NVIDIA B300 in India: bare-metal GPU nodes, real pricing, and who they are for

NVIDIA's B300 is the top of the current training and inference stack, and until recently getting your hands on one in India meant a waitlist or a flight to a US region. That is changing. FlexiCloud now runs dedicated B300 SXM nodes in India, the EU and the US, from $4.50 per GPU-hour. This post explains what a node actually is, what it costs, and how to tell whether you need one.

What the B300 SXM actually is

The B300, part of NVIDIA's Blackwell Ultra line, is a data-centre GPU built for large AI models. The headline number is memory: 288 GB of HBM3e per GPU. That matters because the size of the model you can train or serve without splitting it across machines is bounded by how much high-bandwidth memory each GPU has. More memory per GPU means larger models, longer context windows, and bigger batches before you have to shard.

The unit you rent is not a single GPU but a node: eight B300 SXM GPUs on one board, connected to each other by NVLink so they behave, for many workloads, like one very large accelerator. That is 2.3 TB of GPU memory in a single machine. Eight is the unit because the GPUs share that high-speed interconnect; splitting a node in half would break the thing that makes it fast.

Bare metal, not a managed platform

There is an important distinction that a lot of GPU marketing blurs, so we will be plain about it. FlexiCloud delivers the B300 node as bare metal: dedicated, single-tenant hardware with network and root access. You bring and run your own software stack, the operating system, drivers, CUDA version, container runtime and orchestration. We keep the hardware running underneath, monitoring GPU health and interconnect, handling failures, and giving you an engineer to reach for the whole term.

In short: you run the software, we run the hardware. For teams doing serious training that is usually what they want, full control of the stack, no platform layer between them and the GPUs, and no scheduler deciding when their job runs.

What it costs, and why it is a term not an hourly spot

The rate scales with the length of your commitment: the longer you reserve the node, the lower the per-GPU-hour price. The minimum commitment is six months, and that shorter term sits at the higher end; as the term lengthens the price comes down, to a floor of $4.50 per GPU-hour on the longest reservations. The exact figure also depends on region and how many nodes, and is confirmed in writing before you order. You can see the current rate on the pricing page.

Why a term rather than an hourly spot instance? Because dedicated bare metal is reserved for you. Spot GPU pricing looks cheaper per hour until you factor in the cost of a job that gets pre-empted mid-run, the queue you wait in for capacity, and the bill that swings with the market. For sustained training or production inference, a fixed rate on a fixed term is both cheaper in practice and far more predictable. You know the number for the whole six months.

Who actually needs an 8-GPU B300 node

A full node is a serious amount of compute. It is the right size when:

• You are training or fine-tuning large models and a single GPU can no longer hold the model or the batch you need.
• You are running production inference for a large model at scale, where the 288 GB per GPU lets you serve bigger models or longer contexts on fewer machines.
• You need data residency in India and cannot send training data to a US or EU region.
• You want predictable cost for a project with a known duration, rather than a spot bill that moves under you.

It is not the right size if you are experimenting on a small model, need a GPU for a few hours a week, or are still at the notebook-prototyping stage. For that, a smaller cloud GPU or a shared instance elsewhere is a better fit, and we will tell you so.

B300 vs B200 vs H100, at a glance

If you are choosing a generation, the simplest way to think about it is memory per GPU, since that sets the ceiling on model size:

• B300 (Blackwell Ultra) - 288 GB HBM3e per GPU. Best for the largest models, the longest context windows, and headroom to grow into.
• B200 (Blackwell) - 192 GB HBM3e per GPU. Strong for large-model training and inference.
• H100 (Hopper) - 80 GB HBM3 per GPU. Established, widely available, and lower cost for workloads that fit.

The rule of thumb: if your model and batch fit comfortably on an H100 and you are cost-sensitive, an H100 is the pragmatic choice. If you are memory-bound today, or expect to be as your model grows, the extra headroom on the B300 is what you are paying for. Figures above are NVIDIA's published specifications.

India, the EU and the US

We allocate capacity per region on a first-committed basis, so the practical questions are where your data has to live and where your team works. India nodes keep training data onshore for data-residency requirements. EU and US nodes put the compute next to data and teams already there. The service, the pricing model and the operations team are the same in every region; only the location changes. We confirm availability and lead time for your chosen region before you commit to anything.

Who has B300 capacity in India

India went from almost no Blackwell Ultra silicon to being one of a handful of countries hosting it, in barely a year. A few names are worth knowing if you are mapping the market:

• Yotta - the largest play by far. Yotta has announced a supercluster of 20,736 NVIDIA Blackwell Ultra (B300) GPUs, a roughly $2 billion build, part of which hosts an NVIDIA DGX Cloud cluster, with over 10,000 B300 GPUs committed to the government's IndiaAI Mission.
• Cyfuture - offers B300 GPU servers as part of its enterprise AI cloud.
• Neysa - an NVIDIA-backed Indian AI-cloud company building out Blackwell-generation capacity for training and inference.

Most of that capacity is pointed at hyperscale buyers - sovereign model builders, large enterprises, the IndiaAI Mission - bought in big reservations. FlexiCloud sits at the other end of the same supply: single dedicated B300 nodes you can reserve one at a time, as bare metal, without being a supercluster-scale customer. If you need one node rather than a thousand, that is the gap we fill.

Common questions

What is the minimum order? One node, which is eight B300 SXM GPUs. We do not sell fewer than a full node because the GPUs share NVLink and the node is the unit of allocation.

What is the minimum commitment? Six months. This is dedicated bare metal reserved for you, and the term is what makes the rate possible. Longer terms are priced lower.

Do you manage the software? No. You run the OS, drivers, CUDA and orchestration; we run and maintain the hardware for the term.

Can I scale past one node? Yes, in units of eight GPUs. Multi-node deployments are priced on request.

Getting started

The honest first step is a conversation, not a checkout. Tell us the workload, the framework, how many nodes and which region, and we will come back with a firm price and a lead time. If B300 is not the right fit for what you are doing, we will say so. Start on the B300 page, or talk to an engineer.