Applied Digital Polaris Forge 2 Is Rising Fast, but Its Rural AI Bet Faces a Public Test

Applied Digital Polaris Forge 2 has risen from farmland near Harwood, North Dakota, less than a year after construction began. Its first 200 megawatts of contracted computing capacity are scheduled to enter service in phases through early 2027.

That schedule turns an immense construction site into a test of the AI infrastructure boom. Applied Digital must deliver two specialized buildings, connect them to the regional grid, and satisfy a tenant reportedly identified as Oracle. It must do this while nearby communities question who controls the land, infrastructure, costs, and long-term consequences.

The result resembles a small city built for machines. Yet the central issue is not its physical size. It is whether rural power systems and local governments can absorb industrial-scale AI development without losing public confidence.

Applied Digital has already shown that North Dakota can host dense computing infrastructure. Its Polaris Forge 1 campus in Ellendale provides the closest operational reference. Harwood raises the stakes because the new campus has much greater expansion potential and sits beside the Fargo metropolitan area.

Applied Digital Polaris Forge 2 Is Becoming a Physical AI Factory

The most important change is that Polaris Forge 2 has moved beyond a proposal and into the difficult systems-integration stage of construction.

Applied Digital announced the Harwood development in August 2025. The company initially described a campus with two buildings, more than 900 acres under contract, and over 200 expected permanent employees.

The buildings are designed for high-performance computing, or HPC, which combines many processors to perform demanding calculations. AI training and inference now account for a large share of the demand for this specialized capacity.

By July 2026, construction images showed rooftop heat-rejection equipment being installed while the first data halls approached fit-out readiness. Fit-out is the stage when contractors add the electrical, cooling, networking, and support systems required by the tenant’s computing equipment.

These details matter because an AI data center is not simply a warehouse filled with servers. Its electrical and cooling systems must operate as one coordinated machine. A fault in any layer can reduce the usable capacity promised to the tenant.

Applied Digital’s campus portfolio identifies two 150-megawatt buildings under construction at Polaris Forge 2. However, the current tenant commitment covers 200 megawatts of critical IT load rather than the buildings’ entire potential capacity.

Critical IT load measures electricity delivered to computing equipment. It excludes some supporting consumption, including cooling and power conversion losses. The full facility demand will therefore exceed the contracted IT figure when both buildings are operating.

The company expects the first contracted capacity during the second half of 2026. Full delivery of the initial 200 megawatts is expected by early 2027.

Those dates are management estimates, not completed milestones. Applied Digital’s filings say delivery remains subject to construction, permitting, interconnection, and equipment risks.

Still, the project has passed several commitments that distinguish it from speculative land development. It has an identified power provider, construction financing, a long-term tenant agreement, and major structures already visible above the prairie.

The tenant’s commitment also changes the economics. Applied Digital is not building the first phase and hoping that an AI customer eventually arrives. The initial capacity is backed by a lease covering roughly 15 years.

Cass County Electric Cooperative later identified Oracle as the tenant. Applied Digital’s regulatory filings continue to describe the customer more generally as a U.S.-based investment-grade hyperscaler.

A hyperscaler operates computing infrastructure across very large, distributed facilities. Oracle fits that description through Oracle Cloud Infrastructure, but the distinction in public disclosures remains worth preserving. Applied Digital itself has not consistently named the tenant in its securities filings.

The company’s annual filing reported five contracted campuses totaling approximately 1,410 megawatts as of May 31, 2026. Only about 100 megawatts of its AI-focused capacity was operating and producing lease revenue at that point.

That gap defines the current moment. Applied Digital has accumulated a large contracted pipeline, but much of its future business still depends on converting active construction sites into operating infrastructure.

Polaris Forge 2 is therefore more than another project in that pipeline. It is an early test of whether the company can repeat its Ellendale model while working at several campuses simultaneously.

Why North Dakota Became an AI Data Center Target

North Dakota’s appeal comes from an unusual combination of power access, available land, cool weather, and communities seeking durable investment.

