IP-safe project evidence study / R2 V2 / 2 November 2026
For the same customer workload and service quality, when would floating compute outperform the best feasible land alternative? The investigation challenged the appealing premise that nearby cold water alone creates a business case. It developed a general monograph and a desk comparison of Casco Bay / Portland and Kempenfelt Bay / Barrie. [S1, S2]
| Architecture | Potential advantage | Obligations to investigate |
|---|---|---|
| Floating water-cooled | Modular construction and limited land footprint | Hull, mooring, marine access, cable landing and removal |
| Land water-cooled | Water cooling with conventional site access | Land, intake/outfall, environmental permissions |
| Land air / closed-loop | Established siting and maintenance routes | Power, heat rejection, land and ambient extremes |
Use the same IT workload, accepted results, availability, latency requirements and commissioning assumptions. Compare with no build or defer as appropriate. Separate geographically distributed modules are not automatically a low-latency GPU cluster.
The named areas are research selections, not available parcels, approved berths or ranked investments. The source study did not contact utilities, ports, regulators, landowners or customers. Firm power, diverse fiber, actual rights, costs and ecological acceptance remained unknown. That limits the conclusion to screening and a plan for decisive information.
| Input | Illustrative value |
|---|---|
| IT load | 1 MW |
| Land-air facility load | 1.35 MW |
| Floating facility load | 1.20 MW |
| Operating hours | 8,760 / year |
| Energy price | 100 currency units / MWh |
| Demand charge | 15 currency units / kW-month |
| Extra floating capital | 1,000,000 currency units |
| Recovery period / rate | 10 years / 10% |
| Extra annual marine operating cost | 150,000 currency units |
Load difference = 0.15 MW = 150 kW. Energy saving = 0.15 x 8,760 x 100 = 131,400 per year. Demand saving = 150 x 15 x 12 = 27,000. Total assumed utility saving = 158,400.
Capital-recovery factor = r(1+r)^n / ((1+r)^n - 1) = 0.1627454. Annual capital premium = 162,745.4. Add 150,000 operating premium to obtain 312,745.4. The annual deficit is 154,345.4 currency units. [S2]
An illustrative land-water load of 1.15 MW would use still less electricity than floating at 1.20 MW. Floating would then need another verified advantage: earlier service, lower usable-site cost or a suitable industrial berth. None is established by this screen. All values are assumptions, not Maine/Ontario tariffs or supplier quotations.
| Heat-transfer coefficient fraction | Temperature approach | Heat capacity | IT ceiling |
|---|---|---|---|
| 1.00 | 5 K | 1.200 MW | 1.000 MW |
| 1.00 | 3 K | 0.720 MW | 0.564 MW |
| 0.70 | 5 K | 0.840 MW | 0.673 MW |
| 0.70 | 3 K | 0.504 MW | 0.367 MW |
With exchanger area 240 m2 and reference coefficient 1,000 W/m2/K, Q = U x A x temperature approach. At 5 K, Q = 1.2 MW. The stress model assumes facility heat = 1.10 x IT load + 0.10 MW. Combined coefficient degradation and warmer intake reduce the modeled IT ceiling to (0.504 - 0.10) / 1.10 = 0.367 MW. These four corners are sensitivities, not event probabilities. [S1, section 13]
With assumed thermal capacitance 50 MJ/K, 5 K headroom and 1.2 MW heat input, zero heat rejection reaches the lumped limit in 50 x 5 / 1.2 = 208 seconds. A battery that supports hours of electrical load does not extend that thermal limit. Real local hot spots and control delays require engineering qualification.
Consider two one-hour slots, a 1 MW compute limit, an empty 1 MWh battery and a deadline after slot two. Supply [2,0] MW serves 2 MWh of work; supply [0,2] serves 1 MWh and leaves 1 MWh stored after the deadline. Total energy is identical. Temporal accessibility changes useful output. Extending the deadline can remove the difference. The example transfers a finite-resource insight without importing geological validation claims.
| Question | Casco / Portland | Kempenfelt / Barrie |
|---|---|---|
| Water evidence | Published shallow harbor monitoring | Depth-dependent lake temperature and oxygen monitoring |
| Hazards to assess | Storm surge, waves, corrosion and access | Ice, stratification, internal waves and access |
| Environmental questions | Intake/outfall and marine habitat | Source-water protection and lake ecology |
| Common missing facts | Firm power, fiber, rights, cost | Firm power, fiber, rights, cost |
A shallow station or monthly mean does not establish extreme temperature at a proposed intake. The source study did not download and quality-control a paired depth-by-time dataset for the two areas. No local P95/P99 cooling capacity or downtime comparison is supported. These are source-study screening observations, not independently refreshed site conditions. [S2]
The monograph declares a 14,400 kWh job budget: 2,400 consumed + 4,000 reserved + 8,000 available. A requested 9,000 reservation exceeds availability. Counting the reserved 4,000 as available would break the account. Selected local helper execution reported these deliberately constructed violations. This ledger is not a physical battery measurement. [S1, Appendix D]
The reported helper demonstration retained two compatible supplied histories for a lost acknowledgement and an unknown alternative. Completion of the comparison did not resolve whether the job executed. This shows why replay needs reconciliation. The local patterns were synthetic; no marine scheduler or native enforcement operated.
The R2 monograph moves from facility electricity alone to accepted work. Its screening denominator is W = 8,760 x availability x utilization x accepted-output fraction. That product assumes separable factors. Correlated outages, deadlines and checkpoint losses need chronological simulation. A more favorable cooling ratio can coexist with worse cost per delivered result. [S1, section 15]
The source discusses no research-extension/native execution while later reporting selected local mathematical-helper calls. This document distinguishes those components explicitly: selected local advisory calculations are source-reported; no full-system runtime, native review service, marine controller or independent engineering review is claimed.
Obtain an actual site and binding utility/fiber information; specify the workload and service contract; acquire marine, cooling and ecological SME input; and price both water-cooled architectures against the credible land baseline. The present study is broad screening. Actual site information enables deeper engineering rather than simply more candidate locations.
[S1] Compute_at_Sea_HSDLG_V2_Monograph_R2, sections 12-16 and Appendix D.
[S2] Compute_at_Sea_Casco_Kempenfelt_Feasibility_R1, sections on architectures, thermal screening and economics. Both supplied project documents, dated November 2026.
The economic and thermal arithmetic in this public document was recalculated from declared inputs. Existing helper executions were not rerun. Public source navigation in S2 includes NOAA Portland station 8418150 / CASM1 and Ontario Lake Simcoe monitoring. No site engineering suitability finding follows from those references.
Detailed assumptions, calculation records, standards and selected execution evidence are available for scoped review under an appropriate NDA or contract.