Self-hosted vs cloud: when data residency actually matters

The short answer

For most Indian businesses, cloud is the right answer. Self-hosting only makes sense if you have specific data-residency requirements (regulated industries, sensitive customer data) or you’re running enough volume that cost-per-token becomes painful. The “I want my data on my own servers” instinct is usually less important than people think.

The two questions to ask first

Before deciding, get specific answers to these:

1. Is there a regulatory requirement that your data physically stay in India? If yes (regulated finance, healthcare, government work), self-hosting may be required. If no, your instinct about data residency is probably aesthetic, not legal.

2. How much volume will you actually run? A team running 50 Claude API calls a day pays maybe ₹500/month. That’s not enough volume for the engineering cost of self-hosting to make sense.

If both answers are “no big deal,” cloud is fine. Use Claude.ai Pro or the Claude API directly. Use n8n Cloud. Move on.

What “self-hosted” actually means in this context

Two different things people lump together:

Self-hosting your orchestration layer (n8n). This is common and easy. You run n8n on a ₹500/month VPS in Mumbai or Bangalore. n8n is open source. Most workflows run fine. Your data is on your server, but the AI calls still go to Anthropic, OpenAI, or Google’s cloud.

Self-hosting the AI model itself (Llama, Mistral, etc.). This is harder and rarely worth it for business work. You need a server with GPUs (₹15,000–50,000/month). The models are 6–18 months behind frontier in capability. For most tasks, the quality drop versus Claude or GPT is significant.

When most people say “self-hosted AI,” they actually mean the first one — self-hosted n8n with cloud-AI calls. That’s reasonable.

When self-hosted-everything actually makes sense

A few specific cases:

• Regulated industries with hard residency requirements. RBI-regulated entities, defence work, certain healthcare contexts. The compliance team will tell you if you’re in this bucket.
• Customer data you’ve contractually promised never leaves your infrastructure. Some enterprise customers require this in their contracts. Read the contract before deciding.
• Very high volume. If you’re running millions of AI calls per month and the API bill is meaningful, self-hosting an open-source model can be cheaper. This threshold is high — usually six figures in monthly API spend before the math works.
• No internet at the runtime. If your AI has to run in environments without reliable internet (factory floors, remote sites), local models are the only option.

If none of these apply, you’re probably reaching for self-hosting because it feels responsible, not because it actually is.

The hidden cost of self-hosting

The cost most people don’t price in:

• Someone has to maintain the server. Updates. Security patches. Disk space. Backups. Monitoring. This is a real ongoing job.
• Self-hosted setups break in unique ways. A cloud service has someone whose job is to keep it running. Your server has you.
• Open-source models lag. Whatever you’re running locally today is six months behind Claude. Six months from now, the gap may be bigger.

For a small team, the engineering cost of self-hosting is genuinely larger than the API spend it saves. Just pay the cloud bill and put the engineering time into things that compound for your business.

What we recommend

The pragmatic stack for an Indian SMB:

• Self-hosted n8n on a ₹500/month VPS in India. Your workflow data and credentials stay on your infrastructure.
• Cloud Claude API. Pay-per-use. Data sent to Claude is processed by Anthropic but not used for training (this is in their terms; verify it for your specific contract).
• Cloud Notion / Google Workspace / etc. for the apps you already use.

This gets you most of the “data on your infrastructure” feeling without the maintenance burden of running everything yourself.

The honest version

If you’re choosing self-hosted AI because it “feels safer,” sit with that for a minute. What specifically are you protecting against? Who is the threat? Is the threat realistic for your business?

For 95% of Indian SMBs, the realistic threats are: laptops getting stolen, employees forwarding sensitive emails to personal accounts, and customers sharing screenshots on WhatsApp. None of these are solved by self-hosting your AI.

Solve the threats you actually have. The cloud-vs-self-hosted question is usually a distraction from the security work that would matter.

Last reviewed: 7 May 2026