Two to eight NVIDIA graphics cards (GPUs) in one server, run under full load with your own AI model before it ships. For a team in an office, a two-card workstation from ₹46 lakh.
A team on an office floor: the Multi-GPU workstation. A whole department, company or lab, or the biggest open models: a GPU server, and a server room to put it in.
Our photo: room for a second cardTypical price, GST included₹46–57 lakh
See the Multi-GPU workstation
Studio renderTypical price, GST included₹64 lakh–2.1 crore
See the GPU serversTwo real prices, GST included. September 2026: an office tower with two workstation cards, 192 GB in all, ₹56.5 lakh. October 2026: a server holding 192 GB on two data-centre cards, with a pair of 96-core server CPUs in a server chassis, ₹93.7 lakh.
The office tower’s cards carry their own fans and push heat out of the case. Data-centre cards rely on the server’s fans blowing air through them, which is loud. The server also adds error-correcting server memory and a spare power supply. An eight-card box always goes in a server room.
Two RTX PRO 6000 cards give 192 GB of GPU memory in a tower that stands on the office floor, no server room needed. ₹46–57 lakh with GST.
Pick it when more than 32 people query a mid-size model at once, or to fine-tune a 70B model. That model needs about 48 GB to answer at 4-bit; fine-tuning it at 4-bit (QLoRA) takes three to four times as much, 144–192 GB, so it needs both cards. RAM and SSDs are sized per order.
Get my spec and priceAs many as eight cards and 768 GB of GPU memory on AMD EPYC, built to live in a server room and serve a whole department or company. ₹64 lakh–2.1 crore with GST.
One chassis holds everything. Multi-server clusters are outside what we build.
Get my spec and priceThree NVIDIA data-centre cards, or the workstation RTX PRO 6000. Memory decides how big a model fits; bandwidth, how quickly the card works through it.
The same 96 GB GPU as in our Pro workstation, built to be cooled by the server.
96 GBmemory
1,597 GB/smemory bandwidth
600 Wat most, adjustable
Best when: many people at once, on everyday to very large models. Eight make 768 GB in one server.
Up to 600 W each: eight draw several kilowatts, so we plan power and cooling with you.
The most memory on one GPU. Two or four can be linked together with NVIDIA’s NVLink bridge.
141 GBmemory
4.8 TB/smemory bandwidth
600 Wat most, adjustable
Best when: the biggest models. More memory on each GPU means fewer GPUs for the same model.
One on its own came to ₹57.8 lakh in October 2026, GST included: more than a whole two-card workstation.
A 48 GB GPU that also handles graphics and video work.
48 GBmemory
864 GB/smemory bandwidth
350 Wpower
Best when: mixed jobs: AI plus rendering, graphics or video in the same chassis.
Half the memory of an RTX PRO 6000, so a big model needs more of them.
NVIDIA RTX PRO 6000 Blackwell Max-Q and Workstation Edition
The card in our Pro workstation, in two editions that carry their own fans.
96 GBmemory
1,792 GB/smemory bandwidth
300 / 600 WMax-Q / Workstation
Best when: the same card across your workstations and your server. The eight-GPU, 768 GB server in our price list uses eight of these.
Own cooling or not, an eight-card box is loud and still belongs in a server room.
Specs: NVIDIA’s product pages, October 2026. Product images and video: NVIDIA. All four are sourced to order and priced on the day you ask.
Four jobs that outgrow a single GPU. Training needs far more GPU memory than serving, so say which you plan.
Every department queries one model over your own policies, contracts and files. It runs in your IT room, so nothing goes to an outside cloud.
GPU serverfar more than 32 people querying together
Train an open model on tickets, reports or research data so its answers sound like your organisation.
Multi-GPU workstation4-bit fine-tuning (QLoRA) of a 70B model needs 144–192 GB
A department’s students and researchers book time on one box to run and fine-tune their models.
Multi-GPU workstation for a team; a GPU server for a whole departmentseveral models side by side, up to gpt-oss 120B size
DeepSeek R1-class models on hardware you own, in your server room, instead of renting them by the hour.
GPU servertoo large for any single GPU
Five recent prices, September and October 2026, with GST, priciest first. No customer names. The ranges on this page cover all our quotes since June 2026, so they start lower than this list.
768 GB on 8 workstation-class GPUs
₹2.1 crore
September 2026
282 GB on 2 data-centre GPUs
₹1.41 crore
September 2026
192 GB on 2 data-centre GPUs
₹93.7 lakh
October 2026
144 GB on 3 data-centre GPUs
₹82.9 lakh
October 2026
192 GB on 2 workstation GPUs
₹56.5 lakh
September 2026
Where it can sit depends on the whole box. One or two cards with their own fans can sit in an office. Any eight-GPU server belongs in a server room, whatever cards it holds.
₹57.8 lakh
Above all, the cards. That is what one NVIDIA H200 with 141 GB came to on its own in October 2026, with GST: more than our whole two-card workstation.
Get my spec and priceCard prices change week to week, so we price on the day you ask, and the formal quotation lists every part with its own price. Your company reclaims the 18% GST through our invoice.
When the cards work through most of the day, buying beats renting. The monthly maths for a two-GPU workstation:
Two rented cloud GPUsOne NVIDIA L40S each, AWS Mumbai list price ₹213.96 an hour, used 12 hours daily
₹1.56 lakha month
Owning the workstationPower at ₹10.11 per unit, plus upkeep
₹12,200a month
About 30 monthsis the break-even for a ₹50 lakh Multi-GPU workstation used 12 hours daily. At 8 hours daily it is roughly 46 months.
