One GPU, traced from the data center to the mine — and a map of where its value actually hides.
It costs Nvidia roughly $6,400 to build a B200. It sells for around $40,000. That gap is the most-discussed number in technology — and it's the wrong number to be amazed by. Almost none of that $6,400 is the "chip" you picture. The silicon brain is the cheap part. The expensive, fragile, geopolitically radioactive parts are the memory beside it and the packaging that glues it together — and both funnel through a handful of buildings on two islands.
The B200 is the flagship data-center GPU of Nvidia's Blackwell architecture. One module pairs two reticle-sized dies joined by a 10 TB/s link so they act as a single chip — ≈208 billion transistors, ~2.6× the H100 — built on TSMC's 4NP node, with 192 GB of HBM3e across 8 stacks, drawing 1,000 W.
It is not a monolith; it is an assembly — bonded onto a silicon interposer with TSMC's CoWoS-L packaging, the step that quietly governs how many of these can exist at all.
The object you can point to is the least defensible thing in the box. Nvidia designs the logic and owns the software moat — but the module is an assembly of other people's hard-to-make parts, and the real leverage sits a layer down. Hold one question the whole way: if this part vanished tomorrow, could anyone else make it at volume?
What we're tracking this issue: advanced-packaging capacity, lead-time drift past 50 weeks, and whether the memory share ever cracks…
A B200 is not a monolith; it's an assembly — the single most important fact about a modern chip. Dies and memory are bonded to a silicon interposer, mounted on a substrate, capped, and tested. Without that packaging step, a Blackwell GPU is a pile of loose dies, not a product. The dies come out of the fab and then wait in line to be assembled — backend packaging, once an afterthought, is now the gate.
Pull the assembly apart and you get a who-makes-each-piece graph that spans the globe: logic dies fabbed by TSMC (~90% of leading-edge), HBM3e ~90% from SK Hynix with Micron and Samsung behind, advanced packaging at TSMC with ASE/Amkor as added OSAT capacity. The punchline: nearly every critical edge terminates in Taiwan or South Korea.

The components dissolve into materials: patterned silicon with copper wiring, hafnium gate dielectrics, tungsten contacts; the package adds gold, tantalum, solder; the board and rack add copper by the kilogram and rare-earth magnets. Materials are where substitutability — the real source of pricing power — gets decided.
Down to the periodic table — where the object stops being "technology" and becomes extracted earth. Four of these — tin, tantalum, tungsten, gold — are the legally defined conflict minerals (3TG) under Dodd-Frank §1502 and the EU. The GPU is, definitionally, a conflict-mineral product.

Where the elements come out of the earth — and the brutal concentration of it. China controls the overwhelming majority of refined gallium and imposed export controls in August 2023; gallium is a by-product of aluminum refining, so no one can simply "mine more." The U.S. is 100% import-reliant for arsenic, fluorspar, gallium, germanium, indium, and tantalum.
HBM + packaging ≈ two-thirds of the cost. The "chip" is the cheap part. Chip-level gross margin is ~82% on Epoch AI's central estimate — and Epoch cautions the realized margin runs lower once it's sold as a rack/server. And the hidden ledger: TSMC alone consumes ~6–8% of Taiwan's electricity and ran on ~101 billion liters of water in 2023.
One B200's journey, mine to data center. Even dies fabbed in TSMC's new Arizona fabs are flown back across the Pacific to Taiwan to be packaged — roughly 7,500 miles each way (Arizona CoWoS isn't online until ~2028). Three single points of failure — TSMC, CoWoS, HBM — two of them on one island 100 miles from mainland China.

A data-center GPU lasts ~3–5 years; used H100s already trade at ~$15,000–$28,000. At end of life it becomes e-waste — and a rich ore: a tonne of circuit boards holds on the order of 90–300 g of gold (vs ~1–5 g/t for natural ore) plus ~200 kg of copper. Hydrometallurgy recovers ~95%+ of the copper and ~60–95% of the gold — yet globally only about a third of e-waste's metal value (~$28B of ~$91B in 2022) is actually recaptured; the rest is lost to landfill, incineration, or toxic informal recycling. A GPU is engineered to be nearly impossible to take apart for repair, yet must be taken apart to recover what it's made of.

Irreplaceable, a by-product of aluminum refining, ~100% U.S. import-reliant, and the opening shot of China's 2023 export-control campaign. Tracking: import shares, ex-China projects, price.
The packaging step that now decides how many AI chips can exist. Nvidia's capacity lock-up + the Arizona-to-Taiwan round trip is the whole story.
Glass substrates vs. organic; HBM4 and the 16-high stack; Samsung/Micron breaking SK Hynix's grip; the so-far-failed hunt for non-Chinese gallium. What's real, labeled.
| line_item | cost_usd | pct | method |
|---|---|---|---|
| HBM3e (192 GB, 8 stacks) | 2,900 | 45 | modeled |
| CoWoS-L packaging | 1,100 | 17 | modeled |
| logic dies (2× 4NP) | 1,150 | 18 | modeled |
| total variable mfg | 6,400 | 100 | modeled |
| part | supplier | country | share | conf. |
|---|---|---|---|---|
| HBM3e | SK Hynix | South Korea | ~90% | med |
| leading-edge logic | TSMC | Taiwan | ~90% | med |
| advanced packaging | TSMC | Taiwan | primary | med |