Steve Hammond built a chiller-free, liquid-cooled data center more than a decade before most of the industry needed one. Now he's warning that AI data centers are creating a problem cooling technology alone can't solve.
Hammond has spent nearly 24 years at what's now called the National Laboratory of the Rockies, the Department of Energy renamed the National Renewable Energy Laboratory to that name on December 1, 2025. Readers who know the Golden, Colorado facility as NREL are looking at the same institution.
How One Data Center Changed the Industry
In 2002, Hammond built the lab's Computational Science Center from scratch, eventually growing it to 100 people. The bigger legacy came later, when he and colleague Ben Kroposki led development of the Energy Systems Integration Facility.
Hammond's team made choices nobody else in the industry had tried. They were first to use component-level warm-water liquid cooling, and first to run a high-performance computing data center with no chillers at all, relying on evaporative cooling instead.
They were also first to capture heat generated by the data center and put it to practical use, a system still melting snow on the facility's plaza today.
"The audacity to do what we did, to go with liquid cooling and to pick Hewlett Packard to partner with us, we took on an incredible challenge and we changed the industry," Hammond said.
That partnership helped launch HP's Apollo line of liquid-cooled computing systems and earned an R&D 100 award. A later collaboration with Johnson Controls cut the facility's water use in half without sacrificing efficiency.
The New Problem: Too Much Power, Too Fast
Hammond's early work solved cooling. The challenge facing AI data centers now is different in scale and in kind.
"When I started at the lab, we thought it was incredible to imagine going up to 50 to 60 kilowatts in a rack," Hammond said. "Today, Nvidia's talking about one megawatt in a rack by 2027."
That's roughly the same physical footprint as a household refrigerator, carrying nearly twenty times the power density Hammond once considered ambitious.
The deeper issue is how AI data centers actually consume that power. Ordinary enterprise computing, email, web traffic, streaming video, spreads load evenly across many independent processes.
Training a large language model does the opposite, pushing nearly every processor from idle to full power within a handful of clock cycles.
"That can be disruptive to the grid," Hammond said. "We need to identify and demonstrate the right infrastructure needed to manage and mitigate those transients."
Building Air Traffic Control for the Grid
Hammond's preferred analogy for the coordination problem is blunt. Picture an airport with no air traffic control, every plane trying to land and take off entirely on its own.
Utilities need something equivalent for AI data centers specifically. That means a way to coordinate power draw across many facilities at once so the broader grid stays both reliable and affordable.
NLR's Advanced Research on Integrated Energy Systems platform is built to let industry test that kind of integration at real data-center scale before deploying it, work Hammond says complements the Department of Energy's Genesis Mission to secure both the AI software and the physical infrastructure behind U.S. AI leadership.
What This Means for Miami
Hammond's grid-transient warning is the technical mechanism behind fights already playing out across Florida's data center pipeline, from water permit disputes at Fort Meade to the power-capacity concerns cited in several of the state's new moratorium ordinances.
For Miami's AI infrastructure investors and utility planners, Hammond's framing is useful because it separates two problems that often get lumped together. Cooling technology, the problem his own career largely solved, is mature and well understood.
Coordinating dozens of AI data centers' power swings against a shared grid, the problem he's flagging now, is the one nobody has fully solved yet, in Florida or anywhere else.