An AI data center is full of chips.
Some chips perform the AI calculations.
Others do something less visible: they repeatedly reshape electricity so those compute chips can use it.
AI needs chips to compute—and chips to deliver the right voltage and current to the computing chips.
First, Follow One Watt From the Grid to the GPU
A GPU cannot plug directly into the utility grid.
Electricity arrives at the data center at a voltage and form suited to power distribution.
The processor core, however, needs extremely low voltage and very high current.
So power moves down a staircase.
Original Asset 1: The Voltage Staircase
utility / medium-voltage AC
↓
transform, protect and switch
↓
AC-to-DC conversion
↓
high-voltage DC distribution
↓
intermediate bus: tens of volts
↓
point-of-load conversion
↓
sub-1-volt processor core
Each step needs switching devices, controllers, inductors, capacitors, protection and thermal management.
The semiconductor devices that switch and control this electrical power are power semiconductors.
Quick Answer
Power semiconductors control large flows of electrical energy efficiently.
In an AI data center, they appear in:
- power supply units,
- battery backup units,
- high-voltage converters,
- intermediate bus converters,
- hot-swap and protection circuits,
- and voltage regulators next to the GPU.
The important point is that 800 V, 50 V and below 1 V can all exist in the same power architecture—but at different stages.
800 V is a distribution voltage. It is not the voltage delivered directly into the GPU core.
Why AI Racks Are Forcing the Power Architecture to Change
AI rack power is rising quickly.
The International Energy Agency says AI server power density increased about elevenfold between 2020 and 2025 and could increase roughly fourfold again by 2027. It says one advanced rack in 2027 could have peak power demand comparable to 65 households.[1]
That changes the physical problem.
A low-voltage power bus carrying a modest amount of power is manageable.
The same voltage carrying hundreds of kilowatts demands enormous current.
Original Asset 2: The Current Wall
The starting equation is simple:
Power = Voltage × Current
For the same power:
higher voltage → lower current
Why does current matter?
Because large current means:
- larger conductors and busbars,
- more connector stress,
- more resistive heating,
- more difficult current sharing,
- and harder board-level power delivery.
Resistive heating follows another simple relationship:
resistive loss = I²R
A Common 800-V Misunderstanding
Moving from 54 V to 800 V reduces current by about 14.8 times at the same power.
If you keep the resistance exactly the same, the I²R term becomes about 219 times smaller.
But that is a physics illustration—not a claim that the whole data center suddenly has 219 times less electrical loss.
Real systems also change:
- conductor size,
- busbar geometry,
- converter topology,
- number of conversion stages,
- switching losses,
- protection equipment,
- and cooling.
Higher voltage reduces the current problem. It does not erase every other power-conversion loss.
Why 800 VDC Is Emerging
NVIDIA's current 800-VDC architecture is designed to move more power through dense AI infrastructure with fewer conversion stages and lower current than legacy low-voltage rack distribution.[2]
NVIDIA says an MGX-compatible 800-V power rack is arriving in the second half of 2026 as a hybrid path for existing AC facilities.
For larger new facilities, its roadmap includes a row power center with overhead 800-V busway supporting up to 2 MW per row, expected in 2027.[2]
This matters because 800 VDC is not only a greenfield idea.
The architecture is also being designed with transition paths for facilities that already exist.
This Is Becoming an Industry Architecture, Not Only an NVIDIA Architecture
Google, Microsoft and NVIDIA are collaborating through the Open Compute Project on an open 800-V low-voltage-DC architecture for next-generation AI data centers.[3]
The basic motivation is physical: higher-voltage distribution can move more power with less current and less conductor burden.
Standardization matters because racks, converters, battery systems, breakers, connectors and safety systems have to agree on the architecture.
What Happens After 800 V Reaches the Rack?
The voltage still has to come down.
Infineon's current grid-to-core architecture describes a path where high-voltage distribution feeds intermediate converters, which then create tens-of-volts buses before final point-of-load stages supply processor rails below 1 V.[4]
At those very low processor voltages, current can reach thousands of amps.
That means 800 V solves one part of the path while the final centimeters near the accelerator remain an extreme current-delivery problem.
The Last Inch Can Become the Next Bottleneck
Imagine a processor that needs several kilowatts but operates at roughly one volt or below at its core.
The lower the voltage, the more current is required to deliver the same power.
This is why power delivery is moving closer to the processor.
