Passive Thermal Computing Infrastructure

The data center
has no building.
The river is the cooling system.

Submera submerges GPU compute hardware directly into natural water bodies using passive thermodynamics to reject heat with zero mechanical cooling. No pumps. No fans. No chillers. Just physics.

See how it works View deployment model
~1.0
PUE target
57°F
validated cooling delta
Zero
moving cooling parts
83
units / site
$26B
cooling market (2025)
The Problem

AI compute is hitting a thermal wall.

Modern GPU racks generate up to 120 kW per rack, six times what air cooling can handle. The industry spends billions of dollars, millions of gallons of freshwater, and acres of real estate just managing heat. Cooling has become the bottleneck of the entire AI infrastructure stack.

120 kW

Per rack, GB200 NVL72

Air cooling physically maxes out at 15–25 kW per rack. The newest GPU systems generate 5–8× that load. Air cooling is not a future option, it is already obsolete at the frontier.

1.58×

Industry average PUE

For every watt going to compute, 0.58 watts are wasted on cooling overhead. At hyperscale, that represents billions of dollars in electricity consumed annually without doing any useful work.

40M gal

Water per year, 10MW facility

Evaporative cooling towers consume tens of millions of gallons of freshwater annually per data center. Regulators and communities in water-stressed regions are pushing back hard.

$26B

Cooling market (2025)

The data center cooling market is growing at 22.3% CAGR, expected to reach $128B by 2033. The industry is not looking to optimize existing cooling. It needs to reinvent it entirely.

The Solution

Passive thermal siphon. Zero moving parts.

Submera places sealed aluminum enclosures containing GPU hardware directly into a natural water body. Heat moves from hardware to dielectric fluid to CNC billet lid to pond water, entirely by the laws of thermodynamics. No mechanical system required at any step.

01

6061-T6 aluminum enclosure

Precision-fabricated, hermetically sealed aluminum housing. Aluminum conducts heat at 167 W/m·K, among the highest of any structural metal. The enclosure wall is the heat exchanger. No additional components needed.

02

EDM250 dielectric fluid

Hardware is bathed in EDM250 dielectric fluid, electrically inert, non-corrosive, and safe for direct contact with all server components. Natural convection circulates the fluid inside the enclosure, carrying heat from GPUs to the walls continuously without any pump.

03

Natural water body as heat sink

A river, lake, reservoir, or ocean at 40–65°F acts as an effectively infinite heat sink. No matter how many watts the GPUs generate, the water continuously absorbs and carries that energy away. Cooling capacity scales with water flow, not with mechanical infrastructure.

04

Armored umbilical to shore

Power and fiber run through a single armored conduit from the submerged frame to shore infrastructure. The deployed node is operationally identical to a rack-mounted GPU server, remotely monitored, fully accessible, invisible to the software stack above it.

05

Eco-friendly by design

Submera is engineered to coexist with the water bodies it occupies. The dielectric is water-soluble and biodegradable in the event of an enclosure breach. No PFAS, no persistent contamination, no oil-style spill response required. Heat rejected to the pond raises only a thin 1–3 ft thermal stratification layer immediately around each enclosure, which dissipates back to ambient within feet of the unit. Every pond is treated as an active ecosystem; unit spacing, depth, and thermal load are tuned so fish, amphibians, and aquatic vegetation continue to thrive alongside the deployed compute.

06

Landowner partnership model

Submera deployments mirror the cellular and radio tower model. Private landowners host the infrastructure on their property in exchange for a flat-rate lease payment, priced by pond surface area. Owners keep the land, keep the pond, and gain a recurring revenue stream from acreage that would otherwise generate nothing. For Submera, it unlocks national-scale deployment without ever owning real estate. For the landowner, it turns a pond into a long-term income asset.

What Submera eliminates

Pumps & CDUs
Server fans
Chillers
Facility HVAC
Cooling towers
Water treatment systems
PFAS / refrigerant exposure
The building itself
1.00
Theoretical PUE, zero cooling overhead
How It Works

From pond to live compute cluster.

A single Submera deployment site uses two ~1.5-acre ponds and two small shoreline power huts. That is the entire physical infrastructure required to bring 83 GPU compute units online. No building, no chillers, no cooling towers.

Step 01

Site survey

Candidate sites are evaluated for water body suitability: depth (8–15 ft target), riverbed or pond-floor stability, and year-round water temperature. Submera targets locations where ambient water stays below 65°F, which covers the vast majority of U.S. inland water bodies.

Step 02

Pond pair preparation

Two oval ponds of approximately 1.5 acres each are selected or constructed on the property. Each pond becomes a self-contained cooling environment, sized to dissipate the heat of roughly half the unit count without thermal saturation, even at full sustained load.

Step 03

Power huts on shore

A small equipment hut sits on a concrete pad beside each pond, labeled PWR-A and PWR-B, housing electrical service, network ingress, and monitoring. Power lines run underground; nothing overhead, nothing visible from a distance. The huts are the only above-grade structures.

