NVIDIA RTX workstation rollout india: a Bengaluru analytics team’s Friday install day

06:42 AM. The boxes arrived before the front shutter was open.

This is an NVIDIA RTX workstation rollout india story, told from the floor. The freight elevator at the HSR Layout office had a 2018 sticker on the side. Two folders, one barcode scanner, a watchman with opinions about the rain. Fourteen pallets came in. Each one wrapped in three layers because the data scientist who placed the order had been very particular about scratches on the chassis.

I was there because Sirius Star had specced and sourced the company’s first NVIDIA RTX workstation refresh. Their three year old Dell Precision 5560 mobiles with RTX A4500s had become the floor’s joke. Their AWS bill was not a joke. The data team had been spending between INR 11 and INR 14 lakh a month on AWS H100 bursts during model fine tuning cycles, per their CFO’s MIS pack. Procurement said yes in March. June was install day.

By the time the watchman had his second cup of thanda chai, eleven boxes were on the data team’s floor. The other three were waiting on a different lift because the facilities head had locked the goods lift at 6:30. Arre. That jhamela never shows up on a kickoff slide.

Why we stopped renting H100 hours

I keep coming back to this point. Cloud GPUs do not become expensive on the day you start using them. They become expensive on the day a CFO looks at twelve months of consolidated bills and asks why a 110-person firm is spending more on AWS than on rent.

Sandhya, the data team’s lead, had been running the numbers for a quarter. Their RAG model was fine tuned about once a week. An H100 on-demand at roughly USD 5 an hour per AWS price list, two GPUs for four hours, twelve weeks. Then add the prototyping bursts. Then add the analyst who forgot to spin down an instance over Diwali. The total was uncomfortable enough that her CFO asked me a question I do not usually hear from a CFO. “Can we just buy the boxes.”

Yes. You can. You buy fewer than you think. You size each one larger than the data scientist asks for. You add three years of on-site. And you do not buy DGX for a 14-person team, even when the vendor’s deck makes it look cheap. That deck is for someone else.

Get my free 4-hour quote 200+ businesses across India trust Sirius Star. Response within 8 working hours.

NVIDIA RTX workstation rollout india: what we actually shipped

The per-seat config landed here after three rounds with the data team and one round with their facilities head. Below is the BOM defence sheet I write for a CFO. It reads better one line at a time.

Per-seat all-in cost: INR 6.4 lakh including the chair upgrade their HR head insisted on. The chair was not in scope until 9:30 AM that morning. The chair is always the last line item.

Lenovo’s official ThinkStation P5 page lists the PSU and ISV certifications we screened against. The Dell Precision T5860 comparison sits in our Dell Precision versus Lenovo workstation piece.

11:30 AM. Bay 4’s PSU sang at the wrong frequency.

This is the part of the day that did not appear in the runbook.

Three workstations on the right side of the floor started flapping under sustained load. Power-good signal would drop for a millisecond. The GPU would throttle. A training run would crash. Not a hardware fault. Not a driver fault. It was the building.

The customer’s facilities head had re-wired the floor’s UPS distribution the previous week. Two new APC Smart-UPS units, each carrying half the load instead of the old single 10 KVA setup. The numbers worked on a clipboard. They did not work for three workstations because the new distribution put two RTX 6000 Ada chassis on the same leg as a 2 HP precision AC compressor that cycled every 90 seconds.

I should have asked. I did not. That was on me. We moved bay 4 to a different leg and the PSU stopped singing. Bas. Ninety minutes lost. Yaar, that was avoidable. The Andheri BPO Friday outage story covers the same problem from the other side.

2:15 PM. The BIOS that came in on a Wednesday batch

One chassis from a different delivery batch had skipped a BIOS sequence that turns secure boot on by default. The data team had a hard policy on secure boot. Their security head’s compliance log expected it. We re-flashed on the floor with a USB stick I had brought for an unrelated reason.

Ten minutes if you have the stick. Forty if you do not. Pakka rule for these refreshes now. Carry the stick.

I logged it on the customer’s tracker as “Wednesday batch BIOS missing default secure boot. Re-flashed on-site.” Their security head saw it at 5 PM and replied with a thumbs up. The DPDP Act guidance on technical and organisational measures reads cleaner when secure boot is on at the boot loader, not just in the OS.

