Oluwadi
Investigating the Overlooked
Region
What
Who For
Analysis
Oluwadi
Investigating the Overlooked
Bath, Maine builds Navy destroyers. Bath Iron Works, a General Dynamics subsidiary on the Kennebec River, employs roughly 6,000 people, and what they build is heavy, built in pieces, and moved into place by cranes.
That is also the problem an AI data center poses. This piece is about why a place made to build ships is closer to the right place for one than it sounds, and what stands in the way. It is not a report that anyone plans to put one in Bath. I found no such plan, and the yard is a working Navy shipyard. The test it uses, a cold place with floors that carry weight and power that can be had, also points further north, and the last section before the conclusions takes it to Newfoundland and Iceland.
A traditional data center and an AI data center are both rooms full of computers. The difference is in what each rack asks of the room.
Start with electricity. In its 2024 global survey, the Uptime Institute put the industry's average rack at somewhere under 8 kilowatts, creeping up from about 5. One industry guide puts a traditional rack at 10 to 15 kilowatts. A single Nvidia GB200 NVL72 rack, the current top-end AI system, draws about 120 kilowatts by one account and 132 at peak by another. At the Uptime average, that is about fifteen racks' worth of power in one cabinet. That ratio is my arithmetic, not a figure from either source.
Then weight. Nvidia's own rack is reported at 1.36 metric tons, about 3,000 pounds. The same guide that lists the traditional rack at 10 to 15 kilowatts gives a static load capacity of 2,000 to 3,000 pounds for it. So the heaviest AI rack sits at the top of what a traditional rack was built to carry, and it arrives with more behind it: the coolant distribution units that serve it can weigh as much as three tons when flooded, and the guide calls for a floor rated at 800 kilograms per square meter.
Next, cooling and layout. Air cannot remove 120 kilowatts from a cabinet. The NVL72 is liquid-cooled: coolant enters at about 25 degrees Celsius, leaves at about 45, and flows at roughly two liters a second through each rack. The rack also holds more than two miles of copper cabling, per The Register's account of the system. So the room is no longer aisles of air. It is a plumbing and power-distribution plant with computers attached, and the pipes, pumps and heavy electrical gear take real floor and real wall.
This is where Maine's cold matters, and where it matters less than people assume. A 25-degree coolant supply is cool, not cold. My read is that a loop like that can shed its heat to outside air for most of the year in a place with Maine's winters, and a warmer climate has to run chillers to do the same. That saves electricity, which is the largest running cost. It is not a reason to build there by itself. The Oxford Networks account in 2014 named the same advantage, and the state's own council lists the cold climate among the things that could draw developers, so the advantage is real. I did not find a Maine-specific measurement of how many hours a year it holds, and the water side is a design choice: the council heard that evaporative cooling uses less energy but more water, and closed loops the reverse.
Size is the odd one. Because a rack draws so much, a given amount of power fits in far fewer racks. One megawatt is about eight NVL72 racks at 120 kilowatts, against about 125 racks at 8. That is the computers alone, again my arithmetic. The cooling and power plant around them does not shrink. My read is that the building gets denser and heavier, not obviously smaller.
Retrofit is where this bites. A 2025 industry guide estimates $50,000 to $100,000 per rack to bring an existing facility up to 40 kilowatts, and $200,000 to $300,000 per rack for new infrastructure that can take 100 kilowatts or more. I have run operations inside a data center in Ashburn, Virginia, the densest cluster of them anywhere, and the lesson there matched the numbers: most buildings were not built for this, and the limit is physical.
The companies that built the spine of the industry now own the problem. NTT, which absorbed the Ashburn operator I worked inside, and Equinix, among others, are in effect landlords: they built, own and lease the buildings the cloud runs in. The rack above changes what those buildings are worth.
The age of the stock is the issue. The Uptime Institute's 2024 survey found that roughly half of all data centers are over 11 years old and lack the resources to modernize, and that racks above 30 kilowatts remain rare. The Ashburn building at the center of my earlier Oluwadi post opened in 2012, according to the history in that post, and a campus of that vintage was designed for a rack a fraction of the size of the one above. Retrofit is expensive: the guide I cited puts it at $50,000 to $100,000 per rack to reach 40 kilowatts, and that is before a rack at three times that.