AI developers need more than graphics processors. They need locations capable of supplying continuous electricity to thousands of densely packed systems. They also need enough land for substations, backup generators, cooling equipment, security zones, and future buildings.

Large urban markets often struggle to provide that combination quickly. Land costs more, transmission systems face congestion, and permitting can involve several overlapping jurisdictions.

Rural North Dakota offers a different starting point. Its industrial economy already understands large energy projects. Cooler temperatures can reduce mechanical cooling requirements during parts of the year, although weather alone does not solve the heat produced inside AI racks.

Applied Digital also has local operating experience. It began developing North Dakota facilities for cryptocurrency computing before shifting its growth strategy toward AI and HPC hosting. That history gave the company relationships with utilities, construction firms, and public officials before the current AI buildout accelerated.

The transition from cryptocurrency hosting to AI infrastructure is significant. Crypto facilities can consume enormous amounts of power, but modern AI clusters require tighter engineering tolerances and more complicated cooling and networking systems.

AI servers increasingly use direct liquid cooling, which carries heat away from processors through a liquid circuit. Applied Digital says its designs use closed-loop cooling, meaning coolant circulates through a contained system instead of requiring a constant fresh-water supply.

That description addresses one frequent concern, but it does not make the entire site water-free. Construction, sanitation, maintenance, and some cooling configurations can still require water. Public evaluation should separate continuous cooling demand from total facility use.

The larger constraint is electricity. Two hundred megawatts of critical IT load represents an industrial customer unlike almost anything a small community normally receives. The load must remain available around the clock, even when regional demand is already high.

Cass County Electric supplies the local distribution connection, while Minnkota Power Cooperative provides generation and transmission services. The cooperatives have argued that serving the project can benefit existing members when contracts assign expansion costs to the new customer.

Cass County Electric President and CEO Paul Matthys wrote that the cooperative would use contractual protections, guarantees, and dedicated rates. His member update also identified Oracle as the company leasing the Harwood campus.

Those safeguards are central to the project’s public case. Residents should not have to accept vague assurances that a very large new customer will somehow lower costs. They need clear rules showing which party pays for new substations, transmission equipment, and other required infrastructure.

Applied Digital’s financing offers another signal of commitment. In March 2026, a subsidiary issued secured notes intended to fund construction of the first 200 megawatts. The financing is secured by project-related assets and subsidiary guarantees.

The company’s financing announcement also lists several risks. They include construction delays, equipment failures, supply disruptions, customer concentration, financing conditions, and changes in AI infrastructure demand.

That list is not routine fine print in this context. It maps directly onto the practical challenge visible at Harwood. The company must coordinate capital, equipment, labor, power, and tenant requirements on a fixed delivery schedule.

North Dakota can make site selection easier, but it cannot eliminate execution risk. The prairie provides space. It does not automatically provide every transformer, skilled technician, transmission upgrade, or completed data hall the tenant needs.

The Real Contest Is Speed Versus Local Control

Polaris Forge 2 puts the AI industry’s demand for rapid deployment against communities’ demand for transparent, enforceable oversight.

Applied Digital and its tenant have a strong incentive to move quickly. AI companies are racing to secure powered capacity because suitable grid connections and construction equipment remain limited.

Harwood and neighboring jurisdictions operate on a different clock. Local officials must consider zoning, roads, emergency response, noise, drainage, utility obligations, and the project’s effect on future development.

That conflict surfaced before the buildings reached their current scale. Harwood and Fargo disputed which city should annex land around the project and control the associated development.

Harwood has a population of roughly 900, while Fargo has more than 100 times as many residents. The dispute therefore became a debate about whether a small city could manage a project whose physical and economic influence crossed municipal boundaries.

Fargo pursued annexation of land that included the data center area. Harwood officials opposed that effort and argued the property fell within their extraterritorial jurisdiction.

Fargo later withdrew its annexation attempt. That decision reduced one immediate threat to Harwood’s control, but it did not remove the regional questions surrounding roads, public services, and long-term growth.

The confrontation illustrates why AI campuses cannot remain invisible pieces of cloud infrastructure. They occupy real land and reshape planning decisions far beyond their walls.