4:15 PM. Sandhya opened VS Code on the new bay.

The data team lead sat down at the bay closest to the window. Two folders and a tablet. She had been on a remote SSH into AWS for three years. She opened a notebook that had been failing during long fine tunes. Ran it on the local RTX 6000 Ada. Three hours twenty. It used to be four hours on a rented H100. The H100 is faster per FLOP. The local rig won on round trip because there was no upload, no instance start, no S3 mount, and no instance billing for a meeting break.

The data scientist next to her asked, “Wait. This is on the box. Not in the cloud.” She nodded. He went back to his laptop and closed his AWS console tab. That tab had been open since their last refresh cycle.

When a data scientist stops opening her AWS console, the rollout is done.

Get my free 4-hour quote 200+ businesses across India trust Sirius Star. Response within 8 working hours.

What I would do differently if we did this again

I came into this rollout believing the BOM was the hard part. The BOM is not the hard part. The BOM is two weeks of haggling that ends in a spreadsheet. The hard part is the building.

One. Ask facilities about UPS distribution before you sign the order. Not after.

Two. Spec the PSU one tier higher than the GPU calculator says. NVIDIA’s RTX 6000 Ada product page recommends a 1000W minimum. We shipped 1300W. The headroom paid for itself when bay 4 was flapping at lunchtime.

Three. Do not put DGX on the slide for a 14-person team. The unit price is wrong, the depreciation curve is wrong, the cooling is wrong for a Bengaluru office floor. The Sirius Star NVIDIA India practice page spells out where DGX fits and where it does not.

Four. Read the chassis write up before you commit. Our Lenovo ThinkStation P5 buyer notes capture the decisions we made on this job.

Five. Budget for the chair. The chair is always in scope.

Key takeaways for a 14-seat NVIDIA workstation refresh in India

• Per-seat budget around INR 6.4 lakh including 3 year on-site is the realistic number for an RTX 6000 Ada workstation in India today.
• Validate UPS distribution with facilities before sign-off. Not after.
• Carry a secure boot USB stick on install day. Wednesday delivery batches occasionally skip the default.
• The receptionist test for an AI workstation refresh is when the data scientist closes her cloud console tab.
• DGX is not the default for a team under 30 people. RTX 6000 Ada handles 13B parameter fine tunes per seat without batching.

FAQ

How much does an NVIDIA RTX 6000 Ada workstation cost in India?

Per-seat all-in budget lands between INR 5.8 lakh and INR 7.2 lakh in 2026, depending on chassis, RAM, NVMe, monitor, and warranty term. The GPU alone lists between INR 7.5 lakh and INR 8.2 lakh in the channel as of this run, per our last three OEM quotes.

Do I need DGX, or are RTX 6000 Ada workstations enough?

If your team is under 30 data scientists and you fine tune up to 30B parameter models, workstations work. DGX comes in when you train above 70B, when you need NVLink across many GPUs, or when rack-level fabric latency is the limit. Most Indian teams I have watched hit the workstation ceiling at year three or four.

What about warranty and on-site support?

Buy 3 year on-site, next business day. RTX 6000 Ada parts are not stocked in every Bengaluru OEM service van as of mid 2026. Sealed-drive retention costs around INR 8,000 per seat per year and matters when client data sits on local NVMe.

What power and cooling do I plan for?

Plan for 600W to 750W per chassis under sustained training load. Validate UPS distribution with facilities first. If your precision AC compressors cycle on the same UPS leg, the GPU will throttle.

If you want a similar NVIDIA workstation rollout scoped for your Indian floor, the Sirius Star NVIDIA India practice is the front door. The same team wrote up the 42-Mac creative agency rollout for a different floor and a different brief.

Get my free 4-hour quote 200+ businesses across India trust Sirius Star. Response within 8 working hours.

P.S. Riya here. We shipped a similar setup for a Hyderabad credit-risk team last quarter. Eleven seats, RTX 5000 Ada, smaller datasets. Their CFO had the same question yours probably does. Can we just buy the boxes. Yes. The math holds at fourteen seats. It also holds at six if your cloud GPU bill has climbed for two quarters. Reply on WhatsApp at +91 91375 93228 and I will send the per-seat BOM and facilities checklist by end of day.