Nvidia’s B200 costs somewhere between $30,000 and $40,000 a module, as of July 2025. Epoch.ai put the cost of building one at $5,700 to $7,300. That gap says a lot about Blackwell, the GPU microarchitecture Nvidia announced at GTC on March 18, 2024.
Blackwell replaces Hopper in the data center and covers two families. The B100, B200 and GB200 chips aim at AI training and inference. The RTX 50 series goes to gamers and creators. It is named after David Blackwell (1919-2010), the statistician and mathematician who became the first Black scholar inducted into the National Academy of Sciences, in 1965.
Two dies, one GPU
Hopper used a single die. Blackwell splits the work across two, each as large as chip lithography allows, and joins them with a 10 TB/s link. Software sees one GPU. That is how a B200 reaches 208 billion transistors, which no single die could hold at workable yields. Its Hopper-era predecessor, the GH100, had 80 billion.
The B200 also carries 192GB of HBM3e memory, against 80GB on the H100 SXM5. In practice that lets a single card load a 70-billion-parameter model in FP16, which needs about 140GB, with room left over. An H100 cannot do that on one card.
Blackwell adds 4-bit floating point, FP4, to the precisions earlier chips supported. Halving the memory needed for model weights and roughly doubling inference throughput comes at a price in accuracy, so it suits only workloads that can tolerate it. At FP4 the B200 is rated at 20 petaFLOPS, about five times what the H100 manages.
Nvidia’s published 30x inference gain applies to a narrower case: trillion-parameter models on a fully loaded 72-GPU rack, the GB200 NVL72. On smaller models, one card at a time, the realistic gain is closer to 5x.
Power and plumbing
A B200 draws 700W, the same rating as an H100 SXM5. The trouble starts at rack level. An eight-GPU DGX B200 server pulls about 14.3kW at most, and a full rack of them takes 60 to 100kW, far above the 10 to 30kW typical for enterprise racks. The GB200 NVL72 cannot run without liquid cooling.
Retrofitting an air-cooled facility costs $2 million to $5 million per rack row, according to a 2025 CBRE Group infrastructure report. One data center operator, quoted in a Goldman Sachs report, put it bluntly: “It is not a GPU swap; it is a facility upgrade.”
Using the midpoints, Nvidia clears about $28,500 on each B200, an 81% gross margin on the hardware. Cloud rental ran $8 to $15 an hour in April 2026. Supply stayed tight through most of 2025, held back by TSMC’s 4NP and CoWoS packaging capacity, so most enterprise buyers got Blackwell through multi-year cloud contracts.
Sources
Sources
- NVIDIA Blackwell: 5x H100 Performance at $40K per GPU (2026)
- Blackwell (microarchitecture) – Wikipedia