Step 04

Unit deployment

83 sealed Submera enclosures are deployed across the two ponds, 42 in Pond A and 41 in Pond B, arranged in a staggered grid pattern that maximizes spacing for thermal exchange. Each unit is fully submerged 1–2 ft below the surface.

Step 05

Buoy & anchor system

Each unit is tethered to a bottom anchor and held at consistent submersion depth by a small float marker at the surface. The system is self-leveling. As pond water levels shift seasonally, every unit tracks the surface automatically with zero mechanical intervention.

Step 06

Online & monitored

Armored umbilicals run from each power hut underground to the units. Within minutes of submersion, GPU junction temperatures reach thermal equilibrium. Validated testing showed 50–57°F temperature reduction and full equilibrium in under one hour at sustained load.

Deployment Diagram

Two ponds. Two power huts. Eighty-three units.

A standard Submera deployment site consists of two ~1.5-acre ponds, each holding roughly half the unit count, with a small power hut on the shore of each pond. All electrical and network conduits run underground. The visible infrastructure footprint is intentionally minimal.

Top-down site diagram · Pond A: 42 units · Pond B: 41 units · 83 units total
Scale-up path

The two-pond, 83-unit configuration is Submera's standard site module. Additional pond pairs can be added in parallel on adjacent acreage, with each new pair adding another ~83 units and one or two additional power huts. Cluster-scale deployments of 250+ units on a single property are achievable with no additional facility construction.

Where It Deploys

Any water body. Any waterfront.

Submera is not tied to a specific geography. Anywhere there is a natural or managed water body with sufficient depth and flow, Submera can deploy. The U.S. alone has over 250,000 rivers and streams and thousands of lakes, representing an enormous untapped deployment surface.

Rivers & streams

Flowing water provides a continuously refreshed heat sink with no recirculation required. Moderate flow rates of 2–4 mph are ideal, enough to exchange heat at the enclosure surface without requiring heavy anchoring. The Blue River corridor in Kansas City is Submera's initial target deployment zone.

flowing heat exchangecontinuous refresh2–4 mph ideal

Lakes & reservoirs

Still water bodies with sufficient depth provide stable thermal stratification. Submera's proof-of-concept validation was conducted at Hamilton Lake Reservoir in January 2025, achieving 57°F temperature reduction at 300W sustained load with zero mechanical cooling assistance.

thermal stratificationstable environmentvalidated Jan 2025

Industrial waterways

Managed channels, cooling ponds, and industrial waterways adjacent to power infrastructure offer ideal co-location opportunities. Proximity to grid power combined with continuous water flow makes industrial waterways highly attractive for high-density GPU deployments.

grid-adjacentmanaged depthhigh density potential

Coastal & international

Ocean-adjacent deployments and international markets in power-constrained regions, particularly across Africa and Southeast Asia, represent significant long-term opportunity. Water-cooled compute at near-zero PUE is especially compelling where grid power is expensive or unreliable.

global applicabilityemerging marketspower-constrained fit
250,000+

Rivers and streams in the United States alone, plus thousands of lakes, reservoirs, and managed waterways. Submera's deployment surface is not constrained by real estate availability. It is constrained only by proximity to water.

Validation Data

Proof of concept. Real numbers.

Submera's V1 system, a Dell PowerEdge R610 in a custom-fabricated sealed aluminum enclosure, was submerged at Hamilton Lake Reservoir in January 2025. The results validated the core passive thermal siphon hypothesis with measured data across all sensor positions.

50–57°F

Temperature reduction across all sensor positions, bulk fluid, enclosure wall, and lid, compared to air-cooled sealed baseline at the same 300W sustained load.

Hamilton Lake, Jan 2025
< 1 hr

Time to reach full thermal equilibrium after submersion. Air-cooled baseline never reached equilibrium, temperatures were still rising when the test was terminated.

300W sustained load
1.00

Effective PUE during submerged operation. Zero mechanical cooling of any kind was used at any point in the test. All heat rejection was passive and thermodynamic.

Zero mechanical assist
Measurement Test 1: Sealed / Air Test 2: Lake Submerged Delta
Ambient temp58°F air40°F water,
Bulk fluid temp110°F (rising ↑)60°F (stable ✓)−50°F
Enclosure wall temp95°F (rising ↑)46°F (stable ✓)−49°F
Lid temp107°F (rising ↑)50°F (stable ✓)−57°F
Thermal equilibriumNOT reached< 1 hour ✓,
Mechanical coolingNoneNoneSame
Competitive Landscape

Every competitor still needs active components.

The immersion cooling market is real and growing, Trane Technologies acquired LiquidStack in February 2026 for ~$85B in enterprise value, validating the category. But every existing solution still requires pumps, CDUs, and mechanical infrastructure. Submera is the only architecture that eliminates them entirely.