A large manufacturing plant usually announces what it produces, how many people work there, and which materials enter and leave. AI data centers can be less transparent because customer identities, system layouts, and workloads are treated as commercially sensitive.

Some secrecy is reasonable. Operators must protect tenants, network architecture, and physical security. Yet withholding technical details does not justify obscuring utility costs, environmental permits, tax arrangements, or emergency plans.

The challenge is to define the boundary between protected operational information and facts the public needs for oversight.

Harwood Mayor Blake Hankey has said the buildings will use concrete construction and place noise-producing equipment inside. He has also said the site sits about a mile from town.

Those design choices can reduce disturbance, but they should be verified after operations begin. Cooling equipment, transformers, and generator testing can produce persistent low-frequency sound that behaves differently from ordinary traffic noise.

Water claims also require operational evidence. A closed loop can sharply limit cooling consumption, yet its actual performance depends on the installed design, local weather, maintenance, and backup operating modes.

Power costs demand similar scrutiny. The cooperative says its existing members will be protected, while Applied Digital has agreed to fund infrastructure needed for the project. The public test is whether final rate structures and actual bills support those assurances.

These are not arguments against construction. They are measurable conditions for judging whether the project delivers what its supporters promised.

North Dakota has started building a more formal policy response. A state working group developed a model data center ordinance, and several counties have considered local standards or temporary moratoriums.

A July 2026 legislative memorandum reviewed policies emerging across the country. It identified utility cost allocation, nondisclosure agreements, tax incentives, community benefits, and local land authority as recurring concerns.

The memorandum noted that lawmakers in more than 30 states introduced over 300 data center bills during the first six weeks of 2026. That activity shows Harwood is part of a national governance problem, not a uniquely North Dakotan disagreement.

AI infrastructure is arriving faster than many local codes were written to handle it. Communities must review utility-scale facilities using planning rules often designed for warehouses, farms, and ordinary industrial businesses.

Polaris Forge 2 could help establish better practice. Its scale gives regulators and utilities a real project against which to test disclosure requirements, rate protections, noise limits, and construction agreements.

It could also expose weaknesses. If important obligations remain informal, later projects might cite Harwood as precedent without adopting equivalent safeguards.

Polaris Forge 1 Shows Both the Opportunity and the Concentration Risk

Applied Digital’s Ellendale campus proves that rural AI capacity can reach operation, but it also reveals how much depends on a few tenants and tightly sequenced deliveries.

Polaris Forge 1 is being developed in Ellendale, North Dakota, for CoreWeave. Its contracted plan covers three buildings and 400 megawatts of critical IT capacity.

The first 100-megawatt building entered service in November 2025. Applied Digital announced delivery of the second building, with 150 megawatts, in July 2026. A third 150-megawatt building is scheduled for 2027.

That sequence gives Applied Digital an operating template. The company has experience moving from basic site preparation through commissioning, energization, and tenant handoff in North Dakota.

Commissioning is the formal testing process used to verify that electrical, mechanical, and safety systems perform as designed. It becomes especially demanding in an AI facility because failures can interrupt expensive computing clusters.

Harwood is not a simple copy of Ellendale. Applied Digital has described Polaris Forge 2 as a flatter, single-story design with a more direct electrical path into the data halls.

A repeatable campus model can shorten design work and simplify procurement. However, tenant requirements can still change while construction proceeds. Equipment placement, rack density, liquid-cooling connections, and network layouts must match the customer’s final hardware.

That creates a fundamental tension. Developers want standardized buildings, while AI customers want facilities optimized for rapidly changing processors.

Applied Digital has acknowledged this challenge in its construction updates. The external building can advance while final data-hall details continue to evolve.

The customer concentration is equally important. Polaris Forge 1 depends on CoreWeave, while Polaris Forge 2 is tied to one investment-grade hyperscaler for its contracted first phase.

Long leases provide predictable revenue only when the developer delivers usable capacity and the tenant remains able to meet its obligations. Investment-grade credit reduces counterparty risk but does not eliminate construction or technology risk.

