A first-principles engineering analysis of the system, subsystem by subsystem. The numbers are concept-study estimates, not a final design package; ranges are shown where the inputs are uncertain. Each section states its assumptions, derives a result, and where possible checks it against a known engineering system. The weakest links are flagged explicitly rather than buried.

A. The Demand — Power Per Frontier Run

The power a single training run draws is its compute divided by the efficiency of the hardware and the duration of the run:

$\boxed{\,P \approx \frac{C}{(\text{FLOP/W}) \times t}\,}$

The growth rate of $P$ is therefore the growth rate of frontier compute divided by the growth rate of efficiency. Both have decade-long trends:^[1][2][3]

Quantity	Annual growth	Basis
Frontier training compute	~4–5×	Epoch AI, decade trend
Energy efficiency (FLOP/W)	~1.3–1.5×, slowing	Hardware specs; Dennard scaling ended
Power per run	~1.5–3×	Compute ÷ efficiency

Projecting a ~0.3 GW 2025 frontier cluster forward at the two ends of the compute range:

Year	Aggressive (compute 4×/yr)	Decelerating (compute 2×/yr)
2028	~7 GW	~1.2 GW
2030	~50 GW	~2.5 GW
2035	unphysical (>1 TW)	~10–15 GW

The aggressive column breaks before 2035, which is the point: scaling cannot hold 4×/yr indefinitely because power runs out first. Across every non-absurd path, a single frontier run reaches gigawatts to tens of gigawatts within 5–10 years.^[35]

For scale, total United States generating capacity is ~1,280 GW,^[6] and data-centre load was ~35 GW in 2024, projected to ~80–130 GW by 2030.^[36][7] A single frontier run of tens of gigawatts is a national-scale load behind one fence, which is why it cannot sit on ordinary land near existing transmission. The efficiency denominator is also weakening: with Dennard scaling ended and precision approaching a floor of a few bits, FLOP/W gains increasingly come from packaging rather than transistor physics, so the historical ~1.4×/yr that partly offset compute growth is fading.^[3][4][5]

B. The Pod — Compute, Power Density, Volume

A current frontier rack integrates about 72 top-end accelerators with their CPUs, high-bandwidth memory, liquid cooling, and a short coherent network, at roughly 140 kW per rack:^[8][37]

$N_\text{racks} = \frac{10\,\text{MW}}{140\,\text{kW}} \approx 71 \quad\Rightarrow\quad N_\text{GPU} \approx 71 \times 72 \approx 5{,}100$

Allowing for power electronics, switching, CPUs, and circulation overhead, the conservative count is 5,000–6,500 accelerators per pod. The reference public cluster used for comparison throughout is a 64-rack, 4,608-GPU deployment, so one pod is approximately one such cluster.^[10]

Volume and power density. Stripped of the aisles, plenums, fan walls, and human-access clearances of a land hall, the pod is a horizontal cylindrical vessel roughly 14–16 m long and 4–5 m in diameter:

$V = \pi r^2 L \approx \pi (2.25)^2 (15) \approx 240\,\text{m}^3$

Power density is then $10\,\text{MW} / 240\,\text{m}^3 \approx 42\,\text{kW/m}^3$ at the vessel envelope, rising to 100–150 kW/m³ at the populated-board level. A conventional air-cooled hall runs 1–2 kW/m³ all-in. The pod is one to two orders of magnitude denser, which is the physical basis for the lattice compaction in §D — and is only possible because the heat leaves through fluid, not air.^[38][39]

C. Coherence — The All-Reduce and What Actually Binds

This is the load-bearing question for the topology claim, so it gets the most careful treatment. Synchronous training repeats a collective operation — the all-reduce — on every step.^[23][24][40] For a ring all-reduce of a gradient of size $S$ across $N$ endpoints, each with link bandwidth $B$ and per-hop latency $\alpha$:

$$\boxed{ \,T_\text{AR} = \underbrace{2(N-1)\,\alpha}_{\text{latency term}} \;+\; \underbrace{\frac{2(N-1)}{N}\,\cdot\,\frac{S}{B}}_{\text{bandwidth term}} \,}$$

The two terms behave completely differently, and only one of them carries distance.

The latency term carries distance; the bandwidth term does not. The per-hop latency decomposes as $\alpha = t_\text{switch} + t_\text{prop}$, where switch-ASIC latency $t_\text{switch} \approx 0.1\text{–}0.3\,\mu\text{s}$ and propagation $t_\text{prop}$ is the distance term: light covers about one metre every 5 ns in fibre. The bandwidth term contains no length at all — it is set by the optical fabric, not by how far apart the endpoints sit.

Topology	Span	$t_\text{prop}$/hop	$\alpha$ (incl. ~0.2 µs switch)
10 MW pod	12–16 m	0.06–0.08 µs	~0.3 µs
1 GW dense block	70–90 m	0.35–0.45 µs	~0.5–0.6 µs
10 GW campus	140–180 m	0.7–0.9 µs	~0.9–1.1 µs
Land multi-building campus	500–2,000+ m	2.5–10+ µs	~2.7–10+ µs

The honest conclusion. Submerging the machine collapses the propagation component of $\alpha$ by 10–100× relative to a sprawling land campus. That is decisive for latency-bound collectives — fine-grained tensor-parallel exchanges, small messages, and the long tail of synchronisation barriers where $T_\text{AR}$ is dominated by the latency term. It does nothing for bandwidth-bound collectives, where the bulk gradient all-reduce is governed by $S/B$ and the binding constraint is fabric bisection bandwidth — a design choice independent of whether the machine is wet or dry.^[25] The subsea advantage is real, but it is an advantage on the latency term specifically. A campaign that wins frontier training on this architecture wins it on thermal density and short-path latency together, not on propagation alone. The bandwidth fabric still has to be built, and it is the same problem underwater as on land.

