OptraCloudAI Inference · powered by Globussoft Start free — 150M tokens
Owner-occupied

We own the building the GPUs go in.

Most Indian GPU resellers rent cabinets inside somebody else's data centre and pass that invoice on to you. OptraCloud runs out of the Globussoft building, owned and occupied by the company that operates it. Our costs start at the electricity bill. There is no colocation contract underneath the pricing with somebody else's margin in it.

The Globussoft building: a five-storey commercial block with a
                full-height glazed frontage, company signage on the roofline, and
                covered parking at ground level.
The Globussoft building. Five floors, glazed frontage, covered parking at grade. Photographed April 2025.
The building

An address you can visit.

Globussoft Technologies has traded since 2012 and occupies this building with its own engineering teams. Customers who need to see where their workload runs before they sign are welcome to come and look. Few resellers can say the same.

The floor

The hall.

The same hall fitted out with two rows of server racks forming a
                  contained cold aisle, in-row cooling units between the cabinets and
                  overhead pipework.
NodeHardwareIT loadHeat to reject
Two-card node2 × RTX A5000 0.81 kW0.28 kW
Four-card node4 × PRO 6000 Max-Q 1.65 kW0.58 kW
Four-card node4 × PRO 6000, 600 W parts 2.85 kW1.00 kW
Eight-card node8 × PRO 6000 Max-Q 3.10 kW1.08 kW
Eight-card node8 × PRO 6000, 600 W parts 5.50 kW1.92 kW

Cooling at 35% of IT load, the figure this business plans against. Max-Q parts roughly halve both columns, so the self-built nodes specify them. Passive Server Edition cards go only in the OEM chassis that supplies its own airflow.

The fleet

What we put in a rack, and what each card unlocks.

Tell us the workload and we specify the card for it. Each row says what that silicon is good for, so we can start from your model and your throughput.

CardVRAMWhat it unlocks
RTX A5000 Ampere · 230 W 24 GB 8B–14B language models, 32B at 4-bit, HunyuanVideo FP8, FLUX, Qwen-Image, Whisper, TTS
RTX 5090 Blackwell · 575 W 32 GB Fast image generation, 14B–32B inference at better tokens per second than the A5000
RTX PRO 6000 Max-Q Blackwell · 96 GB GDDR7 ECC · 300 W blower 96 GB 70B and 120B MoE inference, video generation at roughly 3× the L40S, fine-tuning
RTX PRO 6000 Server Edition Blackwell · 96 GB · passive, 4–8 per chassis 96 GB The same silicon in a density build — 235B MoE across two cards, multi-tenant capacity
H200 SXM Hopper · HGX baseboard, 8 per node 141 GB DeepSeek-R1 671B and frontier-class full models, which need an 8-GPU node and nothing smaller

Ask for the full hardware specification at any point and we send it, including the exact card, host platform, interconnect and the throughput you should expect for your model.

Who you are buying from

The company behind the hardware.

OptraCloud is the AI infrastructure arm of Globussoft Technologies, incorporated 2012. We hold the certifications your security review will ask for, and we send the full certificates and scope statements before you commit to anything.

ISO/IEC 27001:2022
Information security, cyber security and privacy protection.
IC-58930/1225 · valid to Dec 2028
ISO/IEC 27018:2019
Personally identifiable information in public cloud.
GCS/ITST/8532
ISO 9001
Quality management across the engineering organisation.
Certificate on request
Since 2012
300+ engineers, serving over two million users across 50 countries.
CIN U72300CT2012PTC000452
Come and see it

Ask for anything a photograph does not show.

Power, connectivity, physical access control, the full hardware specification and the certificates — we send all of it on request, before you commit to anything.

We reply within one business day — [email protected]