The Campus

Picture a stretch of farmland in Grundy County, Illinois, sixty miles southwest of Chicago. The soil is dark and flat. Interstate 80 runs along the northern boundary. Seven miles to the west, the Dresden Generating Station has been splitting atoms since 1970, its two boiling water reactors producing 1,845 megawatts of carbon free electricity, enough to power two million homes. The land sells for $13,565 per acre. In January 2026, a Denver infrastructure developer paid $150,146 per acre for 343 acres off the interstate, an eleven fold premium over farmland prices, to build a 1 gigawatt data centre campus called Morris Technology Park. That transaction tells you everything about where this industry is heading.

This module is the synthesis. We take every data point from the previous five modules, every economic model, every regulatory timeline, every vendor assessment, and design three fully specified nuclear data centre campuses at three different scales. Each campus is an independent, self-contained facility with its own dedicated on-site small modular reactor. The geographic clustering in Grundy County is an editorial device to enable scale comparison, not an operational dependency. Atlas Edge is a 50 megawatt edge facility powered by Oklo Aurora microreactors. Atlas Sovereign is a 500 megawatt sovereign cloud campus powered by eight NuScale VOYGR modules. Atlas Hyper is a 950 megawatt hyperscale complex powered by four BWRX-300 units. Together, they represent the full spectrum of nuclear data centre deployment, from the minimum viable installation to the maximum credible build.

Why Grundy County, Illinois?

The site selection is not arbitrary. Grundy County possesses a convergence of infrastructure advantages that no other location in the Midwest can match, and the research confirms every element of that thesis.

The existing nuclear infrastructure is the anchor. Dresden Generating Station sits on 953 acres in Goose Lake Township, approximately 7.5 miles west-southwest of Morris. It employs 718 permanent staff. Its annual payroll is $74 million. It pays $24.5 million annually in property taxes to local jurisdictions. The December 2025 license renewals extended operating authorisation through 2049 and 2051. Constellation invested over $370 million in relicensing capital improvements. In 2023, Dresden generated over 15.2 million megawatt hours of carbon free electricity. This is not a theoretical nuclear market. It is an operational one with 55 years of community integration, workforce expertise, and grid interconnection infrastructure already in place.

Siting Factor	Grundy County Advantage	Comparable Markets
Existing Nuclear Operations	Dresden (1,845 MW), 55 years operational	Few US counties match
Nuclear Workforce	718 full-time nuclear staff + regional talent	Training pipeline established
Grid Infrastructure	ComEd high-voltage transmission network	Mature interconnection capacity
Transportation	I-80 corridor, 60 miles from Chicago	Rail and heavy haul access
State Policy	IL SB 25 moratorium lifted, ICC planning authority	Most nuclear-friendly state
Land Cost	$13,565/acre (farmland baseline)	11x premium for DC-ready land
Water Supply	Illinois River and municipal water systems	Abundant vs. Southwest markets
Community Precedent	Morris Technology Park approved, 1 GW planned	Local government supportive

The Morris Technology Park transaction validates market demand. Tract, which manages over 25 gigawatts of planned data centre capacity across more than 25,000 acres nationally, chose Grundy County for its first Chicago metro area project. The company typically invests $100 to $150 million per site to bring parcels to operational readiness before selling to hyperscale operators. Initial power energisation is expected in June 2028, with full 1 gigawatt load ramp by 2032. The project will generate over 1,000 construction jobs and approximately 350 permanent positions, representing 0.65 percent of the county's 54,000 population.

Not every community welcomes data centres. Naperville's city council rejected a far smaller 36 megawatt facility citing energy consumption and noise concerns. Morris approved the gigawatt scale campus. That local political reality, the combination of nuclear heritage, economic need, and informed community acceptance, is a siting advantage that cannot be replicated easily.

The Power Density Revolution

Before designing the campuses, we need to understand the load they serve. The power density requirements of artificial intelligence workloads have undergone a transformation that most facility planners did not anticipate and still have not fully absorbed.