This is also why the architecture survives a shift to loosely-coupled distributed training: that shift attacks the latency term (by tolerating it), not the power-and-cooling problem each site still faces.^[21][22]

Serial depth is the opposite case, and the one that favours the lattice most. Test-time reasoning unrolls $K$ sequential steps, each waiting on the one before it, so the run accumulates a wall-clock latency that grows with $K\,\alpha$. At land spans $\alpha$ is microseconds; in the lattice it is a fraction of one. For shallow chains the gap is noise. For the deep recursion the frontier is turning toward, $K$ is large, and the 10–100× advantage on the propagation component of $\alpha$ multiplies by $K$ into the difference between a run that finishes in usable time and one that does not. Here short-path latency stops being an efficiency and becomes an enabler — the binding dimension the body’s §VII rests on.

Mixture-of-experts sharpens this. Routing each token to a few of many experts is an all-to-all exchange across the devices holding them — the most topology-sensitive collective there is.^[26] On training batches it is bandwidth-bound and the lattice helps little. At decode, one token at a time, the messages are small and it turns latency-bound. A master that is both enormous and recursive pays that latency on every expert hop of every step — the workload that binds hardest to a short-path lattice.

D. Geometry — The Three-Dimensional Lattice

A flat campus packs modules across a plane, so its diameter scales as $D \propto \sqrt{N}$. A dense subsea lattice packs them through a volume, so $D \propto N^{1/3}$. The geometric advantage:

$\boxed{\,\frac{D_\text{2D}}{D_\text{3D}} = \frac{\sqrt{N}}{N^{1/3}} = N^{1/6}\,}$

At $N = 10{,}000$ pods, $N^{1/6} \approx 4.6$. This is the packing advantage before overhead. Add the aisle, electrical-room, and service volume a land campus must carry (the pod has none of it; see §B), and the practical path-length advantage runs from roughly 5× to 20×.

Scale	Pods	Lattice edge (≈ $V^{1/3}$)	Worst-case span	One-way $t_\text{prop}$
1 GW	100	~30 m	70–90 m	0.35–0.45 µs
10 GW	1,000	~65 m	140–180 m	0.7–0.9 µs

The lattice edge assumes ~240 m³ per pod (§B) plus an equal allowance for backplanes, structure, and flow channels, i.e. ~500 m³ per pod gross. A 1 GW block is then ~50,000 m³ of compute, a cube ~37 m on a side — the footprint of a single large building, holding a hundred times the compute of one such cluster. That single sentence is the whole geometric argument, and §B and §D are where it is earned.^[41]

E. The Internal Thermal Loop and Pump Power

For a single-phase dielectric fluid (specific heat $c_p \approx 2{,}274\,\mathrm{J\,kg^{-1}\,K^{-1}}$, density $\rho_f \approx 800\,\text{kg/m}^3$, a synthetic-hydrocarbon proxy) rejecting facility heat $Q \approx 10.5\,\text{MW}$ at a 10 K loop rise:^[9][16]

$\dot{m} = \frac{Q}{c_p\,\Delta T} = \frac{10.5\times10^6}{2{,}274\times10} \approx 462\,\text{kg/s} \quad\Rightarrow\quad \dot{V} = \frac{\dot m}{\rho_f} \approx 0.58\,\text{m}^3/\text{s}$

Pump power. Hydraulic power is $P_\text{pump} = \dot V\,\Delta P / \eta$. The loop pressure drop $\Delta P$ depends on architecture — a full-immersion bath runs low (~1.5 bar), a cold-plate loop higher (~3.5 bar):

Loop type	$\Delta P$	Hydraulic power	At $\eta = 0.65$	Fraction of 10 MW
Full immersion	1.5 bar	87 kW	133 kW	1.3%
Cold-plate	3.5 bar	203 kW	312 kW	3.1%

The seawater side adds its own pumping. Sizing the seawater flow for a 5 K rise gives $\dot m_\text{sw} = Q/(c_{p,\text{sw}}\Delta T) \approx 526\,\text{kg/s}$ ($\dot V \approx 0.51\,\text{m}^3/\text{s}$); at ~0.75 bar through the exchanger, ~59 kW, or 0.6%. Total pumping is therefore ~2–4% of IT load — and this is the dominant contributor to the facility-overhead budget in §L.^[42]

F. The Hull Heat Exchanger — The Thermal Crux

Rejecting 10.5 MW into 20–22 °C water is the hardest steady-state problem in the design. The exchanger obeys^[16][43]

$\boxed{Q = U A \Delta T_\mathrm{lm} \quad\Rightarrow\quad A = \frac{Q}{U \Delta T_\mathrm{lm}}}$

The first-order design question is the temperature difference between the pod’s internal cooling loop and the surrounding seawater. A 35 °C bulk loop against ~20 °C seawater gives $\Delta T_\mathrm{lm} \approx 12\text{–}15,\mathrm{K}$, which is a conservative cold-loop case. Liquid-cooled compute can run warmer than this if the processors, memory, seals, connectors, dielectric fluid and reliability model are qualified for it. A more realistic concept range is a 45–55 °C hot loop, with 60 °C as an aggressive case.

For a clean seawater-to-fluid exchanger, using $U \approx 800\text{–}1{,}100,\mathrm{W,m^{-2},K^{-1}}$, the required wetted area falls sharply as loop temperature rises:

Loop case	Effective $\Delta T_\mathrm{lm}$	Area at $Q = 10.5,\mathrm{MW}$ and $U = 900,\mathrm{W,m^{-2},K^{-1}}$
35 °C cold-loop case	~14 K	~830 m²
45 °C design case	~24 K	~490 m²
50 °C design case	~29 K	~400 m²
55 °C hot-loop case	~34 K	~340 m²
60 °C aggressive case	~39 K	~300 m²

The cylindrical hull of §B has an external area of roughly

$A_\mathrm{hull} \sim \pi D L \approx \pi(4.5)(15) \approx 210\mathrm{m^2}.$

So the pod still cannot reject 10 MW through a bare hull. But the exchanger penalty is less severe than the cold-loop case implies. With a qualified hot loop, the design needs roughly 1.5–4× the hull area in wetted, corrosion-resistant exchanger surface, rather than necessarily 3–5×. That surface can be folded into fins, keels, external exchanger panels, or a jacketed flow path around the capsule.

