B200 at a glance
Built on Nvidia's Blackwell architecture. Replaces H100 and H200 as Nvidia's top-tier data centre GPU.
192GB
HBM3e memory
8 TB/s
Memory bandwidth
208B
Transistors
FP4
Native precision
1.8 TB/s
NVLink 5
How does B200 compare to H200?
| Spec | B200 | H200 |
|---|---|---|
| Architecture | Blackwell | Hopper |
| Transistors | 208 billion | 80 billion |
| Memory | 192GB HBM3e | 141GB HBM3e |
| Memory bandwidth | Up to 8 TB/s | 4.8 TB/s |
| Lowest native precision | FP4 | FP8 |
| NVLink | 5th gen, 1.8 TB/s | 4th gen, 900 GB/s |
| Power draw | Up to 1,000W | Up to 700W |
The jump from H200 to B200 sits in three places: memory, bandwidth, and precision. Together, they decide how fast a model generates tokens and how many requests one GPU can serve at once.
Which GPU fits your workload?
Answer three quick questions for a personalised recommendation.
1. Model size
2. Priority
3. Timeline
Who is Nvidia B200 built for in India?
Tap a segment to see how B200 fits.
What will B200 cost to run?
CloudPe hasn't published B200 pricing yet. Here's the logic that will decide if it's worth it.
- Cost per token matters more than cost per GPU-hour
- A higher hourly rate can still mean a lower cost per million tokens, if throughput is high enough
- B200's memory bandwidth and native FP4 are built for exactly that trade-off
- CloudPe's GPU cloud starts at Rs. 14,500/month today (H200, H100, A100, L4, RTX Pro 6000), India-hosted, no egress fees
How to reserve B200 allocation with CloudPe
Frequently asked questions
Large language model training and inference, high-concurrency AI applications, and large-scale simulation. Its 192GB of memory and FP4 support suit trillion-parameter and mixture-of-experts models specifically.
192GB of HBM3e memory, running at up to 8 TB/s of bandwidth.
CloudPe is opening early access to B200 ahead of general availability. Nvidia H200, H100, A100, L4, and RTX Pro 6000 are live on CloudPe today.
B200 carries more memory (192GB vs 141GB), more bandwidth (8 TB/s vs 4.8 TB/s), and native FP4 precision, which H200 does not support. Both run on India-hosted CloudPe infrastructure.
If a workload fits inside 141GB of memory and runs well on H200 today, there is little reason to wait. B200's advantage grows with models that need more memory, higher throughput, or FP4 precision to run cost-effectively at scale.
AI and ML companies building Indic language models, BFSI firms running real-time fraud detection, global capability centres, and government or academic research projects under the IndiaAI Mission.