The operators are responding, and what they say is instructive. Equinix announced support for direct-to-chip liquid cooling at more than 100 of its data centers in over 45 metros, with Washington, DC among the first sites, according to an industry report from the time. That report did not say whether the sites were retrofitted or newly equipped. In March 2026 NTT said it had secured nearly 115 megawatts across three US locations, including more than 90 megawatts at a campus in Gainesville, Virginia, as part of a $10 billion plan through 2027. It described liquid-cooling loops that can be added in one-megawatt increments and cooling capable of handling "north of 200 kW per rack." Those are the companies' own statements, not independent measurements.
Read it as a capital allocation choice for each landlord, and the options are three: retrofit a building that already stands, build a new campus on cleared land, or sign a long lease with a customer who pays for the build. I do not know the split among them, and I did not find a figure. My read is that rack density is re-marking the sector's real estate. A building that can carry the load and take the power keeps its value, and one that cannot is worth the cost of what it would take to fix it. My read is that this is one reason sites far from the usual markets, and old ones built for another kind of weight, are being looked at again; I have no survey of why developers choose them.
I know the other side of this from having operated inside one of those buildings. We were in the first building of the Ashburn campus, which had raised floors and no liquid cooling. The author's account is that a rack like the one above would have gone through the floor or cooked itself. The first half is a judgment I cannot check without the building's floor certificate. Raised-floor panels are typically rated at 1,000 to 3,000 pounds of concentrated load, according to a trade guide, and that rating is per point, not per rack, so a 3,000-pound rack does not map straight onto it; whether a given floor holds depends on the panel, the understructure and where the feet land. The second half follows from the cooling: a rack that needs liquid cannot run on air. Where a customer wanted to deploy real AI hardware, the building's ceiling decided what was possible, as the earlier post describes. The incumbents are the best placed to fix that, since they own the power connections and the customers. The question is which of their buildings they will spend to fix.
Read the facilities page of Bath Iron Works with that in mind. The yard lists a 750-foot dry dock and three shipways, each able to take a vessel up to 780 feet long. It has a 77,000-square-foot climate-controlled assembly building, the Ultra Hall, that can build units weighing up to 4,000 tons. A 4,000-ton unit weighs about as much as 2,700 to 2,900 of those racks, depending on the ton. It has track-mounted cranes of 300 tons, a transfer facility built as a concrete platform for final assembly, and a structural assembly building of 170,000 square feet.
I am not saying a destroyer module and a server rack are the same load. They are not, and a floor rated for one is not rated for the other until someone checks. My read is that the shared quality is the design intent: a site built to carry very heavy things on a few points, under cover, with the cranes to put them there, starts closer to the answer than a warehouse does.
Set that against the Ashburn building above, with raised floors over an air plenum and no liquid loop, and the contrast is the piece in miniature. One was built to carry air-cooled racks of its day, the other to carry ships. The author's account is that the yard's floors are thick concrete, in a cold place. The record I found supports the concrete: the yard describes its transfer facility as an enormous concrete platform for final assembly. It gives no thickness or load rating, so "thick" is the author's word, and the cold is climate, not a measurement I took. The author's further point is that what carries a heavy load is the reinforcing steel in the slab and the ground beneath it, not thickness alone, and a floor built for ships has both by design. I have no drawing to confirm that for any bay at the yard.