Applied Digital’s broader portfolio also increases the demands on management. As of May 2026, the company reported five named AI campuses with contracted capacity either operating or under construction.

These projects stretch from North Dakota to Louisiana. Each requires financing, utility negotiations, equipment, contractors, and tenant coordination.

Scale can improve bargaining power with suppliers and investors. It can also multiply the effects of delayed switchgear, transformers, generators, cooling modules, or specialist labor.

The company has another legacy to manage. Its older North Dakota facilities continue hosting cryptocurrency-related workloads. That background helped Applied Digital secure powered sites, but AI customers expect a different level of reliability and engineering.

CoreWeave offers a useful comparison. It also emerged from cryptocurrency computing before building a large AI cloud business. Its growth demonstrates how assets and skills from one compute boom can move into another.

The shared history does not make the transition automatic. Cryptocurrency systems can often tolerate interruptions or move workloads in response to energy prices. Enterprise AI customers expect contracted availability and predictable performance.

The technology inside AI campuses also changes faster than the buildings around it. A data center can operate for decades, while processors, cooling requirements, and network designs can shift within a few years.

Applied Digital therefore needs flexibility without paying repeatedly for major redesigns. Its long-term returns depend on making today’s campuses useful for several generations of computing equipment.

This is why Harwood’s progress matters beyond one company. AI capacity cannot grow solely through chip announcements. Developers must convert those chips into dependable systems connected to sufficient power.

The physical layer is becoming a competitive constraint for Oracle, CoreWeave, Microsoft, Amazon, Google, Meta, and other large AI infrastructure buyers. The companies that secure usable megawatts gain room to deploy more compute.

Yet contracted megawatts are not the same as operating compute. Construction progress must eventually produce energized racks, stable cooling, network connectivity, and accepted tenant capacity.

Polaris Forge 2 currently sits between those two states. Its buildings are real, but its most important promised output is still ahead.

What the Construction Photos Cannot Prove

Visible progress supports Applied Digital’s schedule, but it cannot settle questions about grid impact, operating noise, water use, or long-term economics.

Aerial images are persuasive because they make the AI boom tangible. They show steel, concrete, cooling equipment, and electrical infrastructure spread across land that recently looked empty.

However, a completed shell does not prove that a data center is ready for service. The difficult final stages include testing power paths, integrating controls, flushing cooling loops, validating backup systems, and meeting the tenant’s acceptance criteria.

Even a building that receives power can take time to host its intended computing load. The tenant must install servers, networking equipment, storage, and software before the campus contributes meaningful AI capacity.

Applied Digital’s deadlines should therefore be assessed through delivered and accepted critical IT load. Construction percentages and photographs offer context, but tenant-ready megawatts are the stronger measure.

The grid impact will also become clearer only after operations ramp. Large data centers usually increase demand in stages rather than switching on their full capacity at once.

That ramp allows utilities to coordinate new infrastructure. It also delays a complete view of how the load affects regional peaks, generation requirements, and transmission planning.

Cass County Electric’s protections deserve attention during that process. The most useful evidence will be whether existing members avoid project-related rate increases while the cooperative maintains reliability.

Applied Digital and the utility have said required infrastructure costs will not be shifted to ordinary customers. That commitment should remain visible through public rate proceedings, cooperative communications, and actual performance.

Noise can likewise be measured after cooling equipment and transformers operate under load. Preconstruction modeling is useful, but field measurements provide a stronger test.

Backup generators add another concern. They do not run continuously, but periodic testing and emergency use can affect noise and local air emissions. Permit limits and testing schedules should remain publicly accessible.

The closed-loop cooling system needs similar verification. The company’s design appears less dependent on evaporative water consumption than many traditional facilities. Actual withdrawals and discharge practices will show how the system performs across seasons.

The project’s employment claims should be evaluated carefully as well. Construction creates a large temporary workforce, while permanent data center operations employ far fewer people than a comparably sized manufacturing complex.

Applied Digital expects more than 200 full-time positions when the campus is fully operational. The composition of those roles will matter to Harwood and the wider Fargo region.

