[ExI] Where should we build new multi-gigawatt AI Data centers?

Sat Nov 7 13:20:16 UTC 2026

On Fri, Nov 6, 2026 at 11:40 AM <spike at rainier66.com> wrote:

*> Alaska is better than Canada. *

*I asked MR.**GPT about that and he seemed to feel that although Alaska has
possibilities Canada would be better. However near the end of his long
response he said something that I thought was very interesting, and I
wanted to know more about it:*

*ME: ** I think what you said at the end to be particularly interesting,
"building a huge AI campus in northern Alaska, close to natural-gas
resources, with its own power station, cooling system, and fiber-optic
connections. It would essentially be an entirely new industrial city
devoted to computation. I wouldn't dismiss that possibility at all ". Could
you expand on that?*

*GPT: " *I think this is worth exploring seriously, because it suggests a
rather different way of thinking about the future geography of AI. Instead
of finding an existing city with enough electricity to support a gigantic
data center, imagine building an entirely new industrial complex in
northern Alaska, specifically designed around one purpose: supplying
enormous quantities of computing power. It would have its own power
generation, cooling infrastructure, water supply, fiber-optic connections
and perhaps even its own housing and transportation facilities. In effect,
it would be an AI industrial city, built where energy is available rather
than where people already live. Let me walk through how such a project
might work.

*1. Choosing the location: Alaska's North Slope*

The first place I would investigate is the region around Prudhoe Bay and
Deadhorse. Why? Because this region already has something extremely
valuable: a major industrial energy infrastructure. Prudhoe Bay is one of
North America's largest oil and natural-gas fields. It has been producing
hydrocarbons for decades, and the region has roads, airfields, industrial
facilities and pipelines.

There are three particularly important advantages:

   -

   Natural gas is available in enormous quantities, although much of the
   gas is currently reinjected into the oil reservoirs rather than exported.
   -

   The region is extremely cold for much of the year.
   -

   There is relatively little competing demand for land and industrial
   infrastructure compared with major metropolitan areas.

The biggest attraction is that a data center could be located near its fuel
supply rather than having to transport electricity over hundreds of miles.
*2. There is now an actual proposal remarkably close to this idea*

Here's something I discovered while investigating: this is no longer merely
a hypothetical scenario.

An Alaska company called *STAK Energy* has proposed a project called *Project
Aaka*, near Deadhorse on the North Slope. Its proposed design is remarkably
similar to the AI industrial city we have been discussing.

The important details:

Feature

Proposed development

Location

About 26 miles south of Deadhorse

Land

715 acres

Power generation

1–3 GW

Fuel

North Slope natural gas

Computing

Modular high-performance computing buildings

Cooling

Primarily dry-air cooling

Fiber

New connections to existing Arctic fiber infrastructure

Construction workforce

   Up to 1,500 workers

Permanent workforce

 Approximately 60 people

The company proposes building its own natural-gas power station and
consuming the electricity on-site, without connecting to Alaska's public
electricity grid.

And there has just been a significant development.

*On November 5, 2026, STAK announced a 15-year natural-gas supply agreement
with ExxonMobil Alaska Production.* The company says the agreement provides
enough gas to support nearly 3 GW of electricity generation. That is a
substantial step forward, although it doesn't mean the project has received
all its permits or secured financing.

*3. How would the power station work?*

North Slope natural gas fields would lead to gas delivered through short
dedicated pipelines. The gas turbine power station would consist of
multiple independent generating units and electrical substations and
distribution would provide redundant power feeds to computing buildings.

I would actually favor a collection of smaller generating units rather than
one enormous power station.For example, 30 generating units rated at 100 MW
each would provide 3 GW of installed capacity.That has several advantages:

   -

   A turbine can be taken offline for maintenance without shutting down the
   entire campus.
   -

   Power generation can be expanded in stages.
   -

   Individual failures are less disruptive.
   -

   Generation can be matched more closely to the computing load.

There is one further possibility: using combined-cycle generation. That
captures some of the heat in the turbine exhaust to produce additional
electricity through a steam turbine. It improves fuel efficiency, although
it adds complexity and water-management requirements.

*4. Cooling could be one of the biggest advantages*

This is where northern Alaska becomes particularly interesting.Modern AI
computers generate enormous amounts of heat. Almost all the electricity
consumed by the computing equipment eventually becomes heat that must be
removed. A 3 GW computing installation would therefore have to continuously
dispose of approximately 3 GW of heat, assuming that figure refers to the
computing load itself.In northern Alaska, the outside temperature is below
freezing for much of the year.

I would design the facility around three cooling stages:

   1.

   Liquid cooling at the processors: Cold liquid circulates through cooling
   plates attached directly to GPUs and other high-power chips.
   2.

   Heat exchangers: The warmed liquid transfers its heat to a separate
   cooling circuit.
   3.

   Outside-air cooling: Large industrial heat exchangers transfer the heat
   into the Arctic atmosphere.

This arrangement could dramatically reduce the need for evaporative cooling
and freshwater consumption.

