Inside a $3.2 Billion AI Data Center: Google, Anthropic, TeraWulf, Fluidstack, 500 MW, and the Accountability Problem
- 48 minutes ago
- 6 min read

Lake Mariner in Somerset, New York, is becoming a useful case study in how modern AI infrastructure is financed and operated. The campus is described as a $3.2 billion buildout with capacity reaching roughly 500 MW, but the companies associated with it do not all control the same layer of the project.
TeraWulf owns and operates the physical data-center infrastructure. Fluidstack holds long-term HPC leases and sits in the compute-delivery chain. Google provides credit support for specified Fluidstack lease obligations and holds warrants that current reporting says could translate into a future stake of about 14% in TeraWulf. Anthropic is a major source of the AI compute demand the infrastructure is being built to serve.
That structure became more consequential after a fire broke out in an unfinished Lake Mariner building in early June 2026. No injuries were reported, but the local fire chief said responders encountered heavy smoke while lacking information they expected to have and found no functioning alarm or suppression system and three hydrants without usable water. TeraWulf later stated that it is responsible for operational safety and emergency preparedness at the campus and described corrective measures implemented after an after-action review.
The issue is therefore broader than a single construction incident. Lake Mariner illustrates how physical responsibility, commercial tenancy, financial risk, AI demand, power-system exposure, and reputational accountability can sit with different organizations even when they are all economically connected to the same AI campus.
··········
WHO ACTUALLY CONTROLS LAKE MARINER.
Ownership, leases, credit support, and compute demand are distributed across different corporate layers.
Entity | Direct role | Exposure and responsibility |
|---|---|---|
TeraWulf | Owns and operates Lake Mariner data-center infrastructure and leases critical IT capacity and related services. | Physical construction, site operations, power delivery, maintenance, safety systems, and emergency preparedness. |
Fluidstack | Long-term HPC tenant and compute provider using contracted Lake Mariner capacity. | Lease obligations and commercial delivery of compute; not identified as the party responsible for Lake Mariner physical safety. |
Backstops specified Fluidstack lease obligations and holds 73.58 million TeraWulf warrants. | Credit support and potential equity exposure; the arrangement does not make Google the physical operator of the campus. | |
Anthropic | Major AI compute customer whose demand supports the economic case for the infrastructure. | Compute-demand and reputational exposure; no direct responsibility for the site's fire-safety systems has been established. |
Local government and fire services | Host community, permitting environment, and emergency response. | First-response and community impact without corresponding control over the commercial contracts financing the campus. |
TeraWulf's SEC filings make the financial chain unusually explicit. Under the Google recognition agreements, a qualifying Fluidstack payment default or insolvency event can give Google the option to pay the relevant termination fee or pay rent and assume the lease. In return for providing the backstop, Google received warrants to purchase 73.58 million TeraWulf shares at an exercise price of $0.01 per share.
TeraWulf also reported that Lake Mariner had 81 MW of revenue-generating critical IT capacity on June 30, 2026 and reached 102 MW after delivery of CB-3 in early July, while another 336 MW across CB-4 and CB-5 remained under construction. That progression shows that the accountability questions are appearing while the campus is still scaling, not after the buildout is complete.
··········
THE FIRE EXPOSED THE OPERATIONAL GAP.
A construction-site incident turned a complicated corporate structure into a practical first-responder problem.
The early-June fire occurred in a building that was still under construction, and nobody was injured. The Barker Fire Department chief nevertheless described a response in which firefighters entered heavy black smoke without being able to identify some of the chemicals involved because the safety documentation they expected to consult was unavailable. He also reported that the building lacked a working alarm and suppression system and that three hydrants were not supplying usable water.
TeraWulf subsequently told Ars Technica that it is responsible for operational safety and emergency preparedness at Lake Mariner, including required systems and coordination with first responders. The company said an after-action review led to Knox boxes, additional hydrants, and portable safety-data-sheet packages.
The remaining problem is the gap between corporate remediation statements and what emergency personnel can verify on the ground. When the fire chief was contacted again in mid-August, he said a Knox box program was being developed but that, as far as he knew, the hydrants were still dry. That does not establish that every corrective measure failed; it does show why accountability has to include verifiable operating conditions rather than only contractual responsibility.
For a hyperscale AI site, emergency preparedness is not a peripheral issue. Dense electrical infrastructure, battery systems, cooling equipment, construction materials, backup power, and rapidly changing building configurations can all alter the risk profile faced by local responders. A site can have sophisticated financing and state-of-the-art compute while still depending on basic physical controls such as alarms, suppression, hydrants, access systems, and current safety information.
