Compute as a Service · CaaS
CPU that
keeps up
GPUs are useless if the data pipeline stalls. Our compute tier sits on the same leaf switches as the accelerators, so preprocessing, serving and databases stay next to the work.
Instance families
Four shapes,
no surprises
| Family | Processor | vCPU : RAM | Local disk | Network | From |
|---|---|---|---|---|---|
| C: Compute | EPYC 9004 / Xeon 6 | 1 : 2 GB | NVMe optional | 25 Gbps | ₹1.10 /vCPU-hr |
| M: Balanced | EPYC 9004 | 1 : 4 GB | NVMe optional | 25 Gbps | ₹1.35 /vCPU-hr |
| R: Memory | EPYC 9004 | 1 : 8 GB | NVMe optional | 50 Gbps | ₹1.90 /vCPU-hr |
| I: Storage | EPYC 9004 | 1 : 4 GB | up to 30 TB NVMe | 100 Gbps | ₹2.40 /vCPU-hr |
RAM billed separately at ₹0.55 per GB-hour. Dedicated hosts and bare metal quoted monthly.
How it runs
Virtual, dedicated
or bare
Virtual machines
KVM with SR-IOV networking and passthrough NVMe. Snapshots, resize in place, and images that boot in under twenty seconds.
Single-tenant hosts
A whole physical host reserved to you, so noisy neighbours and licensing audits both stop being a topic.
Bare metal
PXE, your image, your kernel. IPMI and serial console exposed, with the NIC and disk layout you asked for.
Networking
Private by
default
- VPC per project: routed subnets, security groups, no shared broadcast domain
- Internal traffic is free: between instances, pods and storage, in any volume
- Direct connect: cross-connect to your own cage, or a private circuit to your office
- DDoS scrubbing: always on at the edge, included, not an upsell tier
- Floating IPs: reassign a public address between hosts without a DNS wait
Also included
Scheduled snapshots
Policy-driven, cross-site, restorable to a new instance in minutes.
Metrics and logs
Prometheus-compatible endpoints and 30 days of retention, no agent licence.
Roles and keys
SSO via SAML or OIDC, scoped API keys, and a full audit trail per project.
Spend controls
Hard caps per project so an experiment can't turn into an invoice.
Migrating from
somewhere else?
We do the assessment, the move and the cutover. Most workloads land inside a fortnight.