Biofouling is the dominant long-term degradation. Marine growth on the seawater side adds a fouling resistance $R_f$ in series, so^[44]

$\frac{1}{U_\mathrm{dirty}} = \frac{1}{U_\mathrm{clean}} + R_f$

Using a 50 °C loop case with $\Delta T_\mathrm{lm} \approx 29,\mathrm{K}$, fouling changes the exchanger area as follows:

$R_f$ ($\mathrm{m^2K/W}$)	Condition	$U$ from 1,100 clean	Required area	Penalty
0	pristine	1,100	~329 m²	—
0.0002	light film	902	~401 m²	+22%
0.0004	moderate	764	~474 m²	+44%
0.0008	heavy	579	~626 m²	+90%

At 200 m, reduced light suppresses photosynthetic growth, but microbial films and sessile marine organisms can still colonise surfaces. The design must therefore oversize the exchanger for end-of-cycle fouling, favour smooth high-flow geometries that shed growth, and tie the retrieval interval (§K) to fouling as much as to silicon refresh.

Check against known systems. The duty is not exotic by marine standards. Ships and power stations reject tens of megawatts to gigawatts into seawater through engineered exchangers and condensers. The pod’s exchanger is smaller than a power-station condenser and comparable in thermal duty to marine machinery. The hard part is not the heat-transfer equation. It is doing it sealed, fouled, retrievable, and reliable for years around frontier AI hardware.

G. The Thermal Plume and the Federation Limit

The far-field temperature rise of a heated discharge into a current of speed $U$, mixing thickness $H$, and plume width $W$ is, to first order,^[45]

$\boxed{\,\Delta T \approx \frac{Q}{\rho\,c_p\,U\,H\,W}\,}$

With seawater $\rho c_p \approx 4.1\times10^6\,\text{J/m}^3\text{K}$, $U \approx 0.1\,\text{m/s}$, and $H \approx 50\,\text{m}$, the denominator is $\approx 2.05\times10^7 \cdot W$ (in watts per kelvin), so $\Delta T \approx Q / (2.05\times10^7\,W)$. The plume width $W$ scales with the field footprint, roughly as $\sqrt{\text{area}} \propto \sqrt{P}$:

Field	$Q$	Footprint $W$	Far-field ΔT
1 GW compact	~1.0 GW	~300 m	0.1–0.2 °C
10 GW compact	~10.3 GW	~1,000 m	0.5–0.8 °C
100 GW compact	~103 GW	~3,000 m	~2–4 °C

This is a deliberately first-order model — enough to answer the only question that matters at concept stage: does the system scale as one heat source or as a federation? International-finance environmental guidance commonly limits a discharge to a 3 °C rise at the edge of a defined mixing zone.^[17] A 1 GW and a 10 GW campus clear that with spacing and current. A single compact 100 GW field crosses it. The conclusion is structural, not marginal: 100 GW must be built as ~ten 10 GW campuses spread along the coast, routed through higher-current sectors, and instrumented continuously. The ocean sets the size of the organ; the network (§C, §D) joins the organs into one mind.

H. Power Delivery — Solar Field, HVDC, and Tiered Firming

Solar sizing. Active module area for nameplate $P_\text{PV}$ at module efficiency $\eta_m$ and $I_\text{STC} = 1{,}000\,\text{W/m}^2$:

$A_\text{active} = \frac{P_\text{PV}}{\eta_m\,I_\text{STC}}$

At $\eta_m = 0.25$, every 1 GWp needs ~4 km² of module before spacing; a ~1.5× land factor covers tilt, rows, and access. Oman PV yield is ~1,900–2,000 kWh/kWp/year (~22% capacity factor).^[19][46] Sizing the field at ~1.5× facility peak gives ~9 effective full-load hours per training day.

IT peak	Facility peak (PUE 1.03)	PV nameplate (1.5×)	Active area	Land (1.5×)
1 GW	1.03 GW	1.55 GWp	~6.2 km²	~9.3 km²
10 GW	10.3 GW	15.5 GWp	~62 km²	~93 km²

HVDC export. End-to-end loss is dominated by the two converter terminals, not the short cable:

$\eta_\text{HVDC} \approx 1 - (L_\text{conv,shore} + L_\text{conv,sea} + L_\text{cable})$

Modern VSC/MMC converters lose under 1% per terminal; over a 5–15 km subsea run the cable loss is ~0.1–0.3%.^{[20][47][48][49]} Total end-to-end loss is ~1.5–2.5% — small enough that it sits inside the PUE budget rather than beside it.

Storage is tiered by workload value. The minimum system needs only ride-through storage. Because training can checkpoint and resume, a solar-following pod only needs enough battery capacity to ride through cloud transients and shut down cleanly:

$E_\mathrm{BESS} = 1.03\,\mathrm{GW} \times 0.25\,\mathrm{h} \approx 0.26\,\mathrm{GWh\ per\ GW}.$

That is the proof-machine case. It minimises capital and demonstrates the pod, the thermal system, the subsea power interface, and the station scheduler.

The frontier case is different. If the pod contains current-generation or next-generation frontier silicon, night becomes expensive. A 1 GW field with roughly $15B of accelerators should not sit idle for fifteen hours if storage is cheaper than the lost productive time. A public round-the-clock solar benchmark is approximately 5.2 GW of PV and 19 GWh of batteries to deliver 1 GW continuously.^[50] That is a useful current anchor, not the final Leviathan design.

By the time Leviathan reaches bankable scale, the relevant battery cost is not the 2026 cost. It is the 2031–2036 cost. Battery storage costs have already fallen sharply, and further reductions are expected as stationary storage becomes a larger share of global battery demand.^[51][52] The architecture should therefore be designed around tiered firmness:

Pod class	Hardware age	Power model	Economic logic
Frontier	Years 0–2	Firmed solar plus storage / hybrid backup	Protect the most expensive silicon and maximise iteration rate
Frontier-adjacent	Years 2–4	Partly firmed / scheduled	Run high-value work first; defer what can wait
Plateau	Years 4+	Solar-following	Cheap compute absorbs intermittency

The station does not choose between solar-following and round-the-clock operation. It allocates firmness by value. The newest pods get the night. Older pods follow the sun.