Enterprise (2018)

Industry Average (2025)

H100 GPU Rack

GB200 NVL72

GB300 NVL72

Power per Rack (kW)

The 2018 enterprise bar is a midpoint of a 10–12 kW range, not a single published figure.

The NVIDIA GB200 NVL72 connects 36 Grace CPUs and 72 Blackwell GPUs in a single liquid cooled rack consuming 100 to 120 kilowatts at full load. Each rack delivers 1.44 exaFLOPS of FP4 tensor compute. The engineering specifications are demanding: dual 415 volt three-phase feeds via 160 ampere IEC 60309 connectors, 4/0 AWG copper wire, deionised water cooling with 30 percent propylene glycol at conductivity below 10 microsiemens per centimetre, floor loading of 21 kilonewtons per square metre. The newer GB300 NVL72 reaches 140 kilowatts per rack. Industry projections suggest average global rack density will rise from 12 kilowatts today to 30 kilowatts by 2027, but hyperscale AI facilities are designing for 100+ kilowatts per rack as standard, with some planning for 200+ kilowatts.

Meta's AI data centre in Lebanon, Indiana illustrates the scale. The facility carries a $10 billion price tag. It draws 1 gigawatt of continuous power. It houses 300,000 to 500,000 GPUs. Primary AI training workloads consume 400 to 500 megawatts. Inference adds a further 200 to 300 megawatts. These are the loads that our three campuses are designed to serve.

Workload Type	Power per Rack	Cooling Method	Racks per MW	Campus Scale
Enterprise / Colocation	10 to 20 kW	Air cooled	50 to 100	10 to 50 MW
AI Inference	40 to 80 kW	Hybrid (air + rear door)	12 to 25	50 to 200 MW
AI Training (H100)	60 to 100 kW	Direct liquid cooling	10 to 16	100 to 500 MW
AI Training (GB200/300)	100 to 140 kW	Full liquid cooling	7 to 10	200 to 1,000 MW
Next-Gen (2028+)	150 to 200+ kW	Immersion / advanced liquid	5 to 7	500+ MW

The three campuses, compared

Here are the three campuses side by side — the minimum viable edge installation, the mid-scale sovereign cloud campus, and the maximum credible hyperscale complex. Each is an independent, self-contained facility with its own dedicated on-site reactor; the shared Grundy County siting is an editorial device for scale comparison, not an operational dependency. Select a scenario to compare its reactor architecture, capital, timeline, insurance and buyer.

What separates the three is not only size. Each architecture buys a different kind of resilience, and each phases its capital differently — the reasons the comparison above is a set of engineering choices, not just a price ladder.

Sequential commissioning changes the capital math

Both larger campuses earn revenue while they are still being built, because their reactors commission in sequence rather than all at once. That is what makes the multi-billion-dollar commitments manageable: expenditure is matched to revenue over the build.

Black start is Atlas Sovereign's defining advantage

The NuScale design connects an Auxiliary AC Power Source directly to its 13.8 kilovolt generator buses, so in a total grid collapse the plant can start from cold conditions, reach synchronisation in roughly 8 to 15 minutes, and progressively restore the data centre load — with no diesel fuel storage required. The steam bypass system lets the reactors hold at zero external load indefinitely during a grid emergency, then ramp back up on demand. Combined with NuScale's NRC-certified site-boundary EPZ, which lets the reactors sit directly adjacent to the data halls they serve, it is one of the most resilient power architectures conceivable for a campus of this scale. Atlas Hyper's four BWRX-300 units, by contrast, are independently licensed standalone reactors — an accident in one cannot propagate to another — but they need an external source to black start.

Atlas Sovereign's capital cost of $4.2 to $5.7 billion at first-of-a-kind reactor costs covers eight NuScale VOYGR modules producing 616 megawatts of gross electrical output for 500 megawatts of net IT load. What makes that figure financeable is not the headline number but its shape: sequential commissioning spreads the outlay across roughly eight years and matches capital deployment to revenue. The campus can begin earning from its first two modules while the remaining six are still being installed, so the model is far less exposed to a single up-front capital shock than either the microreactor or the hyperscale design.

