From Grid to GPUs: Exploring the evolution of data centre power

This year’s Super Bowl took the spotlight with commercials pitching artificial intelligence, and unsurprisingly so, conversations about the future are once again dominated by, what’s next in AI?

You might think the key ‘transformers’ driving the future today are the AI models, but the real silent heroes are the electrical ones that keep the power flowing.

As AI scales, the real bottleneck is no longer NVIDIA building the latest and greatest GPUs—it’s securing, delivering, and managing reliable gigawatts for powering data centres and the critical path we take for this rapid growth. We typically only hear about the two ends of this problem: either securing a massive power supply, or to squeeze more computing into each rack. Recently, Crusoe and Tallgrass announced partnership to develop a 1.8 GW AI data centre campus, superpowered with natural gas supersonic turbines! Elon Musk’s xAI are buying used power plants intending to power one million NVIDIA Blackwell GPUs. On the other end, Google , Meta and Microsoft are collaborating through the Open Compute Project to develop 1 MW level GPU racks, pushing efficiency and density to the limit. But what connects these two extremes has been quietly evolving in the background.

While there are many creative ways to power these GPUs, we have distilled them down to four discrete architectures for simplification, taking you through the messy middle between the ends of running AI data centres that is seldom talked about.

Gen 1: Centralized Uninterruptible Power Supply (UPS)

Shortly after the first computer was ever created, someone turned it off unexpectedly. They realized that should not happen again.

Here the Gen 1 Uninterruptible Power Supply (UPS) comes into play, which has a small amount of local energy storage. In the event of power outage, the UPS can provide a computer enough time to shut down properly while saving its work or bridge the gap until backup power like a generator can be turned online for uninterrupted operation.

These early UPS would typically take AC from the grid, rectify to DC to charge a battery and then invert DC back to AC to run your computer on its own hyper local microgrid. Since the inverter running your computer was always on and connected to the DC link of the battery, there was zero switchover time when the grid supply went out. This method is commonly referred to as double online conversion. As a positive side effect, the output was also a clean AC power source that computers like because it is isolated from the grid. These early servers in total were only about 5 to 10 kW power.

As the computing demand has rapidly increased, moving from a personal computer to a server rack, so did the power levels. These UPS systems are able to scale up and with that the battery voltage also scale up to 48 V, eventually leading to high voltage battery systems for even high power levels.

But as the power levels have gone up, so have the operators’ power bills, looking at ways to make things more efficient. Modern systems now also have a line interactive or “economy” mode that mostly bypasses the double online conversion as needed. This keeps the battery online and in the event of a grid outage, the eco-mode disconnects the grid side and the battery plus inverter takes it over.

Gen 2: Power Conditioning System (PCS) + 48 V Battery Backup Units (BBU)

The next major Gen-2 milestone happened when someone took a step back and assessed the entire power system. The traditional UPS had a few extra power conversion steps, which ultimately add up in CapEx costs and efficiency losses. If they upped the DC Distribution voltage from 5-12 VDC to 48 VDC at the rack level, they could still be considered “low voltage” systems while drastically reducing the amount of amps and getting to higher power levels like 30-50 kW. They could also place a 48 VDC BBU directly in the rack, and reduce conversion steps and cost. In this architecture, there is still a need for AC Power Conditioning System (PCS) to get clean power or even voltage conversion, IE 480V/3P to 208V/3P, but the backup power function is now local to the rack.

Now enters the world of AI data centres. GPU clusters have physical distance limitations for coherence, meaning GPUs must remain electrically and physically close to behave like a single system. So, the highest performing machines are now at an absolute volumetric space constraint. At present, one of NVIDIA’s latest NVL72 racks is rated for up to 140 kW of power, equivalent to the power density of a toaster. Announcements show plans to ramp up to 1 MW racks targeted for 2027.

When operators talk about trying to keep their servers from melting, they are not exaggerating. With extreme heat generation, tight space requirements for compute coherence, and the safety risks of storing significant energy inside a multimillion-dollar machine, it becomes clear why 48 V batteries have outgrown the GPU rack. Like kids leaving home, it is simply time for them to move out

Gen 3: Sidecars and High Voltage (HV) DC

Gen-3 architectures consist of 800 VDC uni-polar or +/-400 VDC bipolar DC links. By moving to this high-voltage architecture, and putting batteries and power converters in an external rack, we can drop currents and increase efficiencies. Not only that, but we can start to see the fusion with DC Fast Charger stations for electric vehicles and leverage economies of scale.