I. Duty Cycle, Scaling, and Effective Work

Daily delivered work scales the peak-compute multiple by duty cycle and realised programme efficiency. For a 1 GW campus at ~110× the reference cluster’s peak, 37.5% duty (9 h/24 h), and 70% efficiency after scheduling, checkpointing, and networking losses:

$110 \times 0.375 \times 0.70 \approx 29$

Scale	Peak vs reference	Solar-adjusted daily work
1 GW	~110×	~25–35×
10 GW	~1,100×	~200–350×
100 GW	~11,000×	~1,000–3,000×

The solar-adjusted table is the low-storage case. It is useful because it proves that even a sun-following station can produce extraordinary throughput. It is not the only commercial configuration.

If frontier pods are firmed for round-the-clock operation, the duty-cycle multiplier changes. The limiting factor becomes realised programme efficiency rather than sunlight hours. A 1 GW station at ~110× the reference cluster’s peak and 70% realised programme efficiency would deliver roughly:

$110 \times 1.0 \times 0.70 \approx 77$

current frontier-cluster-days per day.

Scale	Peak vs reference	Solar-following daily work	Firmed-frontier daily work
1 GW	~110×	~25–35×	~70–85×
10 GW	~1,100×	~200–350×	~700–850×
100 GW	~11,000×	~1,000–3,000×	~7,000–8,500×

These are projected operating envelopes. The correct dispatch depends on hardware age, workload value, storage cost, solar yield, customer urgency and the software scheduler. The economic point is simple: as battery costs fall and frontier silicon remains expensive, the newest pods increasingly justify firmed power. The older pods do not need it.

Honest distinction: strong vs weak scaling. The peak multiples are near-trivial — a hundred clusters have a hundred clusters’ compute. The interesting and contested quantity is the 70% efficiency factor, and it means two different things:

• Weak scaling (many independent runs in parallel, or sparse/asynchronous workloads — MoE, self-play, synthetic-data generation, RL): the 70% is easy, the multiple approaches the raw peak, and the work does not depend on the subsea topology at all. You could run these anywhere.
• Strong scaling (one model trained faster across the whole field): the 70% is doing enormous work, and sustaining it across ~500,000 GPUs is an extraordinary claim that lives or dies on the fabric of §C. This is the regime where the architecture’s coherence advantage matters — and where it must be proven, not assumed.

The body’s multiple is defensible for the mixed workload most labs actually run. The headline multiples at 10 GW and 100 GW are weak-scaling figures; read them as throughput, not as single-run speed-up.^[41]

The second multiplier: software. Hardware throughput is one axis. Computational Abundance documents a second running in parallel: the compute needed to reach a fixed capability falls as the software improves, roughly threefold a year in pretraining and faster still in inference.^[31] Over the five-year build of a station, that axis compounds into one to two further orders of magnitude of effective capability per delivered FLOP. The two multipliers are independent and they multiply. A station finishing a thousandfold more throughput, running software that extracts a hundredfold more capability per operation, is the arithmetic behind the body’s claim that 100 GW enters a range with no precedent to measure it against.

J. Marine Survivability — Pressure, Corrosion, Connectors, Retrieval

The pod must survive years sealed at 200 m. The question is whether the marine environment imposes anything novel. It mostly does not — the pressure regime is benign by subsea standards — but the connector and retrieval problems are real.

1. Hydrostatic pressure. At 200 m,

$P = \rho g h = 1{,}025 \times 9.81 \times 200 \approx 2.01\times10^6\,\text{Pa} \approx 20\,\text{bar}.$

Twenty bar is modest. A recreational scuba diver reaches ~6 bar at 50 m. Subsea oil-and-gas hardware routinely operates at 100–300 bar (1,000–3,000 m). The pod is best built pressure-compensated: the dielectric fluid is held at ambient pressure, so the hull sees almost no differential and merely contains the fluid, while the (nearly incompressible) fluid transmits 20 bar uniformly to the boards — which, with no air gaps, tolerate it. This is exactly the oil-filled-housing approach used for ROV electronics and proven by Microsoft’s Project Natick, which sealed servers in an inert atmosphere on the seabed and recorded roughly one-eighth the failure rate of an identical land deployment.^[53][54]

2. Corrosion. Seawater plus dissimilar metals drives galvanic attack. Mitigations are standard marine practice: copper-nickel or titanium for wetted heat-exchanger surfaces, cathodic protection by sacrificial anodes, dielectric isolation of dissimilar metals, and anti-fouling coatings on smooth exterior geometry.^[55][56] The hull refit interval (~8–10 years) is set by corrosion and fouling, decoupled from the faster IT-refresh interval (§K).

3. Wet-mate connectors. Each pod needs subsea power-in and fibre, mated and de-mated underwater by ROV. Subsea wet-mate connectors exist at the required power and data rates in oil-and-gas and subsea-power practice.^[57][58] The gap is reliability expectation: subsea energy tolerates downtime that synchronous training does not. This is one of the two genuinely unproven subsystems and belongs on the gate list.

4. Retrieval mass. The pod is dominated by fluid mass: $m_\text{fluid} \approx \rho_f V \approx 800 \times 240 \approx 190\,\text{t}$, plus hull, exchanger, and hardware — call it 300–450 t. Recovery needs a heavy-lift or construction vessel with ROV support, not a barge winch.^[59] The operating model is retrieve-to-service, not repair-in-place: lift the pod, refit ashore, redeploy.

5. The single-point catastrophe land has no analogue for. A hull breach floods the pod and destroys ~5,000 GPUs at once — a failure mode no land datacentre carries. Mitigation is compartmentalisation, continuous leak and pressure monitoring, and conservative seal/penetrator design. It cannot be eliminated, only bounded, and it must be priced into the reliability model.

K. Reliability and the Refresh Cascade

A subsea pod should not be treated as a static asset. It is a modular compute organ with a lifecycle.

The vessel, exchanger, connectors, power interface and subsea frame are durable infrastructure. They should be designed for a five-to-seven-year deployment cycle, with retrieval driven by fouling, corrosion, seal life, connector qualification and planned refit. The accelerator stack inside the pod lives on a faster clock. Frontier hardware turns over every few years. A pod that is frontier-class on the day it is lowered will not remain the newest machine forever.

