables you already have. Recover heat only after that hall is up. Foundations, optics, and offtake sit in the briefing — under NDA.
Phase 0
Existing site
Months, not years
- Modular halls on land you already have.
- Cooling that can run before the overlay is finished.
- Uptime is the first commercial gate.
- Farm or campus — same order.
Workloads start while cooling is built.
Phase 1
Cooling
6–12 months
- Radiator tables on a slice of the field. The rest stays PV.
- Dry coolers remain as fail-safe.
- Same fence. Same permit class.
- This is the product.
More IT at the interconnect you already have.
Phase 2
AI heat recovery
After the hall is up
- Server heat first.
- Leftover acres take CSP.
- Cooling does not wait on this step.
- Optional — after uptime.
More compute from heat the site already makes.
Phase 3
Engine heat recovery
After heat is booked
- Recovered heat runs more halls — not a bigger engine yard.
- Same fence.
- Engines last.
- Still one overlay, not a new power plant.
Same plant. More IT.
Platform
Four levers on the same land.
Cooling-first thermal infrastructure
Radiator tables, a cold bank, and low-water heat rejection cut parasitic load and raise compute density per available megawatt.
Fixed-footprint densification
On existing PV farms: repower legacy polycrystalline, monofacial arrays with higher-density bifacial modules on reflective foundations, recover aperture for Komorebi CSP, and add duration without expanding the site.
Duration overlay
Optional NPI 4-lens, thermal energy storage, and ORC convert leftover acres into 14 MWe — on the same radiator loop. Dry cooler is fail-safe only, not the Phase 2 sink.
Module recovery
On existing PV farms: on-site triage, refurbishment, and redeployment with Revive PV turns aging modules from a repowering liability into a recoverable asset.
Start with an existing PV farm.
Start with an existing PV farm of the right age. Request a solution briefing. GPS, weather, generation, and soils are shared after NDA.