A shipyard is not one kind of place, and the author has walked three of them, Bath, Norfolk and Philadelphia, which he says were each designed differently. The public record bears that out in outline. Bath is a private builder of new ships, owned by General Dynamics, with assembly halls added as recently as 2001 and 2008. Norfolk Naval Shipyard, in Portsmouth, Virginia, is owned and run by the Navy, founded in 1767, with eight working dry docks, and it repairs and modernizes ships, including nuclear carriers, rather than building them. The Philadelphia Navy Yard closed as a shipyard in 1995 and passed to the city; the EPA's account describes 152 businesses and more than 12,000 people working there, on a site with its own 35-megawatt electric grid. That account says nothing about data centers. The author's way of putting it is that they became different things but started as the same need: somewhere to build and keep ships. Each then bent to its owner, its ships and its century. So "a place like Bath" is shorthand, and a builder's assembly halls, a repair yard's dry docks and a redeveloped campus carry different loads in different places. The argument here is about Bath and the kind of site it represents, and I would not carry it to the other two without their drawings.
The second thing a yard has is people. When Bath Iron Works announced nearly 600 new jobs in 2014, the trades it listed were electricians, pipe fitters, outside machinists, welders and tinsmiths. Set that against the AI room above: heavy electrical, a liquid loop, large steel supports. My read is that the trades that run a destroyer's electrical and piping systems are the same trades a dense liquid-cooled hall needs. The yard's own difficulty is the other side of this: it competes with itself for those workers, and layoffs and recalls have gone both ways at the yard over the years. I did not verify current hiring.
The strongest evidence is a few miles away. Naval Air Station Brunswick closed in 2011 and is now Brunswick Landing, run by the Midcoast Regional Redevelopment Authority. It offers more than a million square feet of industrial space, campus-wide fiber, and utilities the authority owns and runs. More than 130 businesses operate there.
One of them was a data center. In 2014 a Lewiston company, Oxford Networks, announced a $6 million expansion of 7,000 square feet within a 52,000-square-foot facility, the base's former communications center. The reasons the company gave in its own account were that the building was built as a highly secure communications center, that the base ran its own power grid in addition to its tie to Central Maine Power, with backup generation, and that Maine's winters cool the servers naturally.
The author's account of why it worked adds the floor: a base building built for military loads can carry heavy equipment. That is plausible and it fits the logic of this piece, but none of the sources I found states it, and I treat it as the author's read until someone checks the structure. Two cautions on the evidence. That facility served small and medium-sized businesses buying cloud services, and none of the sources give its power in megawatts, so it is a proof of concept for the building type and not for AI scale. And I did not verify who runs it today.
I have a reason to know this site beyond the record, and it is not only the reason of a director or an investor: I run large AI models locally, on my own machines, so I write as an operator. I was going to put servers there to run open-source AI models, machines I could not put in the NTT campus in Ashburn, where the ceiling on a rack was about 22 kilowatts. That is the author's account. It is also why we looked at Bath in the first place: we already ran a colocation business in Ashburn, and what we needed next was a cold place that could support the weight. Those were the two requirements, cold and floor, and the old yards and bases of the Midcoast were the obvious place to look. I have not verified the site's power or floor against any document, and the plan ended for a reason unrelated to the site: we shut down the managed-services company that would have run the machines.
The obstacles are power, money and politics, not the building.
On power, the state's own Data Center Advisory Council heard in July that large-load activity in New England is limited, with minimal near-term effect expected before 2027 to 2028. A researcher from Lawrence Berkeley National Laboratory listed Maine's siting deterrents as higher regional power costs, distance from fiber and population centers, and slow interconnection and permitting. The same presentation named what could attract developers: a colder climate, cheaper land and water. The state's energy commissioner warned that adding large loads without adding supply can raise electricity prices for everyone in the regional market.
The scale is worth setting beside the numbers above. ISO New England projects that large loads could add 110 to 130 megawatts to regional peak demand over the next one to two decades. A single data center campus in Ashburn, by its operator's own count, carries 224 megawatts of critical IT load. A project that size would be larger than the whole region's forecast, if I am reading the council summary correctly, and I would want the forecast's definition before leaning on that.
One answer to the power problem is to put the computers where the power is made. A German company, windCORES, has done this since 2020 by installing server racks inside the towers of wind turbines, which are about 13 meters wide. As reported in a 2023 Slashdot item that cites CNN, the sites run 85 to 92 percent on power from their own host turbines and draw the rest from the grid, and the company claims about 10 grams of carbon dioxide per kilowatt-hour against a German data-center average of 430. Those are the company's claims relayed second-hand, the sites are small, and I did not find a figure for their cooling or capacity. I include it only to show the shape of the idea. Whether it fits Maine, or a yard or base on the Kennebec, I have not researched.
