NVIDIA’s B200 costs around $6,400 to produce, with memory accounting for half

High bandwidth memory isn't cheap

How much does it cost NVIDIA to make a B200 GPU?

By our estimate, around $6,400, implying chip-level gross margins of ~80%.

The counterintuitive part: the logic die, the core component of the GPU, is one of the cheaper parts, accounting for less than 15% of the total cost.

The biggest cost is High Bandwidth Memory at $2,900, nearly half the total. The B200 packs 192 GB of HBM3E—2.4× the H100’s 80 GB of HBM3, and a faster generation of memory. This is what lets the chip move data quickly enough to keep its compute units busy.

Advanced packaging is second at $1,100. B200s use TSMC’s CoWoS-L packaging process to connect multiple chiplets and HBM stacks on a single package using tiny silicon bridges, enabling the massive memory bandwidth modern AI chips need.

Packaging yield loss adds another ~$1,000. When packaging fails, the logic die and HBM cannot be recovered. Blackwell had low packaging yields early in production, which has since improved but likely still trails Hopper. Our uncertainty here is wide: $420–$1,700.

The logic die? Just $900. Two large ~800 mm² chiplets fabricated on TSMC’s 4NP process cost less than a third of the cost of memory.

At reported selling prices of $30–40k, this implies chip-level gross margins around 80%. But NVIDIA’s realized Blackwell margins are likely lower since most revenue comes from servers and racks bundled with other components, which likely carry lower margins. NVIDIA’s overall margins have been running ~73%.

Read more from Venkat Somala’s analysis on our website!