Private AI for SaaS enterprise customers.
Enterprise buyers ask where their data goes and who else can see it. We build the isolated AI stack, hosted where the contract needs it, with the logs and evidence that answer the security review.
What we build
When to offer single-tenant AI.
Not every enterprise customer needs it. Offer it where one of these is true, and keep the shared stack for everyone else.
The contract asks for it
The customer’s paper says prompts and documents must not reach a shared model, or must stay in India. A private setup answers this in the architecture, not in a promise.
The customer is regulated
Lenders, insurers and healthcare buyers send long AI questionnaires and often ask where each component runs. Our guide on why enterprise deals stall in AI security review shows what they look for.
The data is the product
If the AI feature reads contracts, claims, patient notes or credit files, the customer’s risk team will want it kept apart from other tenants.
The deal can carry the cost
A dedicated stack has fixed running costs that a shared one does not. Offer it where the contract can support it, and price it as a separate line.
When a shared stack is enough.
Some questions sound like a request for dedicated hardware but are not. Check what the customer needs before you build it.
“Do you train on our data?”
This is a contract and configuration question. A clear no, written down and backed by settings, often settles it.
“Where is the data stored?”
A shared stack in an Indian region may be enough if the contract asks only for India hosting.
“Can other tenants see our data?”
Strong tenant scoping on a shared stack can answer this. Dedicated hardware is not always the only route.
Four levels of isolation.
Choose the lightest level that meets the contract. Each step up gives the customer more control and gives you more to run. The wider Own Your AI page covers the building blocks.
Shared model, tenant-scoped data
What is kept apartDocuments, indexes, logs and caches are tagged and filtered by tenant. The model and the servers are shared.
When it fitsMost customers, when the rules ask for separation of data but not of compute.
Dedicated index and storage per tenant
What is kept apartEach tenant has its own vector index, object storage and keys. The model is still shared.
When it fitsCustomers who want clean deletion and export, and a clear answer on where their data sits.
Dedicated model deployment per tenant
What is kept apartA model server of its own for the tenant, with its own logs and its own version, inside your cloud account.
When it fitsCustomers who will not let their prompts touch a shared model, or who need a pinned model version.
Inside the customer’s own cloud account
What is kept apartThe stack runs in the customer’s account or data centre, and you operate it under an agreement.
When it fitsThe largest regulated buyers, who want the keys, the logs and the off switch themselves.
What it costs to run.
We give no price here, because it depends on the model, the hardware and the number of tenants. These are the cost drivers to size before you quote a customer.
GPUs that sit idle
A dedicated model server costs money even when nobody is asking questions. Share a GPU pool across tenants where the contract allows it.
Memory, not just model size
As a rough estimate, weights take about the parameter count times the bytes per weight: 2 for BF16, 1 for FP8 or INT8, about 0.5 at 4-bit. A 31B model is about 62 GB in BF16 and about 16 GB at 4-bit, before the KV cache and overhead, which grow with context length and users.
Operations per tenant
Upgrades, monitoring, on-call, backups and evaluation all multiply with the number of dedicated stacks. Automate the setup before you sell the second one.
Evaluation and model changes
Each model update needs a check on the tenant’s own documents. Pin versions and test before you roll out.
Support access
Your own staff reaching into a tenant stack is a data flow too. Log it and limit it.
India data-residency demands.
Indian enterprise buyers often write India hosting into the contract, even where the law does not. The full picture for Indian organisations is in our private LLM in India guide. This is a summary of the law, not legal advice; confirm your position with counsel.
DPDP does not require Indian hosting
The Act does not say personal data must stay in India. Section 16 lets the government restrict transfers to countries it notifies, and as of November 2026 none are notified. Section 16(2) keeps stricter laws in force, such as sector rules on where payment data is stored. Many buyers ask for India hosting anyway, in the contract.
You are usually the Data Processor
Your customer is the Data Fiduciary and stays responsible for the data (section 8(1)). You need a contract with them (section 8(2)), and they must have you erase the data they gave you once consent is withdrawn or the purpose is served, unless a law requires it to be kept (section 8(7)). Most core duties start on 13 May 2027. Read what customers ask SaaS processors for the questions to expect.
Count every place the data lands
The model server, the vector index, prompt and output logs, caches, backups, monitoring, support tools, and any outside API used for embeddings, moderation or routing. One call to an overseas service can undo an India-only promise.
A hosted router is another processor
A routing service that forwards prompts to many providers sees those prompts. For sensitive data, run a self-hosted gateway, such as the open source LiteLLM proxy, in front of models you host in India.
Where it can run in India.
The right home depends on the customer, your team and your cloud contracts. Check current regions, GPU types and terms with each provider before you commit.
Indian cloud regions
AWS has Mumbai and Hyderabad (Hyderabad needs opt-in). Azure has Central India, India South Central and others. Google Cloud has Mumbai and Delhi. GPU types differ by zone, so check before planning.
Indian GPU clouds
E2E Networks lists B200, H200, H100 and A100 GPUs. Yotta Shakti Cloud lists H100 and L40S, with B200 announced. Both have Indian data centres. Check current availability and terms with each provider.
The customer’s account or servers
For the strictest buyers, the stack runs where they already hold the keys and the audit trail.
IndiaAI compute
The IndiaAI portal offers subsidised rates for approved projects, filed monthly, per request. Eligibility follows the portal’s criteria. It is not something to plan production capacity on.
Permissions inside each tenant.
Isolating tenants is half the job. Inside a tenant, search must respect who may read what. Our guide to permission-aware retrieval goes through the options.
Tenant first, then user
Separate tenants, then check what each user may read inside a tenant. A model that can see a document can repeat it.
Filter in the database, not in the prompt
Postgres row-level security can restrict rows per user, but superusers and bypass roles skip it. With pgvector, filters can be applied after an approximate index scan, which can return fewer rows than you asked for.
Synced permissions go stale
If you copy access lists into the index, revoked access may linger until the next refresh. Where revocation must be instant, check the source system at query time.
Model licences for a SaaS product.
You are serving the model to your customers, so the licence matters more than it does for internal use. See open-weight model licences explained for a model-by-model view.
Licences are set per model
Llama 4 is under Meta’s community licence. It has a threshold of 700 million monthly active users, a “Built with Llama” display rule for distributors and a naming rule for derived models. Gemma 4 is under Apache 2.0. Qwen and Mistral differ from model to model.
Embedding or reselling matters
Some licences treat a Model-as-a-Service business differently from a product feature or internal use. Work out which one your product is before you pick a model.
Read the licence on each model card
This is a summary, not legal advice. Have your counsel review the licence for the exact model and version you ship.
What the buyer will ask to see.
A private stack only helps if you can show it. We write the answers with you, backed by how the system is built. Mono does not certify you or audit you. We prepare the evidence others review.
Where each part runs
A one-page map of model, index, logs, backups and support access, with the region for each.
Who can see what
Roles, tenant separation and how support access is approved and logged.
What is retained and for how long
Prompts, outputs, logs and backups, and how erasure reaches each of them.
Which model and under what licence
The version, who updates it and how changes are tested.
How it is tested and reviewed
Results on realistic data, incident handling and human review where it matters.
How it runs.
Phases, not promises of dates. We agree the plan with you once we know your customers, your data and your cloud setup.
- Step 1
Scope
We agree which customers need what, the data involved and the rules in their contracts.
- Step 2
Choose the isolation level
We match each customer type to one of the four levels and decide what stays shared.
- Step 3
Design
We choose hosting, the