Infineon says vertical power delivery—bringing power through the board toward the processor rather than routing everything laterally across the board—is becoming important as AI processor current rises.[4]
The challenge therefore shifts:
low-voltage rack current problem
↓ solved partly by higher-voltage distribution
high-voltage conversion & protection problem
↓
sub-1-V processor-current problem
Original Asset 3: The Power-Semiconductor Job Map
There is no single “best” semiconductor material for every stage.
| Technology | Useful mental model | Where it often fits |
|---|---|---|
| Silicon | Mature workhorse | Low-voltage switching, control, processor-level power stages, broad cost-sensitive applications |
| Silicon carbide (SiC) | High-voltage specialist | Higher-voltage conversion, protection, battery backup, solid-state transformers and breakers |
| Gallium nitride (GaN) | Fast-switching specialist | High-frequency, high-density intermediate conversion where compact magnetics and fast switching matter |
These are useful tendencies, not hard rules.
The right material depends on voltage, current, frequency, thermal limits, reliability, packaging and cost.
Why SiC Matters at High Voltage
Silicon carbide is a wide-bandgap semiconductor, meaning its material properties can support efficient switching at higher electric fields and temperatures than conventional silicon in suitable applications.
That makes SiC attractive in high-voltage power conversion and protection.
Infineon's 2026 24-kW battery-backup reference design, for example, uses 650-V and 1,200-V SiC devices to connect directly to an 800-V DC bus and reports efficiency above 99%.[7]
That number belongs to that specific reference design. It is not a universal 99% figure for every BBU or every data center.
Why GaN Matters for Dense Conversion
Gallium nitride is also a wide-bandgap material.
Its strength in many power applications is very fast switching.
Higher switching frequency can allow smaller magnetic components and denser converters, although the complete system still has to manage control, heat, electromagnetic interference and cost.
Infineon says GaN switching near 1 MHz is being used for ultra-compact bus-converter approaches in its 800-V AI-power roadmap.[6]
Why Silicon Does Not Disappear
New materials do not make mature silicon obsolete.
Silicon remains inexpensive, well-understood and highly capable across many lower-voltage and control stages.
The new AI power stack is therefore less “SiC versus GaN versus silicon” and more:
Which semiconductor technology fits this voltage, current and switching job best?
Original Asset 4: The Conversion-Loss Cascade
Every conversion stage has an input and an output.
The difference becomes heat.
At small scale, one percentage point can look trivial.
At AI-rack scale, it is not.
If a 1-MW rack loses 1% at one conversion stage:
1% of 1 MW = 10 kW
That is 10 kW of continuous heat at that stage.
At a 100-MW facility, 1% equals 1 MW of continuous electrical loss before considering the extra energy needed to remove that heat.
This is why conversion efficiency can affect:
- electricity cost,
- cooling load,
- rack density,
- facility capacity,
- and how much useful compute fits behind a fixed grid connection.
Original Asset 5: The Grid-to-Compute Efficiency Ladder
A useful system model is:
grid power
× facility conversion efficiency
× rack conversion efficiency
× board / voltage-regulation efficiency
=
power that reaches compute
Then the processor's own energy efficiency determines how much useful AI work comes from that delivered power.
This separates two different questions:
- How many megawatts can the grid deliver?
- How much useful compute can the data center extract from each delivered megawatt?
AI Loads Also Move Fast
AI power is not always a smooth, constant load.
Training and inference workloads can create large and rapid changes in demand.
The IEA highlights these rapid power swings and says energy storage becomes important for reliable supply.[1]
That is one reason the AI power architecture includes more than converters.
It also includes:
- BBUs — battery backup units,
- capacitor backup,
- power-path control,
- fast protection,
- and increasingly intelligent digital power management.
Protection Gets Harder Too
High-voltage DC has an important safety challenge: DC current does not naturally pass through a zero crossing every half-cycle the way AC does.
That can make high-energy DC fault interruption more demanding.
Power architectures therefore need fast protection, hot-swap devices and circuit breakers designed for the new voltage/current regime.
Infineon's current roadmap includes SiC-based protection and solid-state circuit-breaker approaches for future DC grids.[5]
Power Management Is Becoming Smarter
The power system also has to respond to changing processor loads quickly.
In August 2026, Infineon acquired C2i Semiconductors, a specialist in software-defined multiphase controllers and high-density AI power delivery.[9]
The strategic point is broader than one acquisition.
As current density rises, power delivery becomes a control problem as well as a transistor problem.
