Quoted post

CV Research · (@CVResearch_)

View quoted post

New deep dive is live on $VICR.

You have probably not seen this name on your feed. That is most of the reason we looked.

Every note on AI power is written in watts. But loss rises with the square of CURRENT, not power. The constraint in the last six inches of an AI machine is measured in amps, and almost nobody
measures it.

A wafer-scale processor pulling 20 kilowatts at 0.6 volts wants 33,333 amps. More current than a lightning bolt, continuously, into silicon the size of a dinner plate.

Move that power across the board at 6 volts, and it burns 2,200 watts getting there. Move it at 48 volts, and it burns 35.

Same chip, same power, one decision about a voltage. Sixty-four times the loss.

So we went and found out why almost nobody can deliver power the second way.

There are two ways to step a voltage down. Averaging, a switch flipping on and off very fast, which is what the entire industry ships. Or transforming, which trades volts for amps at a fixed
ratio, the way a gearbox trades speed for torque.

One company transforms. That choice, made by a CERN physicist in the early 1980s, produces four properties a competitor cannot copy. Not for budget reasons. They fall out of having a turns
ratio, and a switch does not have one.

Also inside:

First-half bookings ran $414 million against $211 million shipped. The factory is the constraint, and it cost $75 million against more than a billion of capacity.

November 2017, when this exact architecture held the world record for computing efficiency. It appears in no current research note on this company. What happened nineteen days later is why.

A hundred published customer applications across fourteen markets, and not one is about AI. Including a two-millimeter mirror in the Atacama, reshaped a thousand times a second by 5,352 magnets,
where an inefficient converter does not cost money. It destroys the image.

A licensing business set up in 1987 that never sues its customers. It stops their supplier at customs. Sixty million dollars closed in three weeks that way.

Six times they built the destination product before the market arrived. Eight to twelve years early, and captured the economics of none of them.

The technology risk was retired in public, more than once. Selling it never was.

Why it's worth your time

Reframes rack power in the unit that binds, amps rather than watts, and puts a number on it: moving the same chip's power at 48 volts instead of 6 cuts the loss 64-fold. First-half bookings of 414 million dollars against 211 million shipped says the constraint is the factory, not demand. Feeds the last-six-inches leg of the power-delivery thesis.