AI Data Centers Have an Accountability Problem for Builders

The AI infrastructure race is building a new class of industrial projects faster than responsibility can be assigned. A $3.2 billion data center campus in New York shows why the hard part is no longer only getting megawatts and GPUs online. It is making sure somebody can answer for safety, power costs, environmental claims, and community impact when the ownership chain is deliberately distributed.

The AI buildout is an accountability stack

The Lake Mariner campus in Somerset, New York, is a useful case study because no single company cleanly maps to the whole project. Ars Technica reports that TeraWulf owns and operates the site, while leasing the land from a company owned by its own chief executive. Fluidstack is expected to run the facility. Google holds warrants for a potential 14 percent stake and guarantees lease payments. Anthropic is among the companies whose compute demand the campus is intended to serve.

That structure may be efficient for financing and deployment, but it creates an accountability stack with gaps between layers. The owner controls the physical site. The operator controls day-to-day execution. A financial backstop has incentives tied to project completion. The compute customer cares about capacity and contractual performance. Local authorities care about emergency access, noise, jobs, water, and grid effects.

Those interests overlap, but they are not identical. When the project succeeds, every layer can claim a role. When something goes wrong, each layer can point to the boundary around its own role. For AI builders, this is the infrastructure equivalent of an undocumented API contract: every component works in isolation until an incident crosses the seam.

The distinction matters because AI capacity is increasingly being assembled through leases, special-purpose entities, cloud partnerships, and power agreements rather than a single vertically integrated owner. The result is a system that can move capital quickly while making operational responsibility harder to inspect.

Safety cannot be delegated into a contract

The immediate warning from Lake Mariner was physical. A fire broke out in an unfinished building in June. Firefighters reportedly encountered no working alarm, no suppression system, and three dead hydrants. They also lacked reliable access to safety sheets identifying the chemicals producing heavy smoke. TeraWulf later said it was responsible for operational safety and emergency preparedness, and described plans involving Knox boxes, additional hydrants, and safety-document “go-bags.” But the local fire chief told Ars Technica in mid-August that, as far as he knew, the hydrants were still dry.

The lesson is not that every AI data center is unsafe. It is that a signed allocation of responsibility is not the same as verifiable readiness. A contract can say who owns emergency preparedness. It cannot, by itself, prove that a fire crew can enter the building, identify hazards, shut down equipment, or reach a live hydrant at 3 a.m.

This is a familiar problem for software teams. A vendor may promise uptime, a cloud provider may define a shared-responsibility model, and an internal platform team may own the deployment. None of those labels substitute for a tested incident path. The AI version needs the same operational discipline: named owners, machine-readable inventories, drills, immutable records, and evidence that controls work under stress.

For an AI campus, that evidence should include more than PUE and commissioning milestones. It should cover emergency access, suppression readiness, chemical inventories, generator behavior, cooling failure modes, network isolation, physical security, and communication with local responders. The customer buying compute should be able to inspect the evidence, not merely rely on a landlord’s summary.

Power claims need an evidence layer

Lake Mariner also exposes the difference between a sustainability claim and a sustainability control. TeraWulf previously described the site as powered by 95 percent zero-carbon energy. Later company disclosures used broader “low-carbon” language and cited a lower figure. The facility draws from a regional grid, while a portion of its power comes through a high-load-factor program that is not itself a renewable-energy product.

Anthropic has committed to covering electricity price increases caused by its data centers and to procuring new generation to match its needs. But the reporting raises a sharper question: who verifies the environmental and safety claims at a leased site operated by another company? Google, meanwhile, has one of the industry’s stricter standards, matching clean power to usage in real time. Whether a market-procured blend meets that standard is not obvious from a generic “clean energy” label.

This is where AI infrastructure needs an evidence layer between marketing and reality. Builders already understand the pattern from observability. A dashboard that says a workflow is healthy is not enough if it does not expose latency, error rates, retries, and the trace behind a failed run. Energy claims need equivalent telemetry: hourly load, generation mix, contractual instruments, water consumption, noise measurements, incident logs, and the party responsible for remediation.

The same principle applies to economics. Data centers can create jobs during construction while delivering relatively few permanent roles. They can also push grid and backup-generation costs onto other ratepayers. If the benefits are public and the costs are socialized, the project needs public reporting that is granular enough to be audited.

What builders should take from this

The first takeaway is to treat infrastructure vendors as part of the system, not as black-box dependencies. When evaluating a GPU host, colocation provider, or AI cloud, ask for the control plane behind the promise: ownership, operators, escalation paths, audit rights, recovery targets, and evidence from real tests.

Second, separate capacity from accountability. A contract for megawatts does not guarantee safe operations, clean power, or resilient service. Put those requirements into measurable acceptance criteria and ongoing reporting, with consequences when evidence disappears.

Third, design for boundary failures. The dangerous incident is often not inside one component. It is between owner and operator, customer and landlord, grid and campus, or cloud provider and tenant. Run tabletop exercises that deliberately cross those boundaries.

Finally, make responsibility observable. The next generation of AI infrastructure will be too capital-intensive and too distributed for trust to rest on brand reputation. It will need the same thing production agents need: explicit permissions, independent monitoring, durable logs, and a clear human owner when the system behaves outside its intended envelope.

The AI buildout is often described as a race to add compute. Lake Mariner suggests a more important race is underway beneath it: the race to build infrastructure whose responsibilities remain legible after the contracts are signed, the power is switched on, and something unexpected happens.