AI infrastructure should create lasting value for the communities that host it.

← Articles index · Home · Infrastructure · Nori · About

Topics: Environmental · Infrastructure · Energy · Cooling

AI infrastructure is becoming one of the largest industrial buildouts of our generation.

It will require enormous amounts of electricity, equipment and capital. Communities are increasingly asking what they receive in return.

The data center industry's answer has largely focused on reducing impact.

How little water does a facility consume? How efficiently does it use electricity? How much renewable energy does it procure? How many construction jobs does it create?

Those are important questions.

But they begin with an assumption worth challenging: that a data center is fundamentally a large industrial object a community must accommodate, and success means making that burden smaller.

We think AI infrastructure can be designed differently.

It can operate without using local freshwater as a heat-rejection resource. The compute itself can have effectively no visual footprint at the surface. Expansion can create recurring local industrial work instead of one large construction cycle. And at sufficient scale, AI's enormous electricity demand can help support new generation for the communities around it.

The question should not only be how little AI infrastructure takes.

It should be what it leaves behind.

A companion essay, Reducing the Community Footprint of AI Infrastructure, focuses more narrowly on the physical footprint: land, visual profile, noise, freshwater and cooling form. This essay takes the next step: how AI infrastructure can become part of a community's industrial and energy system.

Low water consumption is not the same as no water dependency

The data center industry has made real progress reducing water consumption.

Closed-loop cooling is part of that progress.

We increasingly see new facilities described as consuming only the equivalent of a few households' annual water use.

That can be a useful operating metric. But it does not tell the entire story.

A closed-loop cooling system still has to be filled. That water has to come from somewhere.

Large cooling systems can contain substantial volumes of fluid distributed across piping, heat exchangers and other infrastructure. Once filled, that fluid may circulate for years.

A facility can therefore consume relatively little additional water each year while still depending on water as part of the physical infrastructure required to operate.

The lifecycle matters too.

Fluid that has spent years circulating through an industrial cooling system may contain corrosion inhibitors, biocides, treatment chemicals, metals or other constituents. Depending on its chemistry and local regulations, draining that system can require testing, treatment, permitted discharge or disposal.

None of this makes closed-loop cooling bad.

It means we should be precise about what we are measuring.

There is a difference between:

We consume very little additional water each year.

and:

Our data center does not require local freshwater as a heat-rejection resource.

Seabase is designed around the second standard.

The ocean is already the heat sink

AI racks are becoming extraordinarily dense.

Hundreds of kilowatts can increasingly occupy a physical footprint that once contained a fraction of that compute.

Nearly all of that electrical energy eventually becomes heat.

Seabase starts by putting the compute in an environment containing an enormous natural heat sink.

The ocean.

Our compute modules are sealed from seawater. The racks operate in a controlled internal environment using direct liquid cooling. Heat moves from the compute through a heat exchanger and ultimately into the surrounding seawater.

Seawater does not flow through the GPUs.

Local freshwater is not continuously withdrawn to replace evaporative cooling losses.

The engineering challenge instead becomes managing heat transfer to the surrounding marine environment responsibly.

That means modeling thermal behavior, monitoring temperatures and designing deployments around defined environmental limits.

It is a different cooling problem.

We believe it is a better one for many coastal regions.

The best data center may be the one you cannot see

Water is only one part of a data center's relationship with its surroundings.

Land matters too.

So does what people have to live beside.

A conventional AI campus can require enormous buildings, cooling equipment, substations, backup systems, roads and surrounding acreage.

At hundreds of megawatts, the data center becomes part of the landscape.

Seabase moves the compute itself somewhere else.

Underwater.

Supporting infrastructure still exists. There are shore connections, power systems, fiber, marine operations and service facilities.

But the data halls do not need to occupy the surface.

A coastal region can host significant amounts of compute while relatively little of the actual compute infrastructure is visible from shore.

That distinction becomes increasingly important as AI infrastructure moves closer to major population centers.

Construction jobs are not the same as an industrial base

Large data center projects create significant construction employment.

But construction is inherently temporary.

A project ramps. Contractors arrive. Equipment arrives. The facility gets built. Much of that workforce eventually moves to the next project.

At the same time, the broader data center industry is moving toward increasingly modular construction. More infrastructure can arrive at the site already manufactured.

Seabase is modular too, but our physical architecture creates a different opportunity.

We separate the precision compute infrastructure from the large marine infrastructure supporting it.

Compute modules are compact, sophisticated systems. They can be integrated near the supply chains where racks and AI hardware are assembled, qualified and then transported efficiently to deployment sites.

Foundations are different.

They are large marine structures.

Shipping completed foundations around the world makes less sense when qualified regional fabricators can produce standardized structures closer to deployment.

That gives Seabase the opportunity to localize a meaningful part of the physical supply chain.

The next foundation is another local order

This is where modularity becomes more than an engineering decision.

A qualified regional fabricator can build Foundation 1.

Then Foundation 2.

Then 10.

Then 50.

As the compute region expands, the local manufacturing order book can expand with it.

That creates recurring demand across steel fabrication, welding, coatings, assembly, inspection, marine electrical work, heavy lift, vessel operations, deployment and maintenance.

Many of those capabilities already exist in the industrial ports where Seabase intends to operate.

We do not need every community to manufacture GPUs.

The precision compute supply chain can remain global.

But the large physical infrastructure that is difficult and expensive to transport can increasingly be produced near deployment.

The next foundation is not simply another data center construction project.

It is another order for the local supply chain.

Local production changes the expansion conversation

Infrastructure also becomes easier to understand once people have lived with it.

The first Seabase deployment in a jurisdiction is unfamiliar.