On politics, Maine came within one signature of becoming the first state to freeze large data centers. The legislature passed LD 307 on April 9, 2026. It would have paused permitting for data centers drawing 20 megawatts or more through November 1, 2027. The governor vetoed it on April 24, because it did not exempt a $550 million project at the former Androscoggin paper mill in Jay, expected to bring more than 800 construction jobs and 100 or more permanent ones. The legislature failed to override on April 29. That day the governor signed an executive order creating a 15-member Maine Data Center Advisory Council, with recommendations due January 29, 2027. A second law bars data centers from the state's business development tax incentives. A law firm that tracks the matter calls the moratorium vetoed, not dead, and notes the midterm elections could bring it back.
Note what that means for a site like Bath. The state is not saying no to data centers. It is saying: not yet, and not without a plan for who pays for the power. The governor's own objection was to a bill that would have blocked a project built on a closed mill, which is the same logic as a base or a yard. My read is that a former industrial site with local support is the kind of project the state is trying not to freeze.
One more proposal is nearby and unproven. In 2025 Wiscasset's selectboard said an unnamed site assessor had inquired about a town-owned plot next to the former Maine Yankee nuclear plant, which once occupied 820 acres. The board called it conceptual, with no formal proposal, and said it was not known whether the old interconnection could support a large data center.
Everything above is also a question about where a dollar goes.
Take the guide's figures at face value: an average AI rack costs $3.9 million in 2025, against $500,000 for a traditional one, and the infrastructure to host a 100-kilowatt rack runs $200,000 to $300,000. By my arithmetic, the room is about 5 to 8 percent of what goes into it. So the building is not where the money is. The machines are.
My read is that this is exactly why the building decides the return. A $3.9 million rack that cannot be powered, cooled or held up earns nothing, and the cost of waiting falls on the most expensive thing in the room. The scarce resource is time to power. A site that already has the floor, the halls, the cranes, the trades and a utility tie lets the money go into machines sooner, and a site that has to be built or rebuilt first spends the same dollars on shell, structure and delay. Retrofitting an existing building to 40 kilowatts costs $50,000 to $100,000 per rack in the same guide, and that is the cheaper path only if the building can take it, which is why the floor and the structure come first.
There is a second allocation in a former base or yard. The public already paid to build much of it. The Navy built the base, communications center included, and the company that used a part of it in 2014 paid $6 million to expand. A buyer of an old industrial site is buying structure that someone else's capital already placed, and the price reflects what the site was, not what it could now carry. I did not find the price of any site discussed here, so I cannot say whether the discount is real. That is the diligence question.
It also connects to the earlier piece on who buys a company. A company with idle cash has four uses for a dollar. Putting a dollar into a heavy, powered building is the first of them, reinvesting in the business, and the same buyback-or-build question applies to a data center: build what you can power, or return what you cannot.
The risk runs the other way. A site with local support can still be frozen by a state that has not decided who pays for the electricity. Maine's own council does not report until January 2027, and the permitting calendar can strand the capital as surely as a weak floor. I would price that as time, not as a probability, since I have no basis for a probability.
There is a counterexample I came close to buying. In July 2024 a partner and I toured an industrial building in northern New Hampshire that had been fully built out as a data center. The author's account is that the builder never put clients in it. We were looking at buying it. We did not, and I have not said why.
The public record is thinner than that, and partly different. A public listing for the building shows 12,400 square feet on about three acres, built in 1994, with office space over two levels and rear space for light manufacturing or storage. It sold in February 2023 for $1.35 million, was relisted that September at $1.8 million and was cut to $1.75 million by April 2024. An economic-development blog from around 2011, undated in the copy I could read, reported that a managed-services company planned to open a data center in the town's industrial park, buying a former shipping building, citing the state's low taxes, cheaper real estate and responsive government. A data-center directory lists that company at the same address as operating, with two separate power connections, a 60 kVA diesel generator, redundant air conditioning and a fiber ring. The directory gives no capacity, and says the operator supplied none.