Category Submer LiquidStack GRC Direct-to-Chip Submera
Pumps / CDUsYesYes (CDU)Yes (CDU)Yes (CDU)NONE
Fans requiredNoNoNoYES (partial)NONE
Facility coolingCDU + loopCDU + towerCDU + loopCDU + CRACNONE
PUE1.03–1.051.02–1.031.02–1.051.03–1.101.005–1.04
CapEx / kW$3K–$8K$5K–$12K$3K–$6K$1.5K–$4K~$1K–$2K
PFAS exposureNoneHIGHNoneNoneZERO
Water useVariesVariesVariesVariesZERO
Moving partsMultipleMultipleMultipleMultipleZERO
Market signal: Trane Technologies (~$85B enterprise) acquired LiquidStack in February 2026, directly validating the immersion cooling category. Schneider Electric acquired Motivair for ~$850M in October 2024. Samsung acquired Flak Group for $1.5B in May 2025 to enter the space. The largest industrial companies in the world are buying their way into this market. Submera is building the passive alternative that none of them have.
Market Opportunity

A $26B market growing at 22% CAGR.

The data center cooling market is one of the fastest-growing infrastructure segments in the world, driven directly by AI compute density growth. Submera is positioned at the intersection of immersion cooling adoption and passive architecture, a category with no current direct competitors.

T
A
M

Global DC cooling

22.3% CAGR

$26.3B → $128B
2025 → 2033
S
A
M

Liquid cooling (all types)

20.1% CAGR

$6.65B → $29.5B
2025 → 2033
S
O
M

Submera reachable (5-yr)

Years 1–5 build

$4.75M → $62.5M
revenue projection

Market validation signals

Trane Technologies (~$85B) acquiring LiquidStack, February 2026

Schneider Electric acquired Motivair for ~$850M, October 2024

Samsung acquired Flak Group for $1.5B to enter DC cooling, May 2025

AWS launched proprietary liquid cooling for GPU racks, July 2025

Top 4 hyperscalers: $290B+ combined AI/DC CapEx in 2025

AI rack density now 40–120 kW, air cooling physically insufficient at frontier

Revenue Model

Three paths to revenue. One technology.

Submera's business model is sequenced for capital efficiency, the same hardware that validates the technology generates the first revenue, funds the next build, and establishes the track record for product sales and IP licensing at scale.

01
Compute-as-a-Service
Primary revenue stream

Submera operates the deployed units directly and rents GPU compute time to enterprise AI teams, researchers, and inference providers. Because our cooling costs nothing to run, operating expenses are a fraction of traditional data centers, letting us offer competitive pricing while keeping strong margins. Compute revenue funds the next deployment, which funds the one after.

  • Rent GPU compute time to enterprise AI teams, researchers, and inference providers
  • Lower operating costs mean we can undercut market pricing while keeping stronger margins
  • Electricity is the only real ongoing cost, cooling is free
  • Submera retains the compute capacity and the revenue it generates
02
Landowner Partnership Model
Deployment engine for national scale

Submera leases pond surface area from private landowners with lakes, ponds, or large water features on their property. The model mirrors the cellular and radio tower industry: a flat-rate annual payment, priced by surface area, in exchange for hosting infrastructure that the owner never has to operate. This is how Submera scales nationally without ever buying real estate.

  • Flat-rate annual lease, priced per acre of pond surface area
  • Mirrors the proven cellular and radio tower lease model
  • Owner keeps the land, keeps the pond, gains recurring income
  • Submera scales fast on capital we never spend on real estate
Build V2 Rent compute Sign first landowner partnership Replicate site model Multi-site cluster National scale

Compute-as-a-Service generates the revenue. Landowner partnerships unlock the deployment surface. The two together let Submera scale capacity without scaling real-estate cost.

Milestones & Roadmap

18–24 months from funded to commercial.

Submera's execution plan is sequenced to reach GPU-class thermal validation and first compute revenue within 12 months of funding close, then scale into multi-site deployments through landowner partnerships during the second year.

Phase 1

Hardware & Build

Months 1–3

Procure 4× H100 PCIe GPUs and CNC tooling. Fabricate V2 enclosure with precision thermal interface. Seal and fill with dielectric fluid.

Phase 2

Test & Validate

Months 4–9

Deploy V2 in a candidate pond. Run thermal validation at 1.5–2.5 kW GPU-class load. Collect 90 days of production temperature data across the full unit array.

Phase 3

Commercial Revenue

Months 9–15

V2 transitions from validation to production. First compute rental revenue. Scale the initial pond toward full unit count. Refine deployment playbook for replication.

Phase 4

Partnership Scale

Months 15–24

Sign first landowner partnership leases. Deploy a second pond pair on private acreage under the lease model. Build the multi-site pipeline that defines the next phase of growth.

M1

Funding close, hardware procurement begins immediately

M3

V2 enclosure assembled, sealed, and fluid-filled

M4

First pond deployment, thermal testing begins

M9

Thermal validation complete at 1.5–2.5 kW GPU load

M10

First compute revenue, V2 transitions to production

M14

First landowner partnership lease signed

M18

Second pond pair deployed on partnered acreage

M24

Multi-site cluster operational, partnership pipeline active

Get In Touch

Ready to learn more?

Submera is actively working to reshape how the world thinks about data centers, moving compute out of massive, power-hungry buildings and into the natural environment around us. Efficient by design. Out of sight by nature. If that mission resonates with you, we'd love to connect.

Send an email Read our Substack