The $3.2 Billion AI Data Center and the Accountability Gap

Introduction

A fire at the unfinished Lake Mariner data center in Somerset, New York, revealed more than a local construction failure. It exposed a recurring accountability problem in the AI infrastructure boom: the more companies involved in a project, the harder it may be to determine who is responsible when something goes wrong.

Firefighters reportedly found no working alarm or suppression system, along with three hydrants that were not functioning. Safety data sheets that responders were legally entitled to consult apparently burned in the fire, leaving crews to confront heavy smoke without knowing which chemicals might be involved. No one was injured, and construction continued, but the incident raised an important question: who is accountable for the next safety failure?

A project divided among several companies

Lake Mariner is a roughly $3.2 billion campus built on a former coal mine site. Its corporate structure is unusually layered:

• TeraWulf owns and operates the data center, while the land is leased from a company owned by its chief executive;
• UK-based Fluidstack is expected to run the facility;
• Google holds warrants for a future 14 percent equity stake and has agreed to guarantee Fluidstack’s lease payments;
• Anthropic and other AI companies provide the demand for the computing capacity.

TeraWulf later said it was responsible for operational safety, emergency preparedness, required safety systems, and coordination with local responders. The company said it had added Knox boxes, additional hydrants, and emergency “go-bags” containing safety data sheets after an after-action review. However, the local fire chief said in August that, as far as he knew, the hydrants were still dry and that he had not seen the work completed.

Jobs promised, jobs delivered

Employment is another source of local frustration. A 2024 planning presentation indicated that the completed facility could employ between 35 and 40 people. That figure is far below the 165 permanent jobs promised in 2019, when the project applied for a power discount under the same overall plan. The New York Power Authority reviews compliance annually and can adjust benefits if commitments are not met.

The gap reflects a broader feature of data centers. Construction may create substantial temporary employment, but once a facility is operating, the permanent workforce can be small. Without a serious commitment to train local residents for operations and maintenance, a campus consuming hundreds of megawatts may generate far fewer long-term jobs than communities expect.

The wider costs are harder to allocate

A large, continuously operating data center can increase costs across the electricity system, including grid upgrades, congestion management, and backup generation during periods of high demand. If those expenses are not fully assigned to the new load, they may be spread across customers who receive no direct benefit from the facility.

Anthropic has pledged to cover electricity price increases linked to its data centers, including grid infrastructure costs, and to procure new generation matching its energy demand. Lake Mariner is included in that commitment. Yet the pledge addresses electricity pricing; it does not by itself establish who will verify noise limits, clean-energy claims, or on-site safety performance.

Residents have reported a constant low-frequency hum as the campus expands. TeraWulf’s own energy language has also changed. An earlier report described Lake Mariner as powered by 95 percent zero-carbon energy. Later documents used the more qualified phrase “low- and zero-carbon” energy, while a separate disclosure cited 91 percent low-carbon energy. The wording does not itself prove a violation, but it shows why consistent and independently verifiable reporting matters.

Why the structure matters

The Lake Mariner case separates accountability into legal, operational, financial, and reputational layers. Contracts can divide ownership, leasing, guarantees, and computing demand, but they cannot answer residents’ questions about emergency readiness, noise, jobs, or environmental performance.

As AI companies increasingly obtain computing capacity through leases and third-party operators, customers may lack direct control over the sites they depend on. Future projects seeking public support will need to disclose not only investment and power figures, but also the responsible entity, emergency procedures, employment commitments, energy sources, and mechanisms for independent verification.

Source: Ars Technica AI