That is the operating model.

The pod should migrate down the workload hierarchy as it ages.

Pod age	Likely role	Workload
Years 0–2	Frontier / premium	Apex training, dense post-training, high-value synthetic data, urgent frontier experiments
Years 2–4	Frontier-adjacent	Reinforcement learning, distillation, architecture search, fine-tuning, simulation, evaluation
Years 4+	Plateau	Batch inference, student models, simulation farms, data generation, scientific workloads, latency-tolerant compute
Refit	Renewal	Retrieve, clean, inspect, replace silicon, redeploy

This resolves the apparent contradiction between silicon cadence and subsea service life. Leviathan does not need every pod to remain frontier for its whole deployment. It needs the station to remain productive as a portfolio.

New pods carry the frontier. Older pods carry the plateau. The station becomes a living cascade.

That model fits the architecture better than a forced two-year retrieval cycle. It reduces vessel time, lowers operating disruption, extends useful asset life, and lets hardware depreciation become a feature rather than a failure. The same thing that happens across the wider economy in Computational Abundance^[31] happens inside the station itself: the frontier falls, the plateau catches it, and useful intelligence continues to run.

The key design implication is that Leviathan should be built for mixed workloads from the beginning. Its software scheduler must assign work by pod generation, network quality, thermal margin, customer priority and hardware capability. Its commercial model must price the station as a tiered fabric, not as a uniform block of identical capacity.

The remaining risk is reliability. Large accelerator fleets fail.^[60] A pod sealed for years cannot be maintained like a land rack. It must carry spare capacity, reconfigurable networking, graceful degradation, leak detection, thermal monitoring and workload migration. The pilot must measure the true failure rate under immersion, pressure, heat and continuous AI load. That number is still unknown.

But the lifecycle answer is clear.

L. Facility Efficiency (PUE), Built Up From the Parts

PUE is total facility energy divided by IT energy.^[61] Unlike the body’s comparison table, this figure can be constructed from the loss terms derived above:

Loss term	Source	Fraction of IT load
Internal + seawater pumping	§E	2.0–4.0%
HVDC export (sea side share)	§H	~0.5–1.0%
Power conversion / distribution in pod	est.	1.0–1.5%
Monitoring, controls, auxiliaries	est.	~0.3%
Total overhead		~3.8–6.8%
Implied PUE		1.04–1.07

So the 1.03–1.05 target is achievable but sits at the optimistic edge. 1.05 is the safe planning anchor.^[62][63] Against land baselines this yields:

Land baseline	Subsea (1.05)	More IT compute per facility MW
1.20	1.05	~14%
1.15	1.05	~9.5%
1.10	1.05	~4.8%
1.35 (legacy air)	1.05	~29%

The honest baseline is land immersion, which already reaches ~1.05–1.10.^[39][64] Against that, the efficiency edge is small. The durable wins are the ones efficiency tables do not show: zero water consumption and a heat sink of effectively unlimited capacity.^[12]

M. Energy and Capital Floor — Silicon Still Dominates

Per gigawatt, the bill of materials is dominated by the accelerators, not the marine infrastructure. A 1 GW field is 100 pods × ~5,000 GPUs = ~500,000 accelerators:

Component	Cost basis	Per 1 GW
Accelerators	~500,000 × ~$30k	~$15B
Solar field	scaled utility PV, 1.5×	~$1.0–1.5B
HVDC + converters + collector	short route	~$0.5–1.0B
Pods, frames, deployment, wet plant (ex-IT)	parametric	~$0.8–1.5B
15-min BESS (§H)	IRENA ~$192/kWh	~$0.05B
Total ex-silicon		~$2.4–4.0B

Silicon is therefore ~79–86% of capital.^[51][41] The marginal energy cost is smaller still: one pod consumes $10\,\text{MW}\times 9\,\text{h}\times 365 \approx 33\,\text{GWh/yr}$, ~$1.0M/yr at $0.03/kWh solar — under 1% per year of the ~$150M of silicon it contains. The conclusion mirrors Bridge to Infinity’s lunar arithmetic in reverse:^[65] the architecture does not make intelligence cheap. It changes where the most expensive machine in the world can physically exist. The solar-and-sea system removes the grid, water, and land constraints.

We’re basing this on a 2026 concept snapshot. The dollar cost of one electrical gigawatt of silicon may not fall quickly, because each new generation of accelerator is both more powerful and more expensive. But the effective cost of computation per gigawatt should fall sharply. The relevant denominator is not only dollars per installed watt. It is training work per dollar, per watt, and per year. On that basis the architecture improves with time: denser racks, better FLOP/W, cheaper solar, cheaper short-duration storage, and compounding software efficiency all make each future gigawatt more productive than the one before it.

N. Siting

The Gulf of Oman shelf varies sharply. The broad Al-Batinah shelf near Sohar reaches 200 m only ~25 km offshore, lengthening the power cable and weakening the case. The narrow shelf at Sur / Ra’s al Hadd reaches 200 m within ~5 km — the strongest candidate for short export runs and compact fields.^[13] Intermediate-depth water along the slope carries the warm, saline Arabian Gulf outflow (~20–22 °C, the §F design point),^{[14][66][67][15]} and sits within an oxygen-sensitive regime that makes continuous local thermal and ecological monitoring non-negotiable.^[68][18]

The gate list. Pulling the weak links together, an honest programme proves these in order before scaling, and no gate assumes the one above it: continuous 10 MW rejection into 20 °C water with managed fouling (§F); safe plume behaviour (§G); wet-mate connector reliability at training-uptime expectations (§J.3); measured in-pod failure rate and graceful degradation (§K); clean retrieve-and-refit (§J.4); and a resolved answer to the frontier-versus-plateau positioning (§K). The physics works at concept level. The programme is what turns it into a machine.

[1] Pilz, K.F., Sanders, J., Rahman, R. & Heim, L. (2025). “Trends in AI Supercomputers.” arXiv:2504.16026. (500-system dataset, 2019–2025. Finds AI-supercomputer performance doubled every nine months while hardware cost and power doubled annually; extrapolates a leading 2030 system requiring ~9 GW. Load-bearing source for the power-per-frontier-run model in §A.)