Original Asset 6: The 800-VDC Benefit Test
When you see a new AI power architecture, ask:
- Rack density: How much power must each rack or row deliver?
- Current: What current would a lower-voltage bus require?
- Copper: How large do cables, busbars and connectors become?
- Conversions: How many power-conversion stages remain?
- Protection: How are DC faults isolated safely?
- Backup: How do batteries or capacitors connect?
- Semiconductors: Where do Si, SiC and GaN actually improve the design?
- Retrofit: Can an existing AC facility transition through a hybrid architecture?
- Standards: Is the architecture interoperable across vendors?
- Last inch: Does sub-1-V delivery near the GPU become the next limiting step?
What Power Semiconductors Cannot Solve
Efficient conversion does not create new electricity.
If a site cannot obtain a grid connection, power semiconductors cannot manufacture megawatts.
They also do not remove transformer, cooling, construction or generation constraints.
What they can do is help turn a larger fraction of delivered electricity into useful compute—and make much denser racks physically possible.
What to Watch Next
- 800-V deployment. Does the hybrid power-rack approach move from reference architecture into broad deployment?
- OCP standardization. Do equipment ecosystems converge enough to avoid incompatible 800-V variants?
- Protection. Do solid-state breakers and hot-swap designs scale safely and economically?
- Vertical power delivery. Does power move through substrates and boards closer to the processor?
- Last-inch regulation. Can sub-1-V rails supply thousands of amps without consuming too much board area or losing efficiency?
- SiC / GaN manufacturing. Does supply expand with data-center demand without creating a new constraint?
- Software-defined power. Do digital controllers become more important as AI load transients grow?
- Efficiency metrics. Does the industry report end-to-end grid-to-compute efficiency rather than isolated converter peaks?
The Simple Idea to Remember
The GPU does not consume “800-V power.”
The data center uses high voltage to move large amounts of power efficiently, then converts it down stage by stage until the processor receives the very low voltage it actually needs.
The hidden AI power-chip race is really a race to move more watts through less space, lose fewer of them as heat, and still deliver thousands of amps safely at the processor.
Key Vocabulary
power semiconductor
A semiconductor device used to switch, convert or control electrical power.
PSU
Power supply unit: converts facility/rack power into the DC power used by servers.
BBU
Battery backup unit: supplies short-term stored energy when the main source changes or fails.
IBC
Intermediate bus converter: converts one DC distribution voltage into a lower intermediate DC voltage.
VRM
Voltage regulator module: produces the precise low-voltage, high-current rails needed by processors.
SiC
Silicon carbide: a wide-bandgap semiconductor used in many high-voltage, high-efficiency power applications.
GaN
Gallium nitride: a wide-bandgap semiconductor valued in many high-frequency, high-power-density converters.
800 VDC
An emerging high-voltage direct-current distribution architecture for high-density AI infrastructure.
vertical power delivery
Delivering current through the board/package toward the processor rather than routing all power laterally across the board.
Read the AI Hardware Full Stack Series
- Why AI Chips Need Advanced Packaging, Not Just Smaller Transistors
- What Is a Chiplet? Why AI Chips Are Splitting Into Specialized Dies
- Why Silicon Photonics Could Become the Next AI Data Center Bottleneck
- What Is an NPU? Why AI Does Not Have to Run Only on GPUs
- The Hidden Chips Behind AI Power: Why Power Semiconductors Matter
Related Articles
- Why Power, Not Chips, May Limit the AI Data Center Boom
- What Must Be Built to Power the AI Data Center Boom?
- What Is the AI Full Stack?
Sources
- IEA — Key Questions on Energy and AI, 2026.
- NVIDIA — Why Scaling AI Compute Performance Requires a New Power Architecture, August 11, 2026.
- Open Compute Project — Google, Microsoft and NVIDIA 800-VDC standardization, August 11, 2026.
- Infineon — AI data-center power solutions from grid to core, checked October 4, 2026.
- Infineon — We Power AI / grid-to-core architecture, checked October 4, 2026.
- Infineon — NVIDIA MGX and 800-VDC power architecture, May 29, 2026.
- Infineon — 24-kW SiC BBU reference design, June 2, 2026.
- Infineon — AI data-center PSU reference designs, June 2026.
- Infineon — C2i acquisition / software-defined power management, August 24, 2026.
Sources checked through October 4, 2026. Vendor efficiency figures describe specific reference designs, not universal facility efficiency. Silicon, SiC and GaN roles vary by voltage, topology, switching frequency, thermal requirements, reliability and cost.