That fits the author's account better than it first looks. A directory entry records that a room is equipped and open for business, not that anyone rents space in it, and nothing I found shows a client list or a customer there. I cannot confirm the account from the record, and the record does not contradict it. I have not confirmed that the building I found is the one we toured, though the address matches.
What the record does give is scale. A 60 kVA generator is, on any ordinary power factor, well under the 120 kilowatts that one Nvidia NVL72 rack draws. A generator is not the same as the building's capacity, and I treat it only as a sign of a small critical load. The point is how far a room like this sits from AI scale. It was a colocation room for small customers, and the machines that would have filled it are of another order.
The lesson is the one in the capital allocation section, seen from the small end. A room fully built out for computers and never filled with clients is capital placed ahead of demand, and what it earned was nothing. The price history is how that shows up: a sale at $1.35 million, then a relisting a half-million higher that had to be cut. My read is that this is the version of the skeptic's bubble warning that happens to small sites first. The use history is the author's account, and I would check it with someone who knows the room before relying on mine.
The third argument is about people, and it is the one that does not show up in a site brochure.
An earlier post on DoAyni cited an estimate that the US data-center construction industry is short about 499,000 skilled-trade workers, with mechanical, electrical and plumbing trades short in 87 to 90 percent of markets, and electrician pay climbing into a range of $72,000 to $150,000. I did not re-check that source for this piece. If it is close, then the limit on building a dense, liquid-cooled hall is not only the floor and the power. It is who can wire it, pipe it and commission it.
Bath is a place where those trades already live. The yard employs roughly 6,000 people. When it hired nearly 600 in 2014, the list was electricians, pipe fitters, outside machinists, welders and tinsmiths, and the company ran its own on-site training center to tailor their skills. Set that beside the AI room: heavy electrical, a liquid cooling loop, large steel supports. My read is that these are close cousins, not identical trades. A shipyard electrician and a high-voltage commissioning engineer are not interchangeable, and I would not claim they are without someone who has done both checking it.
The author's account is that the people were there. For a would-be tenant at Brunswick, that is the point: skilled labor within driving distance of the site, in a state where the base's closure sent thousands of people out of it. The local news reported five years after the 2011 closure that thousands had left Maine and that Brunswick Landing then held about 921 jobs, with a forecast of 1,600. I have no current count.
Two limits matter. First, the yard is a competitor for the same workers, and a data center that bids against a Navy contractor for pipe fitters will raise the price of both. Second, the permanent staff is small. The Jay mill project is described as 800 or more construction jobs and 100 or more permanent ones, so the trades matter most while the building goes up, and far less once it runs. The long-term jobs argument is weaker than the construction one, and a town should be told so.
My read is that staffing and capital allocation are the same argument seen from two sides. Both say that the cost of a data center is not the shell. It is the time and the hands it takes to turn a building into a working machine, and a place that already has both is ahead.
I put the idea to a chief executive I know in the industry, in October 2025, and his reply is worth setting down because it cuts two ways. In my paraphrase, he agreed with the basics: huge buildings, thick floors, cheap power, cheap housing. He was not persuaded by the fiber link to Boston. And he added that if the idea catches on broadly, with a variety of data centers going up in places like this, that would mean bubble time, even though he did not think it a bad idea in itself.
Take the pieces in turn. The fiber objection is the same one the Lawrence Berkeley researcher gave the state council: distance from fiber and population centers. It matters less for a training or batch workload than for one that serves users in real time, and I have no source on how much it matters for either, so I am not weighing it. The housing point I did not check. And the power point runs against what I found: the researcher listed higher regional power costs among the deterrents, and the state's energy commissioner warned that new large loads can raise prices. If Maine power is cheap, it is cheap relative to somewhere, and I would want to know where before using the word.
