GPU cloud · NVIDIA B200 SXM
Dedicated NVIDIA B200 SXM nodes, in India, the EU and the US
Dedicated 8-GPU HGX B200 nodes on a one-year term, the cost-effective Blackwell node for large-model training and inference. FlexiCloud delivers the bare metal and keeps the hardware running; you bring the workload and run your own stack.
From USD 3.75 per GPU-hour. Minimum one node of 8 GPUs on a one-year commitment.
The numbers that matter
- B200 GPUs per node
- 8
- HBM3e memory per GPU
- 192 GB
- Regions: India, EU, US
- 3
- Per GPU-hour, from
- $3.75
Built for
Where the B200 earns its place
The B200 is the value Blackwell: the newer generation, FP4 and fifth-generation NVLink, at a step below the B300 for workloads that fit comfortably in 192 GB.
Value Blackwell training
Blackwell-class training at a lower rate
You get the Blackwell generation, FP8 and FP4 precision, and fifth-generation NVLink, for less than the B300. For most large training runs that is the sweet spot.
- Large-model and mixture-of-experts training
- 192 GB fits most large models on one GPU
- Lower rate than the B300
Efficient inference
FP4 inference that changes unit economics
The second-generation Transformer Engine’s FP4 path roughly doubles throughput on suitable inference, which lowers cost per token against Hopper and needs fewer replicas.
- High-throughput FP4 inference
- Fewer replicas than the H100 or H200
- Lower cost per token at scale
When to pick it
Choose the B200 for Blackwell without the B300 premium
Reach for the B200 when you want Blackwell performance and 192 GB per GPU is enough. Step up to the B300 only when the extra memory or the longest context windows genuinely need it.
- Best value in the Blackwell line
- 192 GB HBM3e per GPU
- Compare head to head with the B300
What we run, and what you run
The node is yours. The hardware is our job.
Most GPU providers hand you a bare node and a support ticket queue. We keep the hardware running underneath: monitored around the clock, failures handled, and an engineer you can actually reach.
Full control of the stack
Root access to run your own OS image, CUDA version, container runtime and orchestration. No platform layer between you and the GPUs; the software is entirely yours.
Watched around the clock
GPU health, thermals, memory errors and interconnect are monitored continuously. A failing card is caught by us, not discovered by a crashed training run at 3am.
An engineer, not a queue
Support is a person who runs GPU infrastructure, in your timezone. When something is wrong you talk to someone who can fix it.
Data where you need it
Nodes in India for domestic data residency, and in the EU and US for teams and datasets that live there. Same service, same team, whichever region.
Yours alone
Dedicated bare-metal nodes. No shared tenancy, no noisy neighbours, no scheduler deciding when your job runs. The whole node is yours for the term.
Predictable cost
A fixed term at a fixed rate. No spot-market surprises, no burst pricing, no bill that doubles because a job ran long. You know the number for the whole year.
Terms
How it is sold
This is dedicated infrastructure on a term, not an hourly spot instance. The terms are simple and they are the same for everyone.
Minimum order
One node: 8 × B200 SXM
The unit is a full HGX B200 node. Eight GPUs with NVLink between them, dedicated to you. Need more? Nodes are added in units of eight.
- 8 × NVIDIA B200 SXM GPUs
- 192 GB HBM3e per GPU, 1.5 TB per node
- NVLink interconnect across the node
- Dedicated bare metal, single tenant
Pricing and term
From $3.75 per GPU-hour
On a one-year minimum commitment, billed for the term. Final pricing depends on region, term length and how many nodes, and we confirm it in writing before you commit.
- From $3.75 per GPU-hour
- One-year minimum commitment
- Longer terms and multi-node priced on request
Availability
India, the EU and the US
Capacity is allocated per region on a first-committed basis. Tell us where the data lives and where the team sits, and we will confirm availability and lead time for that region before you commit to anything.
- India: for domestic data residency
- European Union: for EU-resident data and teams
- United States: for US-based workloads
- Lead time confirmed per region before order
How it works
From first conversation to a running node
- 01
Tell us the workload
Training, fine-tuning or inference; the framework; how many nodes; and which region. Fifteen minutes with an engineer, not a form that disappears into a CRM.
- 02
We confirm capacity and price
A firm quote for your region and term, with the lead time to a live node. No commitment until you have the number in writing.