··········
500 MW MAKES ACCOUNTABILITY AN INFRASTRUCTURE ISSUE.
Power, local costs, jobs, and environmental claims extend the responsibility chain beyond the data hall.
A campus capable of drawing up to 500 MW is not only a technology asset. It is a long-duration industrial load whose effects extend into transmission constraints, generation planning, electricity pricing, emergency services, land use, noise, tax policy, and community expectations. Those effects can persist even when the entities financing the buildout, leasing the capacity, and consuming the compute are different companies.
The employment picture illustrates the mismatch between scale and local benefit. Documents cited in the investigation say a 2024 planning presentation estimated roughly 35 to 40 jobs at full buildout, while an earlier 2019 project scope associated with the site had referenced 165 permanent jobs. The figures come from different stages of the site's evolution, but the contrast explains why local officials are asking what a community receives in exchange for hosting hundreds of megawatts of new demand.
Accountability layer | Primary mechanism | What remains difficult to verify |
|---|---|---|
Fire and emergency readiness | TeraWulf says it is the responsible operator for safety and emergency preparedness. | Whether every corrective measure is present, operational, and known to local responders at a given time. |
Lease-payment risk | Fluidstack is the tenant; Google provides a contractual backstop under specified conditions. | Financial protection does not itself define who is accountable for physical or environmental performance. |
Compute demand | Anthropic is a major customer driving capacity requirements. | How far a compute customer can independently audit site-level safety and environmental claims made by upstream providers. |
Grid costs | Anthropic has committed to cover specified electricity-price and grid-infrastructure impacts associated with its data centers; Lake Mariner is reported to be covered. | The commitment does not automatically resolve noise, safety, water, land-use, or every other community externality. |
Clean-energy claims | TeraWulf points to the regional generation mix and low-/zero-carbon characteristics of the upstate New York grid. | Grid-average composition is not the same as a dedicated clean-power contract, and public independent verification across the corporate chain is limited. |
Community engagement | TeraWulf is the site operator interacting with local authorities; other stakeholders participate through contracts and financing. | No single stakeholder necessarily answers for every local consequence of the project. |
Energy claims are particularly sensitive to definitions. TeraWulf has described the site using low- and zero-carbon terminology and has pointed to the generation mix of the upstate New York grid. The investigation notes, however, that the company does not buy renewable energy certificates to recharacterize Lake Mariner's grid electricity. A regional grid with substantial hydro and nuclear generation is materially different from a facility being contractually supplied by a dedicated renewable resource.
New York's policy response also shows the timing problem. State lawmakers passed broader data-center legislation in 2026 and the governor later imposed a narrower moratorium on new hyperscale permits, but Lake Mariner's existing permissions mean the current buildout is not stopped by those measures. Regulatory frameworks can therefore change while projects already in the construction pipeline continue scaling.
··········
WHY THIS MODEL MATTERS FOR THE AI BUILDOUT.
Lake Mariner shows how hyperscale capacity can separate economic participation from direct operational control.
The structure is not unique to Lake Mariner. AI infrastructure increasingly combines specialist data-center developers, neocloud tenants, hyperscaler credit support, chip suppliers, model developers, utilities, lenders, and local governments. This architecture can accelerate construction because each participant supplies a different scarce resource: land and power, data-center execution, customer demand, investment-grade credit, accelerators, or capital.
The same specialization can make responsibility harder to map. A financial guarantor may have enormous economic exposure without operating a building. A model developer may consume the compute without controlling hydrants or suppression systems. A physical operator may control site safety while relying on tenants and guarantors to make the project financeable. Local residents experience the power demand, noise, traffic, tax arrangements, and emergency-response consequences regardless of which contract assigns which obligation.
A useful accountability framework therefore starts by identifying the operator of record for each layer rather than treating every company named in a project as equally responsible. Physical safety, cloud service delivery, payment guarantees, compute purchasing, grid interconnection, environmental reporting, and community commitments need separate owners, measurable requirements, and a mechanism for independent verification.
Lake Mariner is not evidence that multi-company AI infrastructure is inherently unsafe. It is evidence that the faster and more financially layered the AI data-center boom becomes, the more explicit the accountability map has to be. Without that map, the party benefiting from a project, the party paying for it, the party operating it, and the party answering when something fails can all be different organizations.
··········
FOLLOW US FOR MORE.
DATA STUDIOS
datastudios.org

