● Mumbai · Pune · Chennai · Noida
GPU capacity
that lives
in India
Rent H200s, H100s and L40S by the second from our own Tier III floors. Your data never leaves the country, and you never pay to get it out.
Per-second billing · Zero egress · No commitment on on-demand
Running production workloads on MSL
What we rent
Three things,
done properly
Everything else is an integration. We build, power and cool the machines ourselves, so the price you see is the cost of the metal plus our margin, not a chain of resellers.
GPU as a Service
Bare-metal and containerised GPUs from a single card to a 512-GPU InfiniBand fabric. Spin up in about 40 seconds.
from ₹44 / GPU-hr↗CaaSCompute as a Service
Virtual machines, dedicated hosts and autoscaling container pools on EPYC and Xeon, billed by the second.
from ₹1.10 / vCPU-hr↗StaaSStorage as a Service
NVMe block, S3-compatible object and archive tiers on the same fabric as your compute. Egress is free.
from ₹0.40 / GB-monthFrom nothing to training
Four commands.
No sales call.
Sign up with a company email, add a payment method, and the API is live. Enterprise procurement exists if you want it; it just isn't in the way.
- CLI, REST and Terraform: the same resource model in all three
- OCI images: bring your own container or start from our CUDA bases
- Persistent volumes: attach the same NVMe volume to any pod in the region
- Idle timeout: pods stop themselves so a forgotten notebook can't cost you a weekend
$ pip install msl-cli
$ msl auth login
# eight H100s, a 2 TiB scratch volume, Mumbai
$ msl pods create \
--gpu h100-sxm --count 8 \
--image msl/pytorch:2.4-cu124 \
--volume scratch:2Ti --region bom1
✓ pod-7fk29d running in 38s
✓ ssh msl@7fk29d.bom1.mslproducts.com
The floor, in numbers
4 sites
Mumbai · Pune · Chennai · Noida
18 MW
Contracted IT load
1.38
Design PUE
99.99%
Uptime commitment
Why teams move here
The boring
reasons matter
Under 10 ms to your users
Inference served from Mumbai reaches most of western India in single-digit milliseconds. A US region cannot do that, whatever the GPU costs there.
Data stays where the law wants it
RBI localisation, the DPDP Act and sectoral rules all point the same way. Our regions are in-country and audited, so residency stops being a design constraint.
Nothing to pay on the way out
Moving a checkpoint set out of a hyperscaler can cost more than training it. We don't meter egress at all: not on object storage, not on pods.
Engineers in your timezone
The person who answers at 3 a.m. IST can see the rack, the switch and the PDU. Escalation is a corridor, not a ticket queue in another hemisphere.
Built for
Workloads we
see every day
Model training
Multi-node runs on 400G InfiniBand with NCCL tuned and checkpointing to local NVMe.
Inference at scale
Autoscaling endpoints with scale-to-zero, so idle traffic costs nothing overnight.
Fine-tuning
Single-node A100 and L40S pods for LoRA and full fine-tunes, priced for iteration.
Render and simulation
RTX 6000 Ada farms for VFX, CAD and CFD, with shared project volumes.
Capacity is live now
Start on one GPU.
Grow to a cluster.
On-demand needs a card and an email. Reserved capacity needs a conversation, usually a short one.