On Friday of last week, our commentary provided a diagram of a data center and each of its relevant components. The piece not only outlined just how complex these systems are, but it also highlighted the wide range of companies and industries that support the buildout of these data centers. It also explained where some of the manufacturing bottlenecks exist, like in the production of power, the supply of leading-edge chips, water usage, and the limited supply of critical minerals. Today’s weekly market commentary drills down a bit further into one of these potential bottlenecks, and that is the energy that supports these factories. More specifically, the enclosed piece examines the actual distribution of power within data centers and of course, what that may mean for investors.
To illustrate the massive power demands on data centers – and how these demands have changed as the need to support AI infrastructure has grown – the enclosed chart shows the estimated rack power (measured in kilowatts) by Nvidia compute architecture. Rack power is the amount of electricity required to operate all of the computing equipment housed within a single data-center server rack.
Starting on the far left, the chart begins by showing the power consumption of a traditional server rack. These types of racks were used with traditional cloud data centers which predated AI server racks. Then we can see each succession of Nvidia generations, with Hopper in 2022, Blackwell in 2024, Rubin in 2026, Rubin Ultra in 2027 and then Feynman in 2028 and beyond. As can be seen in the chart, and as the piece explains, “power consumption has increased with each successive Nvidia computing architecture, with future rack systems projected to exceed 1 megawatt, enough to power 1,000 homes at once, by the end of the decade.” – AOM 483. This is an increase of power consumption of about 100 times.
The piece continues on to explain that existing data centers were not designed for this level of power density. They were produced to distribute power at voltage levels that were appropriate for a traditional rack server. But currently distributing these large amounts of electricity at today’s relatively low voltages is inefficient. It requires bulky copper cables, occupies valuable space, and generates additional heat.
To address these challenges, and what that potentially may mean for investors, the piece explains, “raising the voltage addresses both of these challenges. At higher voltage, the same power flows with less current, reducing both the copper required and the energy wasted as heat. Nvidia has announced that its server racks will shift to a new 800-volt design in the next few years, roughly 15 times today’s standard. A broad ecosystem of semiconductor and electrical equipment suppliers has also committed to the transition. We believe these changes could drive demand for companies providing the equipment that converts, distributes, stores, and safely manages electricity, positioning power delivery as an important layer of the AI buildout.” – AOM 483
The continued expansion of AI will require enormous computing capacity—but increasingly, the ability to power that capacity efficiently may be just as important.
We would ask that you review the attached piece at your convenience and please let us know if you have any questions or if you would like to discuss it further. And as we always end this correspondence, please remember that regardless of current momentum and regardless of the key takeaways in this weekly perspective, we will continue to monitor and manage with a thoughtful approach based on your specific long-term objectives.
