India’s AI Cloud.

It depends on the workload. For training large language models at 70B parameters and above, the H100 and H200 carry 80 to 141 GB of VRAM and support NVLink for multi-GPU scaling. If you want the most VRAM per rupee, the RTX Pro 6000 Blackwell has 96 GB at ₹179/hr. For fine-tuning and inference, the A100 at 40 or 80 GB gives the best performance per rupee. For students and researchers running experiments, the A30 at ₹39/hr and the L4 at ₹41/hr are the cheapest way in. Every GPU ships with CUDA, cuDNN, PyTorch, and TensorFlow pre-installed.

No. The A6000, RTX 6000 Ada, and A5000 are no longer part of the JarvisLabs lineup and cannot be launched in any region. The closest replacements are the RTX Pro 6000 Blackwell at ₹179/hr, which carries 96 GB of VRAM against the A6000's 48 GB, and the A30 at ₹39/hr or L4 at ₹41/hr for the lighter training and inference work that used to run on an A5000.

Yes. Spot instances run on spare capacity at a steep discount and can be preempted when that capacity is needed back. In India, an H200 is ₹189/hr on spot against ₹378/hr on demand, and an H100 is ₹113/hr against ₹255/hr. Spot suits checkpointed training runs, batch inference, and anything you can restart. Save your work often, because a preemption can arrive at any time.

On training GPUs, yes. An H100 is ₹255/hr here against roughly ₹557/hr on AWS and ₹896 to ₹995/hr on GCP and Azure, so 2 to 3.5 times cheaper for the same hardware. An A100 40GB is ₹84/hr here against roughly ₹222/hr on AWS. The gap narrows on smaller inference GPUs, so we would not claim it across the board. The larger practical difference is access rather than rate: hyperscaler GPU quota starts at zero and is reviewed by hand, and large GPU instances frequently return capacity errors, while here you launch as soon as you have credits.

GPU-as-a-Service means on-demand access to NVIDIA GPUs without owning hardware. On JarvisLabs you sign up, add credits from ₹810, pick a GPU and a pre-built template (PyTorch, TensorFlow, JAX, ComfyUI, and others), and the instance is live in under 90 seconds. You get full access through JupyterLab, VS Code Web, or SSH. Pause when idle to stop GPU charges and resume later; your data and environment persist across sessions.

Yes. Billing is per minute and there is no minimum rental period. You can take an H100 for a 30-minute training run, pause it, and come back to it next week. That suits intermittent work: fine-tuning runs, image generation, and weekend research projects.

Payments go through Stripe with supported Indian credit and debit cards, and every top-up comes with a GST tax invoice. If you’re GST-registered, your GSTIN appears on the invoice so your finance team can process input credit per your tax position.

There is no free tier, but ₹810 is enough to start, and at ₹39/hr on an A30 that is over 20 hours of GPU time. In practice that beats free-tier options like Google Colab, which come with session limits, queue times, and disconnections mid-run. Several IITs and other universities use JarvisLabs for ML coursework and research for exactly that reason.

Buying a GPU server with an H100 costs ₹30 to ₹50 lakh upfront, before electricity, cooling, and maintenance. Renting on JarvisLabs, you pay ₹255/hr only while the GPU runs. At 8 GPUs for 8 hours a day, that works out to roughly ₹61,236 a month, a fraction of the purchase price, and you get access to every other GPU type without a procurement cycle.

Yes. Instances go up to 8 GPUs for distributed training, available on H200, H100, RTX Pro 6000, A100, A30, and L4. That is what you need for training large language models, running DeepSpeed or FSDP, and working through large datasets. H100 and H200 instances use NVLink interconnects for maximum throughput.

Yes. JarvisLabs runs GPU capacity in Indian datacenters, so your instances and their data stay in India. You can also create a VPC with your own private IP ranges and run instances inside it, isolated from other tenants; VPCs are available in the India regions today. Each instance has persistent storage that survives pauses and restarts, so your code, datasets, and checkpoints are kept until you delete the instance.

Three steps. Sign up at jarvislabs.ai. Add credits from ₹810 by card. Then pick a GPU and a pre-built template and launch. The instance is ready in under 90 seconds with JupyterLab, VS Code Web, and SSH, and the whole path from signup to a running training job takes under five minutes.

Pausing stops GPU charges immediately. While paused you pay only for storage, at ₹0.0130/GB/hour, and your environment, packages, datasets, and checkpoints are all kept. Resume to pick up where you left off. Deleting is permanent: the data goes and no further charges apply, so download anything you need first.