COST ANALYSIS

Why GPU clusters in India cost 60% less than AWS

A line-by-line teardown of an 8× A100 cluster — the same workload, two providers two streets apart in Mumbai. Same hardware. Vastly different invoice.

VK Vikram Krishnan · CEO 8 min read · 28 Apr 2026

The setup

Last quarter we onboarded a research lab fine-tuning Llama-3 70B for a domain-specific assistant. They came to us with an AWS quote and asked: "is this normal?" The quote was for an 8× A100 80GB SXM4 node in ap-south-1 on a 1-year reserved-instance commit, plus 50 TB of EBS gp3, a Linux AMI, and 30 TB/month of egress to fetch their dataset and ship checkpoints out.

We priced the same workload on Glixy. Here's what came out the other side.

Line-item teardown

1. Raw compute (the GPU node)

An p4d.24xlarge on a 1-year RI in Mumbai works out to roughly USD $11.57/hr × 730 hours$8,446/month₹7.05L. That's the all-up rate before storage, networking, or anything else.

Glixy's equivalent — bare-metal 8× A100 80GB with NVLink, 96-core EPYC, 2 TB RAM — runs ₹2.85L/month on annual commitment. Saving: 59.5% on the compute line alone.

2. Storage

50 TB of gp3 at AWS list price is $0.08/GB/month × 51,200 GB = $4,096/month₹3.42L. Plus IOPS and throughput bumps for ML workloads — usually another $500+.

Glixy ships the cluster with 8 TB of NVMe-oF burst storage included; an additional 50 TB of warm pool runs ₹38,000/month. Saving: ~89% on storage.

3. Networking & egress

This is where AWS gets you. Egress from ap-south-1 is $0.1093/GB for the first 10 TB and $0.085/GB after. 30 TB/month of model checkpoint shipping out costs roughly $2,940₹2.45L. If your training also pulls a 5 TB dataset from a cross-region S3 bucket, that's another $250+ on inter-region traffic.

Glixy includes unlimited egress on Growth and Enterprise plans. The line is literally ₹0. Saving: 100%.

4. Reserved-instance lock-in

The AWS price above assumes a 1-year RI commit. If usage drops, you're paying for capacity you can't use. Glixy plans are monthly with 30 days' notice — you can shrink the cluster the day after a customer churn without burning capex.

This isn't a price difference, but it's a real cost: most teams overcommit by 30–40% to avoid running out, then under-utilize. On Glixy, you pay for what you use this month.

5. The hidden line: DevOps

To make AWS work for ML, you also need someone who knows EKS, GPU device plugins, fair-share scheduling, NVLink topology hints, FSx for Lustre or Weka or whatever your fast filesystem ends up being. That person costs ₹40–80L/year in Bangalore and is hard to hire.

Glixy includes Kubernetes, Kueue scheduling, Grafana dashboards, runbooks, and 4-hour priority support on Growth. You don't hire — we do.

The total invoice

AWS (1-year RI in ap-south-1): ~₹13.4L/month all-up.
Glixy (annual commit, Mumbai): ~₹3.6L/month all-up.

That's a 73% reduction. Even adding the cost of a senior MLOps engineer at Glixy's plan level (which we provide free), AWS still costs ~3× more for the identical workload.

Why is the gap this large?

Three structural reasons:

  1. We don't pay the hyperscaler tax. AWS has to fund retail margin, sales orgs, marketing. We don't. Our customers find us because their friends are already on us, not because of a Super Bowl ad.
  2. We bill in INR. When the rupee weakens against the dollar, AWS's price climbs. Our price doesn't.
  3. We're geographically optimized. Our racks are within 8 km of most Indian customers' offices. Egress is a non-issue because most traffic is internal. AWS prices egress to subsidize cross-region replication for everyone.

When AWS still wins

To be fair: AWS wins on three things we don't pretend to match.

For everyone else — startups training and serving models from India, banks running internal LLMs, product companies building AI features — the math is unambiguous. Run the numbers on your own quote against ours and we'll usually beat it by 50%+ on year-one TCO.

The honest disclaimer

List prices on AWS are negotiable. Enterprise commits with $X/year guaranteed spend get discounts. The 60% gap I started with assumes typical Indian-startup buying patterns: small enough that you don't get a bespoke deal, big enough that the bill hurts.

Bring us your AWS quote. We'll annotate it line-by-line. If we can't beat it, we'll say so.


📞 Get a side-by-side quote →


More from the blog: Build your own private LLM in 14 days · The Indian AI infrastructure opportunity