The bubble point is the one an investor should keep. A single reuse of an old yard or base can be sound, and still be the early example of a crowd. If many sites like it get built on the same logic, the logic itself becomes the thing that is priced, and the sites that arrive last are the ones left holding a heavy building and a power connection nobody needs. My read is that the test is the same as in the first section: is the machine earning, and is the power there? The first piece in this series asked how a company with cash should spend a dollar. The dollar spent on a building ahead of demand is the one a bubble wastes.
I paraphrase the executive without naming him or his company, because I have not asked him whether he wants to be quoted. He has not seen this draft.
If the requirements are cold, weight and power, the same logic reaches past Maine. This is a test for choosing a site, and I am applying it as an investor would, not as a view on any country's policy.
Iceland is the clearest case for power. Its electricity is almost entirely renewable, mostly hydro with geothermal making up the rest, according to a trade report from 2013, and data centers there were estimated to take 5 to 6 percent of national demand in 2022 and 2023, per a 2026 magazine account. The established operators include Verne, Borealis and atNorth. The same account says atNorth agreed in February 2026 to be acquired by the Canada Pension Plan Investment Board and Equinix, so the landlords discussed above are buying there. It also names the limits: new geothermal fields can take 10 to 15 years or more, extraction can lower well pressure, and the landscape is tied to tourism and national identity. The 2013 report put the standard price at 4.5 cents per kilowatt-hour; that is old, and I would not rely on it.
Newfoundland and Labrador is the case where the power exists on paper and the grid does not obviously deliver it. A company called Great North Data proposed a Labrador data center of more than 40 megawatts, citing cold climate and hydroelectricity then priced at about 3.3 cents per kilowatt-hour, according to a CBC report from some years ago. The utility replied that it had no current plans to supply Muskrat Falls power to Labrador's interconnected customers, citing reliability, and provincial law limits connecting specific customers to that power. So the proposal ran into the grid and the law, not the weather. I did not find a current project there.
What the two share with Bath is that the sites are not in the usual markets, and the fiber question comes back. Here the answer is different. The Greenland Connect cable, in service since 2009, links Milton in Newfoundland to Greenland and to southern Iceland, so the two places sit on the same network, and in 2013 the trade report gave about 40 milliseconds of latency from Iceland to the US East Coast. Those figures are dated and I have not seen current numbers. My read is that this makes cold and heavy a spectrum, not a place: Maine is close to the market and short of power, Iceland has the power and is far, and Newfoundland has the power on paper and a grid problem. Each trades a different cost. The author's judgment is that all three make sense for workloads that do not need to be near a user, and I have not tested that against any measurement.
The author's way of putting the fiber question is that the only real limit is physics, and the physics is easy to compute. Light travels about 299,792 kilometers a second in a vacuum and about two-thirds of that in glass fiber, roughly 204,000. On straight-line distances from Ashburn, the round trip in fiber comes to about 8 milliseconds to Bath, 22 to St. John's, 38 to San Francisco and 44 to Reykjavik. Real cable routes are longer than the straight line, so add perhaps a third: about 11, 28, 49 and 58 milliseconds. These are my arithmetic from published distances and a typical fiber refractive index, not measurements, and routers and switches add delay on top. The comparison that matters is with the work itself: an AI model typically takes a good fraction of a second or more to answer, so a network delay of tens of milliseconds is small for chat, training and batch jobs and decisive only for work that is timed in milliseconds, such as trading. I have not sourced the response-time claim and would not print it as fact without a measurement. The author's practice, as someone who has run colocation for customers, is the working rule: put the servers where the customers are close enough that they notice no delay. The author adds an observation from his own network: the servers sit on the same switch, linked by fiber at 10 gigabits, and his laptop reaches them over wifi at 1 gigabit at best, so the last hop, not the distance to the servers, sets what the user experiences. That is a statement about throughput, not delay, and wifi adds its own small delay that I have not measured. The right distance is whatever the customer's work tolerates, which is why a colocation room in Ashburn served one set of customers and a cold site in Maine could serve another.
