Turning an Old Coal Mine into a New AI Data Center

Western Pennsylvania’s former industrial sites, like Greene County’s Robena Mine, have the special ingredients needed for AI development

Editor’s note: This story has been updated to clarify that the data center at Robena will likely use less water than they have requested to draw. Based on reader feedback, it also has been revised to reduce technical explanations. Finally, I have added a disclaimer that I work in the software industry and support responsible use of artificial intelligence.

Modern data centers made for artificial intelligence barely resemble the server rooms of the past from which they evolved. For one thing, data centers like the one being developed at the former Robena Mine site in Greene County have rows of power-hungry monster servers that require industrial-strength cooling to prevent a fire. For another thing, data centers are so big that they are measured in acres instead of square feet. That’s why the data center at Robena is requesting to draw 18 million gallons of water per day from the Monongahela River to cool it.

If that sounds extreme, it is. Robena will use a massive amount of water for cooling. However, actual water use will be closer to 11 million gallons per day. The developers—Essential Utilities and International Electric Power III LLC—noted in an August press release that the data center will get energy from its own 944-megawatt gas-fired power plant. They would also build a new water treatment plant to cool both the data center and its on-site power source. The data center will be connected to the power grid for backup. From an environmental perspective, the projected data for Robena means that the site may:

• Consume three times more water than the average data center
• Use as much water as a large chemical plant
• Emit CO₂ gases comparable to a major steel mill

The data center’s investors like the Robena site because it has special ingredients necessary for an AI-capable data center: a stable local gas supply, skilled workers, a fiber communications line already running along the river, and the river itself. Despite the Robena site being an ideal location for investors, many people in Greene County have questions about it. This blog is not intended to answer all those questions or explore the data center at Robena in depth. Rather, it is an opportunity to understand why these data centers have suddenly been appearing in our neighborhoods and how they might affect us.

Data needs

The development in Greene County is hardly an anomaly. To meet rising demands, data center development has doubled during the past few years. In March, Synergy Research reported that there are 1,136 hyperscale data centers globally, and 54% of them are in the United States. By 2030, global demand for data center capacity may triple, according to business analyst McKinsey & Co. Larger data companies, such as Google and Microsoft, have made sustainability commitments to offset the environmental effects of their data centers.

However, AI is not the only reason data companies are building more of these sprawling centers. Other reasons for more data centers include:

• Use of 5G networks
• Cloud computing as part of a digital transformation
• Continuous streaming of audio and video
• A rise in e-commerce
• The Internet of Things (IoT)
• Work-from-home arrangements
• 24/7 social media applications
• Real-time gaming systems
• Logs for security and compliance

The AI factor

Nevertheless, the investors have already said they are building the data center at Robena for AI purposes. AI is a branch of computer science that studies and develops software programs capable of performing tasks that typically require human skill. Although applications that run GenAI models—like ChatGPT and Claude—seem to magically produce text, they are just multiplying matrices in an algorithm known as a neural network.

The most common type of AI is narrow AI, which is also called weak AI because there is debate on whether it truly is AI. Narrow AI performs singular tasks, such as facial recognition. Other subsets include machine learning, natural language processing (NLP, used for GenAI applications), computer vision, and deep learning. There are also many theoretical types.

AI has been around for 75 years, but its development was minimal until recently. In 2017, Google published its groundbreaking paper, “Attention Is All You Need,”  about generative pretrained transformers using NLP. By the time it was written, several important factors in the industry had already occurred or were starting to happen that would provide a pathway for rapid AI expansion. They include:

• Using graphics processing units (GPUs) instead of the computer processor (CPU) for heavy calculations. As it turned out, GPUs could run processes in parallel to calculate much faster and carry a very heavy load. Early Bitcoin miners frequently used gaming systems because they had high-power GPUs for their time. But some home-built systems of the early 2010s caught fire because they didn’t have adequate cooling. Engineers now design silicon just for GPUs.
• The growth of data centers, more casually known as the cloud, where powerful GPUs could be used to train AI models at a massive scale. Started by Amazon Web Services (AWS), the cloud is the entire group of global remote data centers that typically lease server space to other organizations through web-based portals.
• Faster networks. Before the COVID-19 pandemic, 1-10 Gbps was normal for an office. Today, network links accelerate speeds up to as much as 400 Gbps, and they continue to get faster. These powerful networks let clusters of GPUs act as one enormous machine.

In terms of energy consumption, large language models (LLMs), the foundation of GenAI, consume the most. In the data center, GPUs process calculations from billions to exaflops (10¹⁸) per second for LLM requests. LLM training is very intense and requires an even greater amount of energy and water.

Modern development

From healthcare to retail sales, AI is greatly increasing the amount of data. The need for more data centers is real. To stop the development of new data centers would mean pulling back on years of technological advancements. Such an idea would cripple the global economy. As a result, many people will eventually live near a data center. One of the considerations when choosing a place to build one is its proximity to people and businesses. Regardless of how fast your network is, being physically closer to a data center reduces latency, which means data gets there quicker.

Western Pennsylvania has seen a lot of data center development in the past few years to be closer to Pittsburgh customers. Here’s the status of those data centers:

Data centers recently considered for Western Pennsylvania

Location	Power usage	Water usage	Notes
Greene County, Monongahela riverfront (IEP and Essential Utilities)	~944 (for data center load in filings/coverage)	18 million gallons per day (MGD)	Reported 18 MGD water plant serves on-site power gen and data center cooling. The 18 MGD exceeds the implied need by ~1.7x, which would be 10.8 MGD.
Springdale, Allegheny (former Cheswick plant)	180	~2.05 MGD	Developer said max 180 MW draw. No public water plan.
Zediker Station, Washington (JLL-marketed)	500–700	~5.7–8.0 MGD	Site marketed as 500–700 MW capable. No public water plan.
Upper Burrell, Westmoreland (TECfusions campus)	up to 3,000 (multi-year)	~34.2 MGD if fully built	Press release says 3 GW over 6 years; 12 MW now. Water plan not disclosed.
Homer City, Indiana (ex-coal plant redevelopment)	TBD (power plant 4.5 GW supply)	TBD	They cite up to 4.4–4.5 GW generation to support “AI” data centers, but no committed IT load yet.

Closing thoughts

I am employed by a software engineering company. Therefore, I support the responsible use of AI. I also applaud sustainability efforts to reduce the effect of data on the environment. As a software engineer recently told me, AI is beneficial, but it is like the “bells and whistles” of application development. AI makes some manual tasks easier, faster, and potentially more accurate, but it is a luxury for many people and businesses.

As options go, Robena is a reasonably good place for a data center. Data centers do not create many long-term jobs, but they take a long time to build and must be maintained. A data center—probably more than one—will eventually be built somewhere in the Pittsburgh area. For Greene County, the economic boost will not be as powerful as when Robena mined coal. Still, a data center is too powerful to ignore.