The physics that rules all of this
Ohm's Law, P = IV, I²R — the 8th-grade equations you forgot are now driving hundreds of billions in infrastructure decisions. Every data centre architecture choice traces back to resistance.
Resistance is the universal physics wall. The same fundamental constraint that drove interconnects from copper to optics is now driving power delivery from 48V to 800V. Understanding the equations is understanding why the industry is being rebuilt.
Every frontier AI data centre is ultimately an enormous machine designed to convert electricity into intelligence. The physics governing that conversion isn't quantum mechanics — it's the classical electromagnetism you learn in high school.
But when you scale those simple equations up to a 600 kW rack drawing 12,500 amps, basic physics becomes a violent, structural constraint. This page builds up the physics of power delivery from first principles. If you understand why a magnetic field induces an eddy current, you can understand why the entire data center industry is being forced to rebuild its architecture.
What killed copper?
As data rates climbed, current stopped flowing through the whole wire. It retreated to the skin — and resistance won.
At high frequencies, alternating current doesn't distribute evenly across a conductor. It concentrates on the outer surface — the skin — leaving the core dark. This is the skin effect. The higher the frequency, the thinner the active ring, the higher the effective resistance.
Skin depth (δ) decreases as frequency (ω) increases. At high data rates, only a fraction of the wire carries signal.
When alternating current flows, it creates a constantly changing magnetic field. According to Faraday's Law, this changing field induces small, circular "eddy currents" inside the wire itself. By Lenz's Law, these eddy currents fight the main current in the center of the wire, but reinforce it at the edges.
The result? The center of the wire becomes a high-resistance dead zone. The electrons are forced to crowd along the "skin," leaving most of the copper completely unused.
This is why the interconnect world moved to optics. Copper hit a physics wall — and that wall was resistance. Now, the same villain is back — this time in the power domain.
P = IV — the equation that decides everything
Power is fixed. Voltage and current are a trade-off. The choice of voltage sets the entire architecture.
A Kyber-class rack needs 600 kW of power. That number is set by the GPUs. The question is: how do you deliver it? Power equals voltage times current. If you fix the voltage, current is the only variable — and it can become absurd.
For fixed power, current is inversely proportional to voltage. Double the voltage → halve the current.
Why 48V? It's a legacy of the telecom industry. Alexander Graham Bell's era settled on 48V because it was the highest voltage that remained generally safe for human contact (Safety Extra-Low Voltage or SELV). Data centers inherited it for reliability, but at 600kW, pushing 12.5k amps is absurd.
The choice between 48V and 800V isn't a preference — it's a physics decision with a 17× difference in current. And current is what kills you, because power loss scales with current squared.
The I²R squared trap
Engineers who think linearly about current underestimate the dissipation problem by two orders of magnitude.
Every wire, busbar, connector, and PCB trace has resistance. When current flows through resistance, power is dissipated as heat. The amount of power lost is not proportional to the current — it's proportional to the current squared.
Power dissipation scales quadratically with current. A 17× current reduction yields a ~278× reduction in heat loss.
The squared term is the trap. It doesn't grow linearly — it explodes.
Grid to GPU — the voltage waterfall
345,000 volts at the grid. 0.8 volts at the GPU die. Every step in between is a different technology, a different supply chain, and a different percentage of power lost as heat.
Power starts its journey at hundreds of thousands of volts on high-voltage transmission lines. By the time it reaches the GPU die, it's been stepped down through 5-7 conversion stages. Each stage uses different physics — transformers, LLC converters, synchronous buck converters — and each stage bleeds 1-3% as heat.
This multi-stage "waterfall" is why power delivery is such a complex market. Unlike memory chips or silicon wafers—where a few companies dominate a single, standardized process—power delivery is a wide-open problem. Every single conversion stage requires completely different physical components, different materials (Silicon vs. GaN vs. SiC), and different engineering trade-offs between efficiency, speed, and thermal management.
The topology war
The industry hasn't decided how to step down from 800V. Two paths are competing — and the losers' entire product lines disappear.
When 800V arrives at the rack, it still needs to be converted down to ~1V at the GPU. But how many stages should that conversion take? Every additional stage adds components, cost, and efficiency loss. The incumbent path reuses the existing 48V infrastructure. The disruptor path skips it entirely.
The incumbent path has 3 conversion stages and preserves the existing 48V component ecosystem. Higher cumulative loss, but the supply chain is proven and deep.
Why does every stage lose efficiency? Because stepping down DC voltage requires rapidly turning power transistors on and off — millions of times per second. Even the best semiconductors aren't perfect switches. Every "flip" burns a tiny fraction of power as heat. Fewer stages mean fewer switches, which means less heat.
CPO for power — vertical delivery
The same principle that drove co-packaged optics applies to power: convert at the edge, not at the perimeter.
In the optics world, co-packaged optics (CPO) moved the optical-to-electrical conversion as close to the chip as possible — eliminating long, lossy copper traces. The same insight applies to power: keep voltage high for as long as possible, and only convert to low voltage at the point of load — directly under the GPU die.
In optics: the lossy domain is copper at high frequencies. In power: it's low-voltage, high-current busbars. Same principle.
VRM sits at the board edge. Low-voltage, high-current traces run across the PCB to the GPU.
VRM sits directly under the GPU die. Conversion happens at the chip — minimal low-voltage trace.
Keep signal in light as long as possible. Convert to electrical at the chip edge.
Keep power at high voltage as long as possible. Convert to low voltage at the chip edge.
Same principle: minimise time in the lossy domain. In optics, the lossy domain is copper. In power, it's low-voltage high-current busbars.
Ultimately, this is a geometry problem. At 1V, pushing 1,000 amps across just 5 inches of motherboard PCB creates catastrophic resistance losses. By delivering power vertically—directly through the Z-axis of the silicon package—you reduce the distance of the lossy domain from inches to millimeters.
The thread that ties it all together
From skin effect killing copper interconnects to I²R losses killing 48V power delivery — resistance is the recurring villain. The solutions are structurally identical: move to a domain where resistance hurts less (light for data, high voltage for power) and stay there as long as possible before converting at the edge of the chip. Every billion-dollar infrastructure decision in AI data centres traces back to this 8th-grade physics.