[2] Sevilla, J., Heim, L., Ho, A., Besiroglu, T., Hobbhahn, M. & Villalobos, P. (2022). “Compute Trends Across Three Eras of Machine Learning.” arXiv:2202.05924. (Finds that in the large-scale era, training compute doubled roughly every six months. Supports the claim that frontier compute has grown far faster than hardware efficiency.)

[3] Shankar, S. & Reuther, A. (2022). “Trends in Energy Estimates for Computing in AI/Machine Learning Accelerators, Supercomputers, and Compute-Intensive Applications.” arXiv:2210.17331. (Reviews slowing energy-efficiency gains from geometric scaling. Supports the §A claim that efficiency gains cannot indefinitely rescue frontier compute growth.)

[4] Koomey, J., Berard, S., Sanchez, M. & Wong, H. (2011). “Implications of Historical Trends in the Electrical Efficiency of Computing.” IEEE Annals of the History of Computing 33(3), 46–54. (Foundational source for long-run computations-per-joule improvement. Background for the efficiency denominator in §A.)

[5] Dennard, R.H. et al. (1974). “Design of Ion-Implanted MOSFET’s with Very Small Physical Dimensions.” IEEE Journal of Solid-State Circuits 9(5), 256–268. (Original Dennard-scaling paper. Background for the claim that historical voltage/power-density scaling has ended.)

[6] U.S. Energy Information Administration. “Electric Power Annual.” (Source for U.S. electricity generating capacity and generation. For the comparison between a tens-of-gigawatt training run and national generating capacity.)

[7] International Energy Agency (2025). “Energy and AI.” (Global report on electricity demand from data centres and AI. Sources broad claims about data-centre electricity growth, grid pressure, and the concentration of AI loads.)

[8] NVIDIA. “NVIDIA GB300 NVL72.” (Official product page for the liquid-cooled GB300 NVL72 rack-scale architecture: 72 Blackwell Ultra GPUs, 36 Grace CPUs, 130 TB/s NVLink, 37 TB fast memory, 1,440 PFLOPS FP4. Load-bearing source for the 72-GPU rack-class unit.)

[9] Pambudi, N.A. et al. (2022). “The immersion cooling technology: Current and future development in energy saving.” Alexandria Engineering Journal 61(12), 9509–9527. (Review of single-phase and two-phase immersion cooling, dielectric liquids, and heat rejection. Sources the dielectric-fluid concept.)

[10] NVIDIA Blog (2025). “Microsoft Azure Unveils World’s First NVIDIA GB300 NVL72 Supercomputing Cluster for OpenAI.” (Reports a 4,608-GPU GB300 NVL72 Azure cluster, i.e. 64 rack-scale systems, with 800 Gb/s-per-GPU Quantum-X800 InfiniBand. The reference cluster for the “one 10 MW pod ≈ one current top-tier cluster” comparison.)

[11] LiVecchi, A. et al. (2019). “Powering the Blue Economy: Exploring Opportunities for Marine Renewable Energy in Maritime Markets.” U.S. Department of Energy. (Marine-energy opportunities, including maritime markets and data-centre concepts. Context for offshore power/cooling integration.)

[12] Mytton, D. (2021). “Data centre water consumption.” npj Clean Water 4, 11. (Review of data-centre water consumption and water-use metrics. Supports the zero-water-consumption claim.)

[13] GEBCO Compilation Group. “GEBCO Gridded Bathymetry Data.” (Global gridded bathymetric dataset. Required source for any claim about how quickly the Oman shelf reaches 200 m near Sur / Ra’s al Hadd.)

[14] NOAA National Centers for Environmental Information. “World Ocean Atlas.” (Global climatology of ocean temperature and salinity at standard depths, including 200 m. Required source for Gulf of Oman water-temperature assumptions.)

[15] Xue, P. & Eltahir, E.A.B. (2015). “Estimation of the Heat and Water Budgets of the Persian Gulf Using a Regional Climate Model.” Journal of Climate 28(13), 5041–5062. (Supports the claim that the Persian/Arabian Gulf is highly evaporative and exports warm, saline water through Hormuz, relevant to Oman intermediate-water assumptions.)

[16] Incropera, F.P., DeWitt, D.P., Bergman, T.L. & Lavine, A.S. “Fundamentals of Heat and Mass Transfer.” Wiley. (Standard heat-transfer reference for $Q = UA\,\Delta T_\text{lm}$, convection coefficients, and thermal-resistance networks. Supports the thermal-loop and heat-exchanger modelling in §E–§F.)

[17] World Bank Group / IFC (2007). “Environmental, Health, and Safety Guidelines: General EHS Guidelines.” (International-finance environmental guidance for the regulatory constraint — thermal-discharge and mixing-zone limits — as distinct from the modelling method in [45]. Use with site-specific local regulation; not a substitute for Omani permitting.)

[18] Queste, B.Y., Vic, C., Heywood, K.J. & Piontkovski, S.A. (2018). “Physical Controls on Oxygen Distribution and Denitrification Potential in the North West Arabian Sea.” Geophysical Research Letters 45(9), 4143–4152. (Regional source for oxygen-minimum-zone and ventilation dynamics in the Gulf of Oman region. Supports the environmental caution around oxygen-sensitive waters.)

[19] World Bank / ESMAP / Solargis. “Global Solar Atlas.” (Solar-resource and PV-output mapping tool. Required source for Oman PVOUT values, solar-field sizing, and location-specific yield assumptions.)

[20] Ardelean, M. & Minnebo, P. (2015). “HVDC Submarine Power Cables in the World: State-of-the-Art Knowledge.” European Commission Joint Research Centre. (Technical reference for submarine HVDC cable systems, route lengths, voltage classes, and deployment experience. Supports subsea HVDC feasibility.)

[21] Lian, X. et al. (2024). “Universal Checkpointing: A Flexible and Efficient Distributed Checkpointing System for Large-Scale DNN Training with Reconfigurable Parallelism.” arXiv:2406.18820. (Shows how large-scale training can checkpoint and resume across changing model-parallel configurations. Supports the solar-following/checkpointable-workload argument.)

