Laser-focused on AI workloads only. We are not a general-purpose cloud; every rupee, rack, and process is optimised for modern AI training and inference in India.

• Better ₹-per-FLOP economics – latest GPUs + liquid cooling maximise compute for every rupee.
• Regulatory peace-of-mind – data, keys, and logs never leave Indian jurisdiction.
• Rapid time-to-GPU – containerised 3 MW pods can be live ≈ 6 months after purchase order.

Bottom line: If your workload isn’t AI, we’re not your cloud. If it is - nobody in India can match our cost, latency, sustainability, or sovereignty guarantees at this scale.

GPU portfolio (launch sequence)

• Blackwell + family (GB200 and future refreshes)
• AMD MI350X clusters for cost-optimised training
• Cerebras CS-3 wafer-scale engines for ultra-large models
• Groq LPUs for sub-10 ms, high-throughput inference

Evolving rack densities – pods start at 150 kW and are engineered to scale to 300 kW and 600 kW per rack as chip TDPs climb.

Pricing models

• On-demand ₹/GPU-hour or month-to-month rental
• Reserved instances (1-3 yr) with up to 50 % discount
• Dedicated enterprise clusters billed at cluster size rate
• Pay-per-API-call for managed inference endpoints

Security and sovereignty are built in from day one:

• Certifications – SOC 2 Type II, ISO 27001 / 27017 / 27701, and continuous DPDP Act 2023 audits.
• Always-on protection – 24 × 7 SIEM monitoring, zero-trust IAM, full-volume and in-flight AES-256 encryption, HSM-backed key management.
• Enterprise sandboxing – every tenant runs in a logically isolated VPC with private subnets, dedicated GPU pools, and policy-based egress controls.
• Defence-grade segregation – national-security workloads are housed in a physically separate, air-gapped building tied directly to classified defence networks and staffed by cleared personnel.
• Sovereign safeguards – data, logs, and keys never leave India; change-of-control clauses lock infrastructure to Indian jurisdiction.
• Compliance reporting – real-time audit dashboards plus quarterly third-party penetration tests, with attestation reports available under NDA.

Result: Whether you are a start-up, a listed enterprise, or a defence agency, your workloads stay isolated, encrypted, and fully compliant—today and as regulations evolve.

Minutes for experiments, hours for production.

• Self-service portal – spin up a single-node notebook or multi-GPU job in < 15 minutes.
• Pre-baked ML images – PyTorch, TensorFlow, JAX, and CUDA optimised kernels are one click away.
• Custom enterprise clusters – delivered in 24-48 hours with VPC peering, private subnets, and policy-based egress controls.
• Code-level API – request, monitor, and tear down GPU pools programmatically (see example in the next answer).

Result: you can move from “idea” to “training run” the same day—no GPU waitlists, no long procurement cycles.

Yes—Antriksh Cloud plugs into any cloud or on-prem stack.

• Native peering & VPN options for AWS, Azure, GCP, OCI, and on-prem DCs.
• API marketplace – REST & Python SDKs for workload scheduling, data migration, and cost optimisation across clouds.
• Code-level access – request GPUs in one line:

                        from antriksh.cloud import Cluster

                        # Create an 8-GPU GB200 server in 'delhi-1' region
                        cluster = Cluster(region="delhi-1")
                        job = cluster.request(
                            gpus=8,
                            gpu_type="GB200",
                            duration="24h",            # on-demand
                            storage="2TB_NVMe",
                            network="100G_InfiniBand"
                        )
                        print("Job ID:", job.id)
                        

• Kubernetes-ready – use our CSI & CNI plugins to burst pods to Antriksh while keeping control-plane on your primary cloud.
• Unified billing – single invoice in INR even when workloads span multiple providers.

Bring data from anywhere, burst massive AI jobs to our GPU fabric, and return results—without lock-in.

We combine predictive software and modular hardware to keep queue wait-times under five minutes—even on launch-day spikes.

• Elastic GPU pools : Kubernetes + Slurm control plane can burst spare GPUs across adjacent racks in < 30 s; autoscaler pre-allocates capacity as queue depth grows.
• Tiered job queues :
  1. On-demand (interruptible / non-interruptible)
  2. Reserved (guaranteed start window)
  3. Priority (24 × 7 latency-critical inference)
  Jobs are promoted or throttled automatically based on SLA and credit balance.
• Hot-swap containers : GPU “blade” modules (96–192 kW) let us inject or remove 512–1,024 GPUs at a time without touching core power-or-cooling loops.
• Predictive scaling : Prometheus telemetry feeds a Prophet model that forecasts demand two hours ahead; idle pods are shut down or released to the spot pool.
• Multi-chip flexibility : If NVIDIA inventory tightens, the scheduler transparently routes compatible workloads to AMD MI300X or Groq LPUs, keeping queue wait < 5 min.
• Reserved-capacity SLAs : Enterprise contracts include “GPU credits” that guarantee burst rights during peak events—backed by a 10 % service-credit penalty for any miss.

Sub-50 ms latency nationwide, backed by multi-terabit dark fibre.

• Feasibility complete : end-to-end fibre routes surveyed with five independent carriers already present in the region.
• Tri-path redundancy : dual dark-fibre paths to neutral IXPs in Mumbai & Delhi NCR, plus a third east-bound path for disaster resilience.
• Massive headroom : initial order of 1.6 Tbps of dedicated bandwidth (four 400 G channels) split across at least three providers.
• Metro ring integration : 400 G leaf-spine fabric ties directly into India’s top metro markets, sustaining < 50 ms RTT even under failover.
• Carrier-agnostic cross-connects : single-mode and multi-mode options inside the data-centre meet-me room for tenant direct connects, Wave, MPLS or Internet transit.
• Jumbo frame & RDMA-ready : 100 G/400 G InfiniBand gateways ensure low-latency east-west traffic for distributed training jobs.

Result: AI workloads reach users and other clouds in milliseconds, with no single carrier dependency and ample future bandwidth.

Five complementary revenue streams power our business:

• GPU Infrastructure-as-a-Service (IaaS) – on-demand and reserved GPU hours for training workloads (≈ 50 %)
• AI Platform-as-a-Service (PaaS) – managed training pipelines, low-latency inference, vector DBs, RAG tooling, and MLOps (≈ 20 %)
• Defence & Government Cloud – air-gapped, sovereign clusters on long-term contracts with national-security agencies (≈ 15 %)
• Venture Compute – we allocate otherwise idle, off-peak GPU capacity to promising AI start-ups at a steep discount in exchange for equity stakes (≈ 5 %)
• AI-Ops & Compliance Add-ons – audit logging, model observability, policy enforcement, and premium support subscriptions (≈ 10 %)

These streams blend usage-based revenue with sticky subscriptions, turning idle capacity into equity upside while keeping cash flow predictable.

Seed-Infra round: US $66 million

• 64 %: 3 MW Proof-of-Concept (PoC) capex, incl. GPUs and liquid-cooling modules
• 20 %: Land acquisition with a multi year payment plan, permits, and state-incentive obligations
• 10 %: working capital & early operations
• 6 %: contingency reserve

This capital turns the PoC live and revenue-generating—creating an optional early exit for investors—yet our plan is to scale the campus ourselves.

Next phase of capital: scale from 3 MW to the first 100 MW GPU cluster once PoC revenues and metrics are proven.