- 03
We hand over the node
A dedicated bare-metal node, networked and ready to log into with root access. You bring your own OS image, drivers and orchestration; the hardware underneath stays ours to keep running.
- 04
You run. We keep it healthy.
Your team owns the OS, the stack and the workload. Ours watches the hardware, handles failures, and is a phone call away for the whole term.
Questions
What people ask before they commit
Both are 8-GPU Blackwell SXM nodes on the same terms. The B300 (Blackwell Ultra) carries 288 GB of memory per GPU to the B200’s 192 GB, so the B300 suits the largest models and the longest context windows. The B200 is the more cost-effective choice for large-model training and inference that fits comfortably in 192 GB per GPU.
One node, which is 8 NVIDIA B200 SXM GPUs. We do not sell fewer than a full node, because the GPUs share NVLink and the node is the unit of allocation. Additional capacity is added in whole nodes.
One year. This is dedicated bare-metal infrastructure reserved for you, not a spot instance, and the term is what makes the $3.75 per GPU-hour floor possible. Longer terms are priced lower, so a longer commitment brings the Blackwell rate down further.
Pricing starts at $3.75 per GPU-hour, the floor rate on our longest commitments; it is higher on shorter terms and falls the longer you reserve. The minimum commitment is one year. The exact figure depends on your term length, region and the number of nodes, and we confirm it in writing before you commit.
The B200 is NVIDIA’s Blackwell generation. Each GPU has 192 GB of HBM3e and is built from two dies joined and presented to software as one device. It uses the second-generation Transformer Engine with 4-bit floating point (FP4) for inference, and fifth-generation NVLink at roughly 1.8 TB/s per GPU. An 8-GPU HGX B200 node holds 1,536 GB of HBM3e and is typically liquid-cooled.
India, the European Union and the United States. Blackwell capacity moves quickly, so we confirm B200 availability and lead time per region before you order. Pick the region by where your data has to live and where your team works.
We deliver the node as dedicated bare metal with network and root access, and keep the hardware running: continuous monitoring of GPU health and interconnect, hardware failure handling, and an engineer you can reach directly for the whole term. The software stack on top - OS, drivers, CUDA, container runtime and orchestration - is yours to run.
Yes. Nodes are added in units of eight B200 GPUs (1,536 GB of HBM3e each), and multi-node clusters are wired with a dedicated InfiniBand fabric, priced on request. Tell us the target size when we scope the workload and we will plan the interconnect and capacity for it.
No. We work with partners who own the hardware and hold the capacity in each region. FlexiCloud sources the node, delivers it to you as dedicated bare metal, and keeps the hardware running - monitored, failures handled and supported - for the whole term. The software you run on top is yours.
Go deeper
More on running GPUs
The maths and the trade-offs behind an 8-GPU node, cost, memory, rent vs buy, and the interconnect that makes it one machine.
Hopper vs Blackwell: NVIDIA’s GPU generations compared
NVIDIA’s Hopper (H100, H200) and Blackwell (B200, B300) generations compared: memory, architecture, interconnect, power and price, and a practical guide to which one actually fits your AI workload.
Read itB300 vs B200: which Blackwell node, and why the right one lowers your TCO
The B200 is 17% cheaper per GPU-hour; the B300 carries 50% more memory. Which is lower total cost is not about the sticker rate; it is about whether your workload fits in 192 GB, because memory decides how many GPUs and nodes you need. An honest B300-vs-B200 TCO guide.
Read itWhat it costs to fine-tune a model on an 8-GPU B300 node
The cost of a B300 job is GPU-hours × the rate, nothing more. Here is the actual arithmetic, three worked examples from a short LoRA run to a two-week training job, and an honest note on when a node is the wrong tool.
Read itWill your model fit? A practical guide to GPU memory for AI
Whether your model fits is arithmetic, not a benchmark. The bytes-per-parameter maths for inference and training, the levers that shrink it (quantization, LoRA, checkpointing), and a back-of-envelope guide to size any job, mapped to what fits on a B300.
Read it
Tell us the workload and the region
Fifteen minutes with an engineer gets you a firm price, a lead time, and a straight answer on whether the B200 or the B300 is the right fit. No commitment until you have the number in writing.