For an investor, the lesson is that the scarce asset is often not the chip. It is a building that can carry the load, a power connection that can feed it, and a permit. Underwrite the interconnection and the permitting calendar before the brochure, and treat time to power as the cost that matters most, since the building is a small share of what the machines cost and the machines earn nothing until it works. In Maine the calendar has a date on it: the council reports in January 2027, and the moratorium's proposed end was November 2027.
For a founder, the question to ask a site is the one a naval architect would ask: what does the floor carry, where does the power come from, and who is within driving distance that can wire and pipe it?
And for anyone reading a region, look at what the place was built for. A town that spent a century building warships on a river is a town with heavy floors, big halls, skilled trades and a grid hookup that someone paid to build. Whether it becomes a data center depends on the power, not the pedigree.
A note on how this was made. This piece was drafted with Claude, the AI model made by Anthropic. The author points, the model drafts from sources it can show, and the author decides what is true and what stays. Where the piece says "my read" or "the author's," that is judgment and not reporting. Figures for rack power, weight and cost come from trade guides and one press account, not from the manufacturer's own datasheet, and I did not check them against it.
A note on who publishes this. The author runs Osparna, a diligence firm that also makes early investments, has worked inside a data center operation, runs large AI models on his own hardware and so is an operator and not only a director or investor, and once planned to place servers at Brunswick Landing, so the author is a former prospective tenant of the site the piece discusses. This piece is published on Osparna and on Oluwadi, which belong to the same company. The author is not a neutral observer of this subject.
Sources: Uptime Institute 2024 Global Data Center Survey, as summarized by Upsite Technologies; Introl, "High-Density Racks: 100kW+ Designs for AI Data Center Infrastructure"; The Register, Nvidia DGX GB200 NVL72 (Mar. 21, 2024); General Dynamics Bath Iron Works, manufacturing facilities page and strategic investment projects page; Wikipedia, "Bath Iron Works" (employment, ownership); International Association of Machinists, "Bath Iron Works Announces 600 New Jobs, Training Center" (May 2014); News Center Maine, "Brunswick Naval Air Station closing, five years later" (2016); DoAyni, "Data Centers Have Roughly 499,000 Unfilled Skilled-Trade Jobs" (Sept. 8, 2026; its underlying source not re-checked); Brunswick Landing, "Why Choose Brunswick Landing"; Bangor Daily News, Oxford Networks data center expansion (Sept. 17, 2014); Data Center Dynamics, former Maine Yankee site (Sept. 2025); Maine Data Center Advisory Council, July 9 meeting summary; Maine Governor's Office, LD 307 decision (Apr. 24, 2026) and Executive Order 5 (Apr. 29, 2026); Maine Morning Star and WABI, veto override fails (Apr. 29, 2026); Thompson Hine, "Vetoed, Not Dead" (2026); The Energy Mag, Jay project (Apr. 2026); TechTarget, "Selecting raised floors, panels for the data center"; Data Center Dynamics, Equinix direct-to-chip liquid cooling at 100+ data centers (2023); Data Center Knowledge, NTT DATA 115 MW and $10 billion AI plan (Mar. 13, 2026); Data Center Knowledge, "Iceland's Renewable Power Play" (Oct. 15, 2013); The Parliament Magazine, "Could Iceland become Europe's AI powerhouse?" (2026); CBC News, Great North Data and Muskrat Falls power (undated in the copy read); Wikipedia, "Greenland Connect"; U.S. EPA, Philadelphia Navy Yard site spotlight; Wikipedia, "Norfolk Naval Shipyard"; Slashdot, "Project Cuts Emissions by Putting Data Centers Inside Wind Turbines" (Dec. 7, 2023, citing CNN); public real-estate listings for the New Hampshire industrial building (size, price history; retrieved Sept. 30, 2026); New Hampshire economy blog, report on a managed-services company planning a Littleton data center (c. 2011, undated in the copy read); DataCenterMap, Littleton listing; NTT Ashburn campus figure as cited in an earlier Oluwadi post.