bitboy.ro:nvda<GO>

Before a single transistor is drawn, engineers need software just to design the chip. EDA (Electronic Design Automation) tools let designers lay out billions of transistors and simulate how they'll behave — catching errors here costs a bug fix; catching them after fabrication costs a $100M wafer run.

This is where NVIDIA itself sits. A "fabless" designer owns no factories — its product is the chip's architecture, not the physical silicon. NVIDIA designs the GPU; someone else has to actually build it.

Manufacturing starts with raw materials: ultra-pure silicon wafers, plus a long list of rare-earth elements and specialty chemicals that never make headlines but without which nothing else on this list works.

Turning silicon and chemicals into a working chip takes some of the most complex machines ever built — lithography systems that print circuits smaller than a virus, and the etching and deposition tools that build a chip up layer by layer.

The foundry is where the chip is actually fabricated — hundreds of process steps over weeks, on wafers that cost more than a car. NVIDIA doesn't own a foundry; it pays someone else to run theirs.

A smaller group of companies straddle both worlds, fabricating logic chips and memory side by side under one roof.

A GPU without memory next to it is just an expensive calculator. High-bandwidth memory (HBM) sits inches from the chip, feeding it data fast enough to keep up with it.

A freshly fabricated wafer isn't a product yet — it has to be cut into individual dies, tested, and packaged into the chip you'd recognize. This step (OSAT — outsourced assembly and test) is where defective dies get caught before they ship anywhere.

Individual chips get assembled onto boards and into full servers here — the unglamorous, extremely high-volume work of turning components into machines.

The finished servers get a brand on them and go out the door under a name enterprise buyers recognize and order directly from.

For GPUs sold to consumers rather than datacenters, board partners build the actual graphics card around NVIDIA's chip — the cooler, the PCB, the box you'd see on a shelf.

A datacenter full of GPUs is only as fast as the network connecting them. High-speed switches and network cards move data between thousands of GPUs working on the same problem simultaneously.

None of this runs without power and cooling. As GPU racks get denser and hotter, keeping them running has become an industry of its own.

Most people never buy a GPU directly — they rent time on one. Cloud providers buy GPUs at enormous scale and resell compute by the hour.

And for the GPUs that do end up in someone's own PC, retail is the last stop between the factory and the desk.