Regulators need data. Ports need operating experience. Marine users need to understand the system. Environmental performance needs to be measured. Local contractors need to learn the deployment process.

Then imagine the conversation several years later.

The infrastructure has been operating.

Environmental monitoring has produced real data.

Deployment and retrieval have been demonstrated.

Local marine contractors service the system.

Regional fabricators have already built the foundations sitting on the seabed.

Local workers are earning their living from the program.

Then Seabase wants to expand.

The question is no longer:

Should we allow an unfamiliar technology into our waters?

It becomes:

Should we add more of the standardized infrastructure our community already knows how to build, deploy, monitor and maintain?

The next expansion creates another local order.

That is how we think infrastructure should scale.

Electricity demand can become an asset

AI infrastructure consumes enormous amounts of electricity.

There is no point pretending otherwise.

But large, predictable electricity demand also has economic value.

New generation needs customers.

Utilities need confidence that new capacity will be used.

Energy developers need long-duration offtake to finance projects.

And many of the ports where Seabase can operate are likely to need considerably more electricity themselves.

Ports are electrifying. Ships increasingly require shore power. Harbor equipment is becoming electric. Autonomous maritime systems will require energy and connectivity. Industrial operations are electrifying.

Those loads can be large, but they are not always constant.

Compute can be.

A large Seabase region can become a predictable 24-hour anchor customer for new generation.

Instead of asking:

Where can we find another 100 MW for our data center?

we want to eventually ask:

Can our commitment to purchase 100 MW help this region build substantially more than 100 MW of new generation?

That changes the relationship between AI infrastructure and the electrical system around it.

Become the anchor, not the energy company

Seabase does not need to build power plants.

Our job is compute infrastructure.

But we can provide something energy developers need:

Demand.

A Seabase region might begin with 5 or 10 MW using available power.

Then 20 MW.

Then 50 MW.

As the region grows, Seabase becomes capable of making increasingly significant long-term power commitments.

That demand can help support new generation.

The right answer to a 100 MW Seabase requirement does not necessarily have to be 100 MW of new generation dedicated to Seabase.

It could be 200 MW.

Or 300 MW.

Seabase can become the anchor customer while additional electricity supports the port, ships, industrial customers and surrounding grid.

Future generation technologies, including small modular reactors where appropriate, make this model particularly interesting around industrial ports.

But the principle is technology agnostic.

We do not need to predict which generation technology wins.

We need to create enough predictable demand to help make new generation economically possible.

Put generation where generation works best

There is no fundamental reason the power plant and the computer need to occupy the same physical environment.

We think the opposite can make more sense.

Put the compute underwater, where the ocean provides an enormous heat sink and the data center's physical footprint can largely disappear below the surface.

Put generation in the environment where generation works best.

For future small modular reactors, that could mean controlled industrial sites at or near ports, where security, skilled labor, heavy logistics, grid infrastructure and marine transportation already exist.

A reactor that arrives by water does not need to remain on the water.

It can be installed at the port and connected to the broader electrical system.

Seabase can consume part of its output.

The port can consume another part.

Ships can use it.

Industry can use it.

The surrounding grid can use it.

The generation asset becomes useful to more than the data center.

Community energy benefits can be measurable

There is an even stronger version of this model.

At sufficient scale, communities should not have to rely on the hope that new generation eventually benefits them.

The benefit can be structured directly.

Large-scale AI compute creates substantial economic value from every megawatt it consumes. If a community enables Seabase to reach significant scale, there can be room to deliberately return part of that value through the energy system.

The mechanism will differ by jurisdiction.

It could mean supporting more generation than Seabase itself requires.

It could mean funding grid improvements.

It could support defined community energy credits.

It could reserve part of a generation project's output for local users under agreed economics.

Or it could involve Seabase directly contributing to a measurable reduction in community energy costs as part of a long-term development agreement.

The exact structure should be designed with the community, utility, port, generator and regulators.

The principle is straightforward:

If a community allows Seabase to reach sufficient scale, that community should participate directly in the value created by that scale.

Not through vague promises.

Through measurable commitments.

A different kind of infrastructure flywheel

Put these pieces together.

A Seabase region starts operating.

Local companies begin producing foundations.

Marine operations grow.

Seabase expands.

More foundations are ordered.

Compute demand becomes large enough to help anchor new generation.

That generation gives the port and surrounding region additional energy capacity.

Additional energy supports port electrification, maritime activity and new industry.

The local supply chain gets stronger.

Seabase expands again.

More infrastructure is ordered locally.

More generation becomes economically viable.

AI infrastructure does not have to arrive, consume scarce resources and exist separately from the economy around it.

It can become part of the industrial system.

Keep the investment above water. Put the footprint below it.

The AI infrastructure buildout is going to be enormous.

Communities will feel it.

The question is what they should have to give up in exchange.

We do not think the answer has to be vast amounts of land.

We do not think local freshwater should have to become a recurring heat-rejection resource for the data center.

We do not think construction employment should be the end of the local jobs story.

And we do not think AI's enormous electricity demand should only be discussed as a burden on the grid.

There is another model.

Local workers can build the physical infrastructure that expands with the region.

Local marine companies can deploy and maintain it.

New generation can be supported by long-duration compute demand.

Ports can modernize around that generation.

Communities can receive measurable energy benefits.

And billions of dollars of compute can operate nearby without billions of dollars of data center buildings dominating the landscape.

Above the surface should be the things communities want:

Jobs. Manufacturing. Energy infrastructure. Port modernization. Investment.

Below it can be the thing they do not need to see:

The data center.

AI infrastructure should not simply become better at taking less from the communities that host it.

It should leave those communities with more than they had before.