[22] Wan, B. et al. (2024). “ByteCheckpoint: A Unified Checkpointing System for Large Foundation Model Development.” arXiv:2407.20143. (Industrial checkpointing system reporting up to 54× reductions in checkpoint stalls. Supports the claim that checkpointing is a real system concern, not a rhetorical convenience.)

[23] Li, S. et al. (2020). “PyTorch Distributed: Experiences on Accelerating Data Parallel Training.” arXiv:2006.15704. (Explains distributed data-parallel training, gradient communication, bucketing, and overlap of communication with computation. Supports the synchronous-training and all-reduce discussion in §C.)

[24] Sergeev, A. & Del Balso, M. (2018). “Horovod: fast and easy distributed deep learning in TensorFlow.” arXiv:1802.05799. (Describes ring-allreduce distributed deep learning and the communication overheads of scaling from one GPU to many. Supports the all-reduce model and the latency/bandwidth distinction.)

[25] Wang, G. et al. (2023). “ZeRO++: Extremely Efficient Collective Communication for Giant Model Training.” arXiv:2306.10209. (Shows that communication volume from weight-gathering, the backward pass, and gradient averaging limits throughput at scale. Supports the §C caution that bandwidth fabric remains load-bearing.)

[26] Fedus, W., Zoph, B. & Shazeer, N. (2022). “Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity.” arXiv:2101.03961. (Sparsely-activated mixture-of-experts model selecting different parameters per input at roughly constant compute, with up to 7× pretraining speed-ups. Sources the §VII claim that the newest models are sparse, with the total network growing while active computation stays small.)

[27] Shazeer, N. et al. (2017). “Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer.” arXiv:1701.06538. (Foundational conditional-computation paper: increases model capacity by over 1000× with only minor losses in computational efficiency. Sources the mixture-of-experts mechanism behind sparse frontier models.)

[28] Wang, Y. et al. (2022). “Self-Instruct: Aligning Language Models with Self-Generated Instructions.” arXiv:2212.10560. (Bootstraps instruction data from a model’s own generations, then finetunes on it. Sources the §VII claim that synthetic data produced in bulk is now a standard part of the frontier’s second economy.)

[29] Geiping, J. et al. (2025). “Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach.” arXiv:2502.05171. (A model that iterates a recurrent block, unrolling to arbitrary depth at test time. Sources the §VII description of reasoning that loops a small block through depth, and the serial, latency-bound character of recursive inference in §C.)

[30] Snell, C., Lee, J., Xu, K. & Kumar, A. (2024). “Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters.” arXiv:2408.03314. (Shows that allocating compute at inference can outperform a much larger model, establishing test-time reasoning as a scaling axis distinct from parameter and pretraining scale. Sources the §VII claim that the frontier is learning to scale the depth of reasoning before answering.)

[31] Age of Wonders. “Computational Abundance.” (Internal cross-reference. Sets out the slot hierarchy — frontier capability discovery, plateau diffusion, edge access — and the depreciation logic Leviathan builds on.)

[32] OpenAI (2025). “Announcing The Stargate Project.” (Primary announcement for the planned USD 500 billion AI-infrastructure programme. Sources the claim that AI infrastructure is moving from ordinary cloud capacity toward national-scale compute buildouts.)

[33] Reuters (2025). “OpenAI, Oracle, SoftBank plan five new AI data centers for $500 billion Stargate project.” (Reports Stargate’s 10 GW target and nearly 7 GW / over USD 400 billion of announced site commitments. Supports the claim that frontier AI infrastructure is already discussed in gigawatt units.)

[34] Reuters (2025). “Vantage Data Centers plans $25 billion AI campus in Texas.” (Reports Vantage’s 1.4 GW, 1,200-acre Frontier campus in Texas with ten data centres. Land-campus comparator for the city-scale/sprawl argument.)

[35] Cottier, B., Rahman, R., Fattorini, L., Maslej, N., Besiroglu, T. & Owen, D. (2024). “The rising costs of training frontier AI models.” arXiv:2405.21015. (Cost models for frontier-model training; estimates amortised training cost grew about 2.4× per year since 2016. Links compute scaling to capital intensity.)

[36] Shehabi, A. et al. (2024). “2024 United States Data Center Energy Usage Report.” Lawrence Berkeley National Laboratory. (U.S. data-centre electricity-use estimates and projections. For national-scale load comparisons and conventional data-centre energy context.)

[37] NVIDIA. “NVIDIA Vera Rubin NVL72.” (Official product page for the next-generation Vera Rubin NVL72 architecture. Supports the thesis that post-bust hardware may be denser, more power-efficient and more rack-coherent, strengthening the topology argument.)

[38] Open Compute Project. “Advanced Cooling Solutions.” (Industry working group and specification hub for data-centre liquid cooling, including immersion and cold-plate approaches. Grounds liquid-cooling terminology and requirements.)

[39] Haghshenas, K., Setz, B., Bloch, Y. & Aiello, M. (2022). “Enough Hot Air: The Role of Immersion Cooling.” arXiv:2205.04257. (Compares air and immersion cooling on energy, PUE, power density, cost and maintenance. Supports claims that immersion enables higher densities but creates maintenance and reliability trade-offs.)

[40] NVIDIA. “NCCL User Guide: Collective Operations.” (Official documentation for all-reduce, reduce-scatter, all-gather and all-to-all. Technical support for the communication-model section.)

[41] Atlas / Leviathan internal model. No external link. (The 2D-versus-3D geometry model, $D_{2D}/D_{3D} = N^{1/6}$, the 10 MW pod-to-1 GW/10 GW scaling tables, the solar-adjusted frontier-cluster-day calculations, the first-order plume model, and the facility economics are original modelling assumptions. They are derived estimates, not sourced facts. The sources above support the inputs; they do not independently validate the full Leviathan architecture.)

[42] Jadhav, S. & Liu, Z. (2026). “Digital Twin-Based Cooling System Optimization for Data Center.” arXiv:2603.01198. (Uses operational data from the liquid-cooled Frontier supercomputer cooling system to model and optimise pumping and supply-temperature control. Supports the §E and §L pump-power and facility-overhead reasoning.)