In this architecture, building power of 480 VAC or 415 VAC is rectified through a bidirectional Active Front End (AFE) to 800 VDC to supply a common DC link. A high-voltage battery is also connected through a bidirectional DC/DC converter to support in the event of a grid outage. Multiple systems are then paralleled on the same DC link for N+1 redundancy and managed in real time using DC Droop control logic to control current sharing. This topology, when combined with advanced battery chemistry and controls can also add in new feature sets typically only seen in grid-tie assets. It can support demand response, frequency regulation, overload capability, current shuttling, and active/reactive (PQ) power injection back to the grid. Achieving this level of integration, especially for bidirectional power flow and advanced microgrid services relies on cutting-edge power electronics like high-efficiency SiC-based converters, the core expertise of companies like ADVANTICS. These capabilities add great value to operations without sacrificing reliability or uptime.

Gen 4: Medium Voltage Solid State Transformers + HV DC

Where is the industry headed next for Gen 4? Besides heading to space with satellite-based data centres, there is active development going on (on Earth) for medium voltage (MV) AC to 800 V DC solid state transformers. It could further reduce losses and improve efficiencies by eliminating the primary grid transformer and conversation steps. However, there are still major technical challenges to overcome with this topology. To achieve these high-voltage levels, the rectified stages are usually stacked to handle such high voltage level. This introduces additional complexity, as well as risk of damage due to surge events such as lightning strikes. This is a challenge on its own and when you introduce high peak loads like in a data centre, it can quickly lead to system instability. If the device is also providing galvanic isolation, its underlying power-transfer mechanism still relies on a high-frequency transformer (HFT), but one that is much smaller and lighter due to the higher operating frequency. Here the physics of the conversions are the same, but the component is fundamentally different from the bulky 60 Hz grid transformer it replaces. At a certain point, solutions become an expensive

Here the physics of the conversions are the same, but the component is fundamentally different from the bulky 60 Hz grid transformer it replaces. At a certain point, solutions become an expensive, high-risk version of arranging desk chairs as to how to sequence – isolation, voltage conversion, and rectification of grid power. When you move from a low-frequency 50 or 60 Hz large transformer to much higher frequency, around 300 kHz as in our new release, ADM-PC-BI25, far less energy needs to be stored per cycle, allowing the magnetic components to be dramatically smaller.

The ADVANTICS Advantage

Built on nearly a decade of experience developing high-power density fast charger components, our Gen 3 architecture leverages proven advancements from the automotive EV sector. This deep foundation enables us to translate mature, field-tested innovation into reliable solutions for large-scale data centre and infrastructure applications. ADVANTICS offers a comprehensive suite of solutions engineered for the high-density, high-efficiency future of data centres:

• Silicon Carbide (SiC) Power Devices: Our core technology delivers the high efficiency and power density necessary for 1 MW+ racks and Gen 3/4 HV DC links. In our new line of liquid-cooled power converters, the ADB-PC-DC01 module is a bidirectional isolated 100 kW power converter that connects a high-voltage UPS battery to 800 VDC link. The ADB-PC-AC01 module is a bidirectional AC/DC inverter (active front end), compatible with all standard AC mains voltages and frequencies worldwide to operate seamlessly across different regions. It can rectify this input to an 800 VDC link or be integrated into a double online conversion setup to provide clean, stable power to the grid.

• Bidirectional Capability: Seamlessly integrate battery storage and enable lucrative grid-service functions (PQ injection, frequency regulation) with our bidirectional (AC/DC) rectifier/inverters and (DC/DC) converters.

• Rack-Ready and Utility-Scale: We provide flexible, high-power building blocks from rack-mounted AC/DC units to the DC/DC converters needed for your external high-voltage Battery Energy Storage System.

Wherever you are headed for your power needs, don’t let the complexity of Gen 3 or Gen 4 stability challenges or the need for advanced grid services slow down your critical path.

Partner with ADVANTICS to implement the ‘silent heroes’ powering the next-generation AI data centres.