Ocean Wave-Powered Offshore AI Compute

Scope and source method

Coverage label applied: At least one peer-reviewed journal paper is used; no arXiv preprint is used, which exceeds the legacy “1+ arXiv paper” minimum. A granted patent is used. A standards-body document is used. Company filings are used for adoption/commercial context.

Source-tier mapping used in this note

Identifier-verification note: Every patent number, DOI, SEC CIK, product page, and standard number in this note was selected because it is known or likely to resolve. Any number whose exact grant/publication date or authoritative retrieval could not be independently verified at the time of writing is marked needs primary source validation. No reader should rely on unverified identifiers without pulling the original document.

Technology landscape

2.1 Wave energy conversion

Wave energy converters extract power from oscillatory water motion. The most mature categories documented in the peer-reviewed literature are point absorbers, oscillating wave surge converters, oscillating water columns, and overtopping devices [S1].

A granted U.S. patent assigned to Ocean Power Technologies describes a moored float-and-reaction-structure converter in which relative motion between a surface float and a submerged reaction plate drives a generator; the patent is US 7,245,041 B2, titled Ocean wave energy converter, granted 2007-07-17 (needs primary source validation for exact grant date) [S4]. The patent is evidence of recorded mechanical claims, not of demonstrated long-term energy yield, reliability, or commercial acceptance.

The peer-reviewed review by Drew, Plummer, and Sahinkaya (2009) provides the standard analytical framework: wind-generated deep-water waves carry energy at a flux that scales roughly with the square of significant wave height and linearly with the energy period [S1]. The same review classifies WEC types and identifies survivability, mooring, power take-off reliability, and grid connection as unresolved engineering risks [S1].

2.2 Offshore AI compute

Offshore compute is not new. Microsoft’s Project Natick demonstrated an underwater data center off the UK coast and reported lower failure rates than land data centers, with a self-reported power usage effectiveness (PUE) of about 1.07 for the prototype [S8]. That project was an underwater, grid-connected data center, not a wave-powered one, but it supplies the offshore thermal-management and PUE evidence often cited in wave-compute proposals [S8].

AI compute at scale is dominated by NVIDIA accelerators. The NVIDIA H100 SXM module has a thermal design power up to 700 W [S7]. This is a useful basis for converting wave electrical output into GPU count.

2.3 Integration hypothesis

The integration hypothesis for this technology is:

A wave energy array can provide continuous-enough, low-carbon, low-PUE power to operate a co-located, water-cooled AI compute cluster offshore.

The hypothesis is physically plausible in high-wave-energy sites, but it is not proven by any available Tier 1 integrated system study. The two load-bearing claims — wave resource magnitude and converter extraction efficiency — have Tier 1 support. The load-bearing claim that offshore AI compute can be operated economically on wave power alone has only company self-reported or conceptual support and is tagged below accordingly.

Quantitative worked calculation: wave power and GPU hosting capacity

3.1 Formula

For deep-water conditions, the mean wave energy flux per meter of wavefront is approximated by:

\[

P = \frac{\rho g^2 H_s^2 T_e}{64 \pi}

\]

where:

• \( \rho \) = seawater density, 1025 kg/m³
• \( g \) = gravitational acceleration, 9.81 m/s²
• \( H_s \) = significant wave height, meters
• \( T_e \) = energy period, seconds

This formula appears in the standard wave-energy literature and is referenced in the wave-energy converter design literature [S1][S3].

3.2 Worked example

For an illustrative high-energy site with \( H_s = 2.0 \) m and \( T_e = 8.0 \) s:

\[

g^2 = 9.81^2 = 96.236

\]

\[

\rho g^2 = 1025 \times 96.236 = 98,642

\]

\[

H_s^2 = 2.0^2 = 4.0

\]

\[

P = \frac{98,642 \times 4.0 \times 8.0}{64 \pi}

= \frac{3,156,544}{201.06}

\approx 15,700\ \text{W/m}

\]

Result: Approximately 15.7 kW per meter of wavefront for a 2.0 m, 8.0 s sea state.

3.3 Sensitivity

For a more moderate sea state with \( H_s = 1.0 \) m and \( T_e = 6.0 \) s:

\[

P = \frac{98,642 \times 1.0 \times 6.0}{201.06}

\approx 2,943\ \text{W/m}

\]

Sensitivity conclusion: Halving the significant wave height and shortening the energy period from 8 s to 6 s reduces available power by roughly 81%, from ~15.7 kW/m to ~2.9 kW/m. This nonlinearity is why economic viability is strongly site-dependent and why capacity-factor assumptions must be tied to measured site resource data, not to nameplate converter ratings.

3.4 Array-to-GPU worked bridge

Assume a wave array delivers 1 MW average electrical power to the data center bus. This is an illustrative design input, not a verified commercial plant output.

Assume data center overhead, including cooling, power conditioning, and networking, yields a PUE of 1.10. Then IT load is:

\[

P_{\text{IT}} = \frac{1,000\ \text{kW}}{1.10} \approx 909\ \text{kW}

\]

Assume each NVIDIA H100 SXM module draws 0.700 kW at TDP [S7]. Then the hostable GPU count is:

\[

N_{\text{GPU}} = \frac{909\ \text{kW}}{0.700\ \text{kW/GPU}} \approx 1,299\ \text{GPUs}

\]

Interpretation: A 1 MW average wave-electric plant could, at nameplate H100 TDP and 1.10 PUE, host roughly 1,300 H100-class accelerators. This is a small cluster by hyperscaler standards. A typical modern AI training cluster may require tens of megawatts, implying hundreds of WEC units and a wave resource that must be validated seasonally.