[43] Shah, R.K. & Sekulić, D.P. (2003). “Fundamentals of Heat Exchanger Design.” Wiley. (Reference for sizing, log-mean temperature difference, fouling resistance, overall heat-transfer coefficients, and compact exchanger design.)

[44] Bott, T.R. (1995). “Fouling of Heat Exchangers.” Elsevier. (Reference on fouling mechanisms and fouling resistance. Supports the §F treatment of biofouling as the long-term thermal-performance risk.)

[45] U.S. Environmental Protection Agency (2024). “PLUMES2.0 — Dilution Model.” See also the PLUMES2.0 Model Theory and User Manual and the legacy Visual Plumes (4th edition). (EPA’s current public plume/dilution model for surface-water effluent discharges. It computes near-field dilution governed by source buoyancy, momentum, and ambient currents, then Brooks far-field dilution as the plume is carried by turbulence and currents — and is used to set NPDES mixing zones. Confirms that the §G estimates are concept-stage screening calculations, not permit-grade modelling, where real dilution depends on discharge flow, density, diffuser geometry, depth, stratification, and currents.)

[46] ESMAP / World Bank (2019). “Global Solar Atlas 2.0: Technical Report.” (Methodology for Global Solar Atlas irradiance and PVOUT data. Supports the use of GSA for concept-stage solar modelling.)

[47] Wikipedia. “High-voltage direct current.” (Secondary overview for typical HVDC loss ranges and the submarine-cable advantage of DC over AC. Prefer JRC / CIGRÉ / vendor sources for formal work.)

[48] Worzyk, T. (2009). “Submarine Power Cables: Design, Installation, Repair, Environmental Aspects.” Springer. (Core engineering reference for submarine power-cable design, installation, repair, and environmental considerations.)

[49] IEC. “IEC 62895: High voltage direct current (HVDC) power transmission — Cables with extruded insulation and their accessories.” (Standards reference for HVDC cable systems. For later engineering diligence, not headline claims.)

[50] Reuters (2026). “How falling battery costs are igniting race for round-the-clock solar power.” (Reports Masdar’s UAE project using ~5.2 GW of solar and ~19 GWh of batteries to deliver 1 GW of round-the-clock power. Benchmark for firmed-solar sizing and the night-storage-versus-ride-through comparison in §H.)

[51] International Renewable Energy Agency (2025). “Renewable Power Generation Costs in 2024.” (Source for solar-PV and battery-cost trends. Supports solar-field and BESS cost assumptions in §H and §M.)

[52] International Energy Agency (2024). “Batteries and Secure Energy Transitions.” (Battery storage costs are projected to fall substantially by 2030 as stationary storage becomes a larger share of global battery demand. Supports the §H/§M case for modelling 2031–2036 storage economics, not only 2026 costs.)

[53] Microsoft Research. “Project Natick.” (Primary Microsoft Research site for subsea data-centre experiments. Use carefully: Natick supports subsea feasibility at hundreds of kilowatts, not 10 MW AI pods.)

[54] Microsoft Source (2020). “Microsoft finds underwater datacenters are reliable, practical and use energy sustainably.” (Retrieval and assessment of the Natick Phase 2 module after two years at 117 ft, with 864 servers showing one-eighth the failure rate of a land control group. Reliability and sealed-environment precedent.)

[55] DNV. “Recommended practices and standards for offshore/subsea systems.” (DNV standards ecosystem for offshore and subsea reliability, corrosion, qualification, and inspection. A diligence pointer rather than a single-claim source.)

[56] AMPP (formerly NACE). “Standards for corrosion control.” (Standards source for marine corrosion, cathodic protection, and materials selection. Supports the §J corrosion-control gate.)

[57] Teledyne Marine. “Subsea Connectors.” (Commercial reference for wet-mate and subsea connector families. Shows subsea power/data connectors exist, without proving Leviathan-level reliability.)

[58] Siemens Energy. “Subsea Power Grid.” (Offshore reference for subsea power-distribution architecture. Analogue for the wet electrical infrastructure, though oil-and-gas uptime expectations differ from synchronous AI training.)

[59] DNV. “Rules and Standards.” (Certification and offshore-engineering standards. For later vessel, lifting, subsea-frame, pressure-envelope, and operational-safety diligence.)

[60] Uptime Institute. “Annual Outage Analysis.” (Reference for conventional data-centre reliability and outage context. Frames the difference between data-centre uptime and sealed subsea pod reliability.)

[61] The Green Grid. “PUE: A Comprehensive Examination of the Metric.” (Defines and contextualises Power Usage Effectiveness. For comparing land PUE, immersion PUE, and the constructed facility-overhead model in §L.)

[62] Green500 / TOP500. “Green500 List.” (Tracks measured energy efficiency of high-performance computing systems. Background for FLOP/W and system-efficiency comparisons, with care for AI accelerators and training workloads.)

[63] ASHRAE. “Datacom Series.” (Thermal-guidance series for data centres and IT equipment environments. For later thermal design, liquid-cooling, and operating-envelope diligence.)

[64] Oak Ridge National Laboratory. “Frontier.” (Reference point for liquid-cooled exascale HPC facility design and high-density compute operations. Analogue for facility cooling, not AI-specific cluster topology.)

[65] Age of Wonders. “Bridge to Infinity.” (Internal cross-reference for the launch-economics arithmetic mirrored in reverse in §M: the architecture changes where the most expensive machine can exist, not what it costs to build.)

[66] Copernicus Marine Service. “Copernicus Marine Data Store.” (Operational and reanalysis ocean datasets for temperature, salinity, and currents. For site-specific Oman thermal-plume, current, and seasonal-variability modelling.)

[67] Pous, S., Lazure, P. & Carton, X. (2015). “A model of the general circulation in the Persian Gulf and in the Strait of Hormuz.” Continental Shelf Research 94, 55–70. (Regional-circulation model for the Persian Gulf and Strait of Hormuz. Background for saline outflow into the Gulf of Oman.)

[68] Argo Programme. “Argo.” (Global network of profiling floats measuring ocean temperature and salinity. For checking World Ocean Atlas and Copernicus profiles near Gulf of Oman candidate sites.)