For a server: eight of those cloud GPUs at 12 hours a day come to about ₹6.25 lakh a month. The ₹2.1 crore eight-GPU server is about ₹1.78 crore after the GST credit, so roughly 28 months of that rent, before its power and upkeep, which depend on your site.
Break-even is the price after your company reclaims 18% GST, divided by the monthly cloud saving once power and upkeep are paid. A rough guide for planning, not a quotation.
Draw is taken as about 1.05 kW, with two Max-Q cards: each at the 300 W measured on our RTX PRO 6000 Max-Q build, the CPU at its rated figure, and roughly 100 W for everything else. ₹50 lakh sits inside the ₹46–57 lakh range; the cloud rate is AWS Mumbai’s published hourly price for one NVIDIA L40S. That card has 48 GB to the RTX PRO 6000’s 96 GB, so this compares card for card, not memory for memory.
Built for a purchase committee: spec and price inside 4 working hours, a formal quotation for the file, and an invoice with GST in your company’s name. Authorised partner/dealer of AMD and NVIDIA.
Send the models, the user count and where the machine will live: office floor, IT room or server room.
Every component priced, and a straight answer if a smaller machine would serve you better.
On NVSX letterhead with the full spec sheet for the file. Your team raises the purchase order against it.
Lead time depends on the cards, which are sourced to order, so the quotation gives the dispatch date before you sign. Then assembly in Baner, a full-load run and your model on it.
Inside Pune our team installs it on-site. Anywhere else in India it travels insured, and we bring it up with your IT team over a video call.
Setup is free, and the 3-year service warranty works as on every NVSX machine: free labour, diagnosis and claim handling, parts under each maker’s warranty. Warranty and paperwork in full
Our buyers range from government and research labs and universities to AI startups and large companies.
Four AI machine sizes; the top two live on this page. When one card is enough, a Pro workstation at ₹20–45 lakh does the job, and we’ll say so in the quote.
32 GB on one card
Serves up to 16 users at once from a desk
₹3.5–11 lakh
Open its page96 GB on one RTX PRO 6000
Serves up to 32 users at once from a desk
₹20–45 lakh
Open its page192 GB across two RTX PRO 6000
A team beyond 32 users, or bigger models; office floor
₹46–57 lakh
On this page768 GB max, across eight cards
Company-wide, from the server room
₹64 lakh–2.1 crore
On this pageStraight answers on price, space, power and support. Your question isn’t here? Ask us on WhatsApp.
From our quotes since June 2026: ₹64 lakh to ₹2.1 crore with GST, the top figure being an eight-GPU, 768 GB build. The two-GPU office workstation sits at ₹46–57 lakh. Every line of the quotation carries its own price.
One NVIDIA H200 with 141 GB, as a bare card, came to ₹57.8 lakh including GST in October 2026. Card prices change weekly, so we price on the day you ask. Cards are available alone or inside a machine.
Yes, up to eight GPUs inside a single chassis, with error-correcting server memory and a spare power supply that takes over if one fails. Clusters of several servers are outside what we build.
For a GPU server: space in a cooled server room, a dedicated power line and a UPS for several kilowatts, and a network port. The two-GPU workstation needs only an office room and a UPS. We plan power, cooling and placement with your facilities team.
Usually, if we know now. Say so on the form and we choose a board, chassis and power supply with free slots and headroom, so more cards go in without a rebuild.
It depends on the cards, which are sourced to order. The quotation gives the dispatch date before you order, and the spec and price behind it arrive within 4 working hours of your message.
Yes. After the spec and price (within 4 working hours), ask for a formal quotation on NVSX letterhead with the full specification sheet. The invoice carries GST in your company’s name, so the 18% comes back as input credit.
We can’t give a number yet: no two-GPU machine or server has been through our benchmarks. What we have measured is one RTX PRO 6000 serving 32 people at once on a 27B model. Your machine is sized by memory and runs your model before dispatch.
Loud enough that it belongs behind a door: plan a server room for any GPU server. The two-GPU workstation is fine on an office floor.
Our estimate for the Multi-GPU workstation at full load is about 1.05 kW with two 300 W Max-Q cards, or ₹10.60 for every hour it works at ₹10.11 per unit; with two 600 W cards, about 1.65 kW or ₹16.70 an hour. An eight-GPU server pulls several kilowatts; we work out its power and cooling with your team. Add a UPS to the budget.
Inside Pune, yes: we deliver and install it. Elsewhere in India it travels insured, ready to switch on, and our team brings it up with your IT staff over a video call.
The 3-year service warranty covers labour, diagnosis and claim handling; parts carry their makers’ warranties, and we run every claim for you. Anything dead on arrival we replace within 72 hours. Your IT team gets one WhatsApp number for the people who built it.
Yes. The model lives on the machine’s own GPUs, so it keeps answering offline, and prompts and files stay inside your network.
Spec and price inside 4 working hours, a formal quotation for the committee when you want it, and a straight answer if a smaller machine or cloud rental suits you better.
We’ll send your itemised build spec within 4 working hours.
Hey. I'm VOLT — NVSX's AI assistant.
Ask me about products, build recommendations, compatibility — or anything PC hardware.