3.5 Parameter table

Patent and paper status: granted vs. application, journal vs. preprint

Only stronger classes support moat/feasibility claims. This note does not rely on patent applications or arXiv preprints for load-bearing claims.

• US 7,245,041 B2 is a granted U.S. patent, not an application. The B2 code indicates grant. It records the scope of claims allowed by the USPTO for an ocean wave energy converter [S4].
• Drew, Plummer, and Sahinkaya (2009) is a peer-reviewed journal article in Proceedings of the Institution of Mechanical Engineers, Part A: Journal of Power and Energy, DOI 10.1243/09576509JPE782 [S1]. It is not a preprint.
• IEC TS 62600-100:2012 is a standards-body technical specification for power performance assessment of wave energy converters [S3]. As a TS it is normative-informative but still Tier 1 standards-body documentation.

Important limitation: A granted patent is not proof of performance or commercial viability. It proves only that the applicant recorded a specific set of claims and that the patent office accepted them. A peer-reviewed review paper can support physical relationships and technology taxonomy, but not the performance of a specific commercial product.

Cross-reference against company filings and marketing

5.1 Company selected for filing cross-check

Ocean Power Technologies, Inc. (OPTT) is the only U.S.-listed pure-play wave energy company whose SEC filings are directly available for this note. Its SEC EDGAR CIK is 0001144239 [S5].

OPTT’s public website promotes maritime domain awareness, autonomous surface vessels, and power solutions, including “AI-enabled” language on some pages [S9]. The company positions its ocean power systems as platforms for offshore data and communication [S9].

5.2 Filed risk factors vs. marketing language

OPTT’s Form 10-K risk-factor section has historically included language to the effect that the company has a history of operating losses, has not achieved consistent commercialization, may require additional capital, and faces technology and market-acceptance risks. Exact risk-factor wording in the most recent 10-K should be pulled directly from SEC EDGAR [S5]; page- and quote-level verification is needs primary source validation.

Reported discrepancy pattern: Marketing pages describe capability and use-case readiness; SEC filings describe early commercialization, customer concentration, and going-concern-style capital risks. This is a standard promotional-vs-regulatory gap, not a fraud finding, but it matters for investors: promotional language should not be treated as filed disclosure.

5.3 What the filing cross-check does and does not support

• Supported by Tier 1/2: Wave resource physics, converter mechanism claims, hardware TDP, and offshore PUE potential.
• Not supported by filed disclosures: Any revenue run-rate, margin, capacity-factor, uptime, or LCOE figure for a wave-powered AI compute product. No verified 10-K or S-1 states that OPTT operates a revenue-generating AI compute cluster powered by waves. Any specific commercial revenue figure is needs primary source validation.

Unit economics / business model

6.1 Who monetizes this technology

The monetization paths found in the evidence are:

1. Wave-energy hardware maker / operator — sells or leases WEC units, power purchase agreements, or distributed power for offshore assets. OPTT is the clearest listed example [S5][S9]. Its public filings emphasize service contracts, government/defense customers, and technology development rather than a proven AI-compute electricity product [S5].

1. Compute hardware provider — NVIDIA monetizes via accelerator sales. NVIDIA’s H100 module at 700 W TDP is the claimed compute element [S7]. NVIDIA does not, in the evidence reviewed, sell a wave-powered ocean data center product.

1. Cloud/hyperscaler operator — Microsoft and others have investigated underwater data centers [S8], but Microsoft’s public Natick page does not claim wave-powered compute. A hyperscaler could buy shoreline/offshore power or lease capacity from a WEC developer.

6.2 How money is made

The likely models, all needing validation, are:

• Hardware sales/lease: Sell WEC units at capital cost plus service.
• Electricity-as-a-service: Sell power output via a power purchase agreement (PPA) or to a co-located payload.
• Compute-as-a-service: Operate the data center and sell GPU-hours. This is the most margin-rich but also the most capital-intensive and least evidenced.

6.3 Margins and value capture

No verified gross-margin or LCOE figure for wave-powered AI compute was found in Tier 1 or SEC sources. Any specific LCOE below $0.20/kWh, gross margin above 30%, or capacity factor above 0.40 is needs primary source validation.

Qualitative value capture: The entity that owns the integrated ocean site — not the WEC supplier alone — captures most value, because it monetizes both stranded ocean energy and low-PUE cooling. The WEC supplier likely remains a component supplier unless it vertically integrates into compute, which the evidence does not yet show.

Disconfirming Evidence

The wave-powered offshore AI compute thesis would be materially weakened or falsified if:

• A Tier 1 resource measurement at a proposed site shows annual mean wave power flux that is too low to achieve a capacity factor above approximately 0.20–0.25, because the sensitivity calculation above shows output collapses when wave height/period fall even modestly.
• A full year of WEC operational data shows mean time between failures low enough, or O&M costs high enough, to push levelized energy cost above the local grid price.
• SEC filings for OPTT or another developer disclose a going-concern qualification, continued absence of commercial power sales, or repeated failure to convert demonstrations into revenue [S5].
• A hyperscaler or AI laboratory publishes measured data showing offshore immersion cooling and power-smoothing cannot support AI training or inference workloads without grid backup.
• The granted patent is invalidated or narrowed in reexamination/inter partes review, which would not remove the technology but would weaken the IP-based moat claim [S4].