The Rise of Power at the 2025 OCP Global Summit

Blog Summary

• AI data centers are scaling rapidly: The OCP Global Summit highlighted the shift from Megawatt to Gigawatt-scale data centers, driven by the exponential growth in AI workloads and processor power demands. Companies are investing heavily in high-density, high-power infrastructure to support this evolution.
• HVDC power architecture gains momentum: High Voltage Direct Current (HVDC) solutions were prominently featured, with new products like HVDC power shelves, racks and distribution boards showcased to meet the rising power needs of next-gen AI systems.
• Power shelves are still crucial for today’s infrastructure: Despite the push toward HVDC, traditional and inferencing racks continue to rely on ORv3-based power shelves due to their practicality and efficiency for sub-300 kW applications. Advanced Energy is leading with scalable shelf solutions to bridge current and future power needs.

---

The OCP Global Summit has grown to become one of the major conferences for the Artificial Intelligence (AI) Data Center market. This year’s event attracted over 10,000 attendees. The keynotes were hosted in a massive tent with thousands of people in attendance. At times, it felt more like Coachella than a tech conference, with high-profile presenters showcasing the industry's advances in AI to the audience (fans). These advancements are music to our ears, and we can’t wait to see the breakthroughs our industry will achieve to further advance our collective goal towards this AI revolution.

At this year’s OCP Global Summit, we witnessed the rise of power (along with cooling) in this new AI world, power now being viewed as important as compute and networking in enabling AI. The presentations, product demos, and conversations at the show confirmed several merging trends in power:

From Megawatt (MW) to Gigawatt data centers. With AI being adopted across a wide range of applications and use cases, companies from hyperscalers and enterprise to neocloud and sovereign AI are looking to build data centers with a higher density of AI factories, resulting in higher power data centers. Recent announcements of multi-billion-dollar AI investments by multiple companies highlight the push for these Gigawatt data centers. We expect more announcements and plans for greater number of these large data centers, which will require new solutions to supply the high raw energy supply. In the push for higher power, we believe companies with track records of delivering production-proven, reliable, high power, high density, and high efficiency power solutions will capture outsized opportunity in this market.

Compounding effects of accelerating rack power. There are three drivers accelerating rack power rating from tens of kW per rack to over 1MW per rack: 

• Increased power rating of AI processors. Power ratings for each generation of AI processors have been increasing exponentially. As an example, the Thermal Design Power (TDP) of NVIDIA’s next-generation Rubin Graphics Processor Units (GPU) is projected to reach over 2,000 W vs. Blackwell at 1,350 W and Hooper at 700 W. Similar increases are seen for other processors known as xPUs.
• More AI processors are being packed into each rack. NVIDIA’s Blackwell NVL72 has 72 GPUs each with two dies, or 144 GPU dies, but Rubin Ultra Kyber rack is expected to have four times the density with 576 dies. Similar density increases are seen in other rack designs.
• Redundancy requirement. Increased importance of power means that more racks are designed with N+N redundancy instead of N+1 or N+2. This means that the total power within a rack needs to be two times higher than the requirement to ensure redundancy.

HVDC power architecture was front and center at OCP. HVDC as a power architecture has been proposed over the last few years at OCP. This year, however, there was a noticeable increase in the number of HVDC enabling products on display at the show. This acceleration is partially driven by NVIDIA’s plan to launch the Rubin Ultra Kyber requiring up to 1MW of power, necessitating the need to move to a higher power and more efficient solution. At the show, we saw several product demonstrations related to HVDC:

• HVDC power shelves. Several companies displayed HVDC-to-bus DC power shelves for IT rack to enable faster adoption of HVDC power architecture. Since all AI servers today are still based on 54V bus DC input, these shelves enable data center to adopt HVDC architecture without waiting for native HVDC servers to become available. However, the high-power requirements of these shelves at 100 kW can be challenging to achieve for most vendors.
• HVDC power rack. These are racks that convert data center AC distribution power to 800V or plus and minus 400V HVDC to be fed into an IT rack. These racks contain AC to DC and DC to DC PDUs (Power Distribution Units), AC to DC Power Shelf, and in some cases BBU shelves and capacitor shelves. Suppliers of both power conversion solutions and data center power infrastructure are developing these racks and showed them on the exhibition floor.
• HVDC Power Distribution Board (PDB). PDB for converting 800V or plus and minus 400V HVDC to 54V or even 12V bus power were seen in the NVIDIA demo area. Most of these are initial reference designs produced by semiconductor vendors and are not optimized for density, efficiency, and other real-world production requirements yet. We expect real products to become available by next year’s OCP.

Power shelf remains the “sweet point” even as the industry prepares for HVDC. As exciting as the efficiency benefits offered by HVDC power architecture are, there are still many use cases and benefits of using current or future generations of power shelves. At the show, many racks, servers, and data center infrastructure companies show off their new designs utilizing power shelves in their booths. For example, AMD showcased their newly announced Helios rack on the floor with the next generation 72kW ORv3 HRPv2 shelf on display. Similarly, at the Meta booth, the company showed off its new IcePack networking rack with two power shelves and two BBUs shelves. For racks with power density of up to 300 kW, we believe power shelf remains primary option for IT gear designers without overloading a rack with too many shelves. We believe the current roadmap for ORv3-based power shelves will intersect a large portion of the market building these racks, with meaningful demand ahead.

AI inferencing is believed to be less power intensive than AI training, requiring less dense racks that will likely be below the 300 kW threshold mentioned above. While moving to an HVDC architecture could still be more efficient for inferencing, the net benefits are potentially less pronounced, and designers will likely focus on improving training racks while still using power shelves for inferencing racks.

In addition, given the insatiable demand to build AI infrastructure as fast as possible but a constrained supply of xPU, the largest infrastructure owners will continue to build racks with whatever processors they can get their hands on in the near future. As a result, a lot more racks will be built with today’s power shelves in parallel to the denser HVDC racks.

Power of traditional servers is also increasing. While most of the talks were focused on AI, traditional servers will still be needed alongside AI-accelerated racks for many workloads. In fact, Dell’Oro estimated that by 2029, only 16% of server units being built will be AI-accelerated servers. General-purpose servers will account for 84%; although, the AI-accelerated servers will require more dollar investment and power consumption. Some of the traditional servers could have a small number of xPUs working with the most advanced CPUs. As a result, power required per server could be increasing to over 3kW range. Many of these servers will still be designed with in-server power supply units (PSUs), while some might move to power shelf architecture. While these server racks might not need as much power as AI racks, power efficiency remains an important factor to consider for these racks. With the increased focus on power efficiency in AI servers, designers of these traditional servers are also looking for more efficient solutions to minimize power usage. 

Advanced Energy is leading the industry with server rack power solutions that are designed to help data center operators meet the rising energy and cooling requirements of AI infrastructure — delivering higher efficiency and reliability. At OCP, we showcased a new 100 kW, 48V ORv3 HPR-compliant power shelf designed to enable the industry’s transition to HVDC power architecture. Additional generations of ORv3 shelves, ranging from 18 kW to 72 kW, were also on display. The company also presented board-mounted DC-DC converters with up to 98% efficiency for converting 48V to 12V applications, and its portfolio of M-CRPS compliant Titanium efficiency power supplies.

Advanced Energy has delivered multiple generations of ORv3 compliant power shelves in the production environment. With proven technology, reliability, and capabilities, Advanced Energy remains committed to power the world’s growing demand for AI. For more information on Advanced Energy’s data center capabilities, please visit: Data Center Computing | Advanced Energy.