“Inclusion of a power plant utilizing NuScale Power Modules in a micro-grid reduced potential power interruption to mission-critical facilities from approximately eight hours per year to one minute, and when the macro-grid was also included, power disruption was reduced to approximately one-hundredth of a second per year.”Oak Ridge National Laboratory (2020)

The Insurance Question

Every nuclear data centre campus faces an insurance framework that has not yet fully adapted to SMR deployment. The Price-Anderson Nuclear Industries Indemnity Act requires each reactor site with capacity above 100 megawatts electrical to carry $500 million in primary liability insurance through the American Nuclear Insurers pool. A secondary tier provides approximately $15 billion in additional coverage through retrospective premium assessments across all operating US reactors.

For sites below 100 megawatts electrical, the liability requirement scales from $4.5 million to $74 million based on thermal power output and local population density. Atlas Edge, with four Oklo Aurora units at 60 megawatts total, would fall below the $500 million primary threshold, significantly reducing insurance costs. Atlas Sovereign and Atlas Hyper, exceeding 100 megawatts, require full Price-Anderson coverage.

The critical insurance transition occurs when nuclear fuel arrives on-site. Before fuel loading, the campus operates under conventional builder's risk insurance covering standard construction perils. After fuel loading, the facility transitions to full Price-Anderson liability requirements and NEIL property insurance obligations. Insurance professionals emphasise that SMR projects require early underwriting engagement during conceptual design, not during final financing. The review and binding process for ANI nuclear liability coverage can require several months.

Battery Energy Storage: The Essential Complement

Every campus includes battery energy storage, and the economics have reached a tipping point. Installed capital costs for complete utility-scale BESS have fallen to approximately $125 per kilowatt-hour for four-hour systems. The levelised cost of storage has reached approximately $65 per megawatt-hour, achieving cost parity with many conventional peaking generation sources.

For nuclear data centre campuses, BESS serves three functions. First, frequency regulation: nuclear reactors operate optimally at steady-state baseload. Data centre loads, while more stable than most industrial loads, do exhibit transients during training job startup and completion, maintenance windows, and facility ramp-up. BESS absorbs these transients, maintaining system frequency within the statutory ±0.1 Hz of nominal 60 Hz without requiring the reactor to load follow. Second, bridge power: during scheduled module maintenance or refuelling outages, BESS provides hours of backup allowing orderly load management. Third, revenue generation: during periods when data centre load is below reactor output, BESS can store excess generation and sell frequency regulation or energy arbitrage services to the grid, generating revenue that offsets the storage capital cost.

The Texas BESS market provides a reference case. ERCOT added 6 gigawatts of new battery capacity during 2025 alone, bringing total operational capacity to 13.9 gigawatts. The Calpine Nova Power Bank in California deployed 680 megawatts of power capacity. Its energy capacity is 2,720 megawatt-hours. Total project cost was approximately $1 billion. Vistra's Moss Landing reached 630 megawatts and 2,500 megawatt-hours. It was built across three construction phases. Construction timelines have compressed to 12 to 16 months for large projects. BESS is no longer experimental infrastructure. It is mature, financeable, and deployable at the scale our campuses require.

The Unresolved Questions

Intellectual honesty requires acknowledging the genuine gaps in the nuclear data centre thesis. These are not arguments against the investment case. They are risks that must be priced, monitored, and managed.

Fuel supply for the microreactor tier. Atlas Edge depends on HALEU, the high-assay low-enriched uranium the Oklo Aurora requires, and HALEU is the most constrained link in the fuel chain. Atlas Sovereign and Atlas Hyper avoid this entirely because both run on standard low-enriched uranium — but the microreactor campus carries a fuel-availability risk the two larger designs do not.

The regulatory timeline is not uniform. Only NuScale holds an NRC design certification (January 2023). The Oklo Aurora and the BWRX-300 are both still in pre-application review, so their paths to a construction permit are estimates rather than commitments. If NRC throughput slips while the agency works through an unprecedented pre-application caseload, the 6 to 10 year Atlas Hyper schedule moves to the right, and the microreactor path — the least regulatorily mature — is the most exposed.