Technical positions may attract workers from outside the immediate community. Maintenance, security, administration, electrical work, and vendor services can spread benefits more widely, but that outcome depends on hiring and procurement practices.

Tax revenue presents another uncertainty. Large projects can expand a local tax base, but incentives, abatements, and infrastructure obligations determine the net result.

The annexation dispute showed how valuable local governments believe that revenue could become. It also showed that public benefits can be difficult to separate from competition among jurisdictions.

A regional agreement could reduce that tension by clarifying responsibility for roads, emergency services, and future development. Without coordination, each government has an incentive to pursue revenue while shifting costs elsewhere.

There is also a broader demand risk. Current AI spending assumes companies will continue paying for more training and inference capacity over many years.

The long lease reduces Applied Digital’s direct exposure to short-term changes, but it cannot remove the industry’s dependence on sustained AI demand. A customer can remain obligated under a contract while changing its deployment strategy.

Technology efficiency could alter future requirements. More efficient models and chips can reduce computing needed for a specific task, while lower costs often encourage greater overall use.

No single forecast resolves that tension. The safer conclusion is that Applied Digital has secured contractual demand for its first Harwood phase, while the campus’s larger expansion remains dependent on future customer commitments.

The possibility of expanding Polaris Forge 2 toward one gigawatt makes these questions more urgent. The first 200 megawatts establish a base, but later phases would transform the project’s regional footprint.

A right of first refusal gives the current tenant priority over additional capacity. It does not mean that another 800 megawatts have been leased, financed, permitted, or scheduled.

That distinction should remain clear. The operating first phase is the immediate test. The one-gigawatt vision is an option whose feasibility still depends on power, capital, construction, and public approval.

Three Signals Will Decide Whether the Rural AI Model Works

The next judgment should rest on delivered capacity, transparent utility results, and the treatment of any expansion proposal.

The first signal is whether Applied Digital begins delivering Polaris Forge 2 capacity on its stated schedule. Initial service during the second half of 2026 would support the company’s claim that it can repeat the Ellendale construction model.

The stronger milestone will be full delivery of 200 megawatts by early 2027. That would require both buildings and their supporting systems to pass tenant acceptance.

A delay would not prove that the campus has failed. Large infrastructure projects frequently encounter equipment and commissioning problems. It would, however, weaken the argument that Applied Digital can translate its growing contracted portfolio into operations at the promised pace.

The second signal is the project’s effect on cooperative members and regional infrastructure. Cass County Electric has said existing customers will be protected from costs created by the new load.

Future member communications, rate decisions, and reliability data should show whether those protections work. Evidence that the tenant funded required upgrades would strengthen the rural hosting model.

Unexpected cost transfers or service problems would have the opposite effect. They would give other communities reason to demand stricter contracts before approving similar projects.

The third signal is how officials handle further expansion. Applied Digital has designed Polaris Forge 2 for capacity beyond its initial phase, and its tenant holds priority rights covering much of that potential.

Any new phase should trigger a fresh evaluation of grid capacity, roads, water, noise, emergency planning, and local benefits. Officials should not treat the first approval as permanent consent for every later building.

Expansion accompanied by public documentation and enforceable conditions would show that rapid AI development can coexist with local oversight. Expansion negotiated behind closed doors would deepen the trust problem surrounding the industry.

The significance of Applied Digital Polaris Forge 2 will not be measured by its footprint alone. Its real contribution will depend on whether the campus delivers contracted compute without making surrounding communities absorb hidden costs.

For AI users, the outcome sits far behind the interface, but it still matters. More operating capacity can affect cloud availability, competition, and the cost of running AI services.

For local residents, the measures are more immediate: electric bills, roads, noise, jobs, tax revenue, and confidence in public decisions.

Harwood now hosts a practical test of both sets of interests. Watch the delivered megawatts, the utility evidence, and the rules governing expansion. Those signals will reveal whether this machine-scale city becomes a repeatable infrastructure model or a warning about building faster than public institutions can respond.