$VICR.US held a world record in 2017 for the exact architecture the entire AI industry is racing toward now. Then the story disappeared.
Here's the setup. Power loss scales with the square of current, not power. Double the current, you burn four times the energy. That's just what copper does, no amount of capital fixes it.
A wafer-scale AI processor pulling 20 kilowatts at 0.6 volts needs 33,333 amps. More current than a lightning bolt, continuously, into a chip the size of a dinner plate.
Deliver that at 6 volts and you lose 2,200 watts getting it there, 11% of the entire power budget turned into heat. Deliver it at 48 volts and you lose 35 watts. Same chip, same power, one decision about voltage, 64 times the difference.
There are two ways to step voltage down that far. A switch flipping on and off fast, which is what almost everyone ships, Monolithic Power, Infineon, Analog Devices, TI. Or a transformer that trades volts for amps at a fixed ratio, the way a gearbox trades speed for torque.
Vicor is one of the only companies doing it the second way, and they've been building this architecture since the early 1980s.
Now the world record. In November 2017, a Vicor-powered supercomputer called Gyoukou held the record for computing efficiency using this exact setup, co-packaged power conversion, direct 48V to sub-1V at the chip.
That's the entire 2026 AI power pitch, and it was in production and world-record-holding nine years ago. customer's CEO got arrested 19 days later on an unrelated fraud charge and the whole program died with him. It's vanished from every current research note on this company.
There's also a licensing business here running near 100% margin. They don't sue customers, they go to the ITC and get customs to block the infringing supplier at the border. One deal closed in three weeks for $60 million because a hyperscaler's supplier got physically stopped at the port.
Proven technology, a world record to back it up, and a licensing weapon that gets paid regardless of who wins the sockets. Question is whether they can finally convert it.
Here's the setup. Power loss scales with the square of current, not power. Double the current, you burn four times the energy. That's just what copper does, no amount of capital fixes it.
A wafer-scale AI processor pulling 20 kilowatts at 0.6 volts needs 33,333 amps. More current than a lightning bolt, continuously, into a chip the size of a dinner plate.
Deliver that at 6 volts and you lose 2,200 watts getting it there, 11% of the entire power budget turned into heat. Deliver it at 48 volts and you lose 35 watts. Same chip, same power, one decision about voltage, 64 times the difference.
There are two ways to step voltage down that far. A switch flipping on and off fast, which is what almost everyone ships, Monolithic Power, Infineon, Analog Devices, TI. Or a transformer that trades volts for amps at a fixed ratio, the way a gearbox trades speed for torque.
Vicor is one of the only companies doing it the second way, and they've been building this architecture since the early 1980s.
Now the world record. In November 2017, a Vicor-powered supercomputer called Gyoukou held the record for computing efficiency using this exact setup, co-packaged power conversion, direct 48V to sub-1V at the chip.
That's the entire 2026 AI power pitch, and it was in production and world-record-holding nine years ago. customer's CEO got arrested 19 days later on an unrelated fraud charge and the whole program died with him. It's vanished from every current research note on this company.
There's also a licensing business here running near 100% margin. They don't sue customers, they go to the ITC and get customs to block the infringing supplier at the border. One deal closed in three weeks for $60 million because a hyperscaler's supplier got physically stopped at the port.
Proven technology, a world record to back it up, and a licensing weapon that gets paid regardless of who wins the sockets. Question is whether they can finally convert it.
