Loading benchmark data
Every accelerator hour this site prices began as a lens, a machine, a wafer and a memory stack. These are the companies in that sequence and the relationships between them, each one carrying the sentence that states it.
A relationship is recorded only where a source names both parties. Where one of the two says so, it is confirmed; where neither will and a named publication does, it is marked as reported and drawn differently. A foundry does not publish its customer list, so that distinction is not a technicality.
Search companies in Atlas SearchMakes the optics inside those machines.
Makes the machines the fabs are built out of.
Makes what the package is built on.
Turns a design into silicon.
Joins the die to its memory.
Makes the stacked memory without which an accelerator is not a product.
Designs the accelerators, and manufactures none of them.
Makes the switches, optics and interconnect that join accelerators into one machine.
Builds the finished racks.
Makes the power and cooling equipment a datacenter is built around.
Generates or delivers the electricity the racks run on.
Owns the buildings the racks go into, and sells no compute at all.
Rents compute by the hour, at hyperscale.
Rents GPU compute by the hour as a specialist.
Trains frontier models, on compute it mostly buys.
Buys accelerators at scale and consumes every one of them itself.
Holds stakes in the companies here and builds nothing itself.
479 sourced relationships across 185 companies. The number beside each company is how many it has, in both directions.