There is a caveat: extremely cold temperatures require sophisticated
controls to prevent freezing, condensation and equipment damage. Summer
temperatures also mean that cooling cannot be designed around winter
conditions alone.
The permafrost problem

This is perhaps the most unusual engineering challenge.The ground in much
of the region is permanently frozen. Building an enormous facility on it is
not like building on ordinary soil.The proposed STAK site would require
approximately 7.1 million cubic yards of gravel fill to create a stable
construction platform.

Engineers would need to:

   -

   Insulate the buildings and foundations.
   -

   Prevent waste heat from warming the underlying ground.
   -

   Use foundations designed for frozen soil.
   -

   Monitor ground temperatures and settlement continuously.

If the permafrost thaws unevenly, buildings and pipelines can shift or
become unstable. This is not a theoretical concern: it is a major
engineering issue for existing Arctic oil infrastructure.

*5. Communications: surprisingly, Alaska already has a head start*

You might think that locating an AI campus near the Arctic Ocean would
create a serious communications problem. It certainly creates challenges,
but there is already an important piece of infrastructure.The company
Quintillion operates a subsea and terrestrial fiber-optic system connecting
Prudhoe Bay and Deadhorse with Fairbanks. Its published network
specifications describe:

   -

   Approximately 30 terabits per second of capacity.
   -

   Potential expansion to 90 terabits per second.
   -

   Connections onward to networks reaching Anchorage, Seattle and Portland.

That is a useful starting point. However, a 3 GW AI campus would need
considerably more than a single fiber connection. I would want multiple
independent fiber routes, separate cable landing points, redundant
connections to major internet exchanges and substantial additional capacity.

There are two different networking requirements:

   -

   Inside the campus: Extremely high-speed connections between thousands of
   GPUs. This would be a dedicated, enormous optical networking system.
   -

   Outside the campus: Connections to customers, other data centers, cloud
   networks and the rest of the internet.

The first is built as part of the campus. The second requires a major
expansion of Alaska's external fiber infrastructure.
6. How big would this project actually be?

Let's consider the upper-end proposal of 3 GW.

3,000 MW Installed generating capacity. 26.3 TWh annual electricity if
operated continuously at 3 GW. 715 acres proposed land parcel. 1,500
potential construction workers. 60 estimated permanent employees.

One important distinction: 3 GW of power-station capacity does not
necessarily mean 3 GW of computing capacity. Some electricity is consumed
by cooling, power conversion and other infrastructure, and generators need
reserve capacity. Also, the developer's gas-supply announcement doesn't
disclose the actual gas price or all the commercial terms.The Northern
Alaska Environmental Center, citing STAK's project analysis, reports
projected gas consumption of approximately 350–500 million cubic feet per
day under full simple-cycle operation. That illustrates just how
substantial the fuel requirement would be.

*      7. What could go wrong? *

I see five major risks.

Risk                    Why it matters

Gas pipeline       New pipelines must be constructed and separately
authorized.

Environmental impact    Gas combustion produces substantial CO₂ emissions
and other pollutants.

Permafrost     The huge gravel platform and buildings must remain stable
over decades.

Fiber connectivity      Additional redundant communications infrastructure
is necessary.

Financing and customers      A project this large needs enormous investment
and long-term computing customers.

The environmental issue deserves particular attention.

One attraction of the project is that it would monetize natural gas that
otherwise has limited commercial outlets on the North Slope. But burning
that gas to produce electricity would create substantial greenhouse-gas
emissions. The project would not have the environmental advantages of
Icelandic hydroelectric-powered computing.

There is also a significant local political dimension. The proposed 50-year
land lease attracted more than 500 public comments, and the comment period
was extended. The Northern Alaska Environmental Center reports that fewer
than a dozen comments supported the project. That is an advocacy
organization's account, rather than an official tally of public opinion.

*8. How close is this to actually happening?*

Here is the distinction I think is essential.

Milestone

Status as of November 7, 2026

Proposed 715-acre land lease

          Preliminary state approval only

Natural-gas supply

15-year ExxonMobil agreement announced November 5

Gas pipeline

Separate approval and construction required

Power plant

Proposed, not operational

Computing customers

None publicly identified

Full project financing

Not publicly established

Construction      Not yet started

*My assessment: this could be an important new model for AI infrastructure*

I find the underlying concept quite compelling.

Instead of competing with existing cities for electricity, land and water,
a developer could create an entirely new industrial energy complex in a
remote location.

There are some intriguing parallels with the oil industry itself: pipelines
bring fuel to an isolated industrial site, large power installations
convert it into usable energy, and specialized infrastructure supports a
relatively small permanent workforce.

I would actually consider three variations of this idea:

   1.

   North Slope Alaska: Natural-gas-powered AI campuses, as proposed by STAK.
   2.

   Northern Quebec: Hydroelectric-powered campuses built near new
   generating stations.
   3.

   Northern Manitoba: Large computing facilities supported by new
   hydroelectric generation and transmission.

The first would prioritize access to fuel; the other two would prioritize
low-carbon electricity. One thing I would particularly like to see is
whether AI companies become willing to locate their most
electricity-intensive training operations thousands of miles from major
population centers. For training a frontier AI model, the physical location
of the computers may matter much less than the price and reliability of
electricity, the availability of chips, and the quality of the network.

If that proves true, northern Alaska could become a much more interesting
location for the AI industry than most people currently imagine."

*  John K Clark*
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20261107/4bd6f377/attachment.htm>