Co-location rules are unsettled. The FERC investigation opened on February 13, 2026 is explicit that the tariff treatment of a reactor co-located with a large data centre load in PJM territory is not yet resolved. Every campus here should carry that uncertainty in its development timeline, and insurance underwriters will want the question settled before they bind coverage.

Grid interconnection is the fallback, and it is the thing being escaped. Each campus is designed to island, but a first-of-a-kind nuclear build can slip and a data centre cannot wait for power. The realistic fallback is a grid connection — precisely the 3 to 7 year queue the campus was built to avoid. The prudent developer secures a provisional grid position early even while pursuing the nuclear path, and prices the cost of holding both.

The Decision Matrix

Not every operator should build every campus. The choice depends on scale, capital access, timeline, and risk tolerance. This matrix maps the three campuses to the actors most likely to build them first.

	Atlas Edge (50 MW)	Atlas Sovereign (500 MW)	Atlas Hyper (950 MW)
Primary Actor	Defence / Sovereign Cloud	Sovereign Wealth Fund + Colo	Hyperscaler (AMZN, MSFT, GOOG, META)
Capital Required	$583M to $848M	$4.2 to $5.7B (FOAK)	$16.6 to $19.5B
Reactor Technology	Oklo Aurora (HALEU)	NuScale VOYGR (LEU)	BWRX-300 (LEU)
Fuel Supply Risk	Higher (HALEU constrained)	Low (standard LEU)	Low (standard LEU)
NRC Design Certification	Pre-application	Certified (Jan 2023)	Pre-application review
EPZ Advantage	Expected (small source term)	Approved (site boundary)	Expected (under review)
Black Start	Design dependent	Yes (AAPS, first nuclear)	No (external source needed)
Timeline to Power	6 to 8 years	5 to 8 years (phased)	6 to 10 years (phased)
Risk Profile	Higher (FOAK microreactor)	Moderate (certified design)	Moderate (construction proven at Darlington)
Learning Curve Benefit	Limited (small fleet)	High (modular manufacturing)	High (sequential units)

“Advanced reactors will be available in various sizes to match the data center power demands and can be built in a phased approach to power the data center blocks as they become available.”EPRI Advanced Reactor Roadmap

The Convergence

This series began with a lightning arrestor failing on a 230 kilovolt line in Northern Virginia, dumping 1,500 megawatts of data centre load off the grid in milliseconds. It ends here, with three campuses that make that vulnerability obsolete.

The arithmetic is straightforward. Grid interconnection in the markets where data centre demand is greatest takes 3 to 7 years. PJM capacity charges have risen from $28.92 to $329.17 per megawatt-day in two auction cycles, with the 2027/28 auction hitting the FERC price cap while still leaving a 6,623 megawatt reliability shortfall. A 500 megawatt campus pays $60 million per year in capacity charges alone, accumulating roughly $750 million in net present value over 20 years. The grid literally does not have enough committed capacity to meet projected demand, even at the maximum price regulators will allow.

Nuclear changes that equation. A NuScale VOYGR campus with eight modules provides 24/7 carbon free electricity at a fixed cost for 20 or more years. It eliminates capacity charges entirely. It can operate islanded from the grid, maintaining power through events that would take any grid-connected facility offline. It can black start without external power. And with NuScale's NRC-certified site boundary EPZ, it can sit directly adjacent to the data halls it serves.

The regulatory path exists. Part 53 provides technology-inclusive licensing effective April 2026. Executive Order 14300 mandates 18 month licensing timelines. The ADVANCE Act cut NRC fees by 53 percent. Illinois has lifted all nuclear moratoria and established the ICC's authority to mandate nuclear in utility resource plans. The vendor landscape includes designs that are NRC-certified, under construction, and backed by billions in hyperscaler commitments.

The question for every data centre operator, infrastructure investor, and sovereign wealth fund is no longer whether nuclear power makes sense for data centres. The question is which of these three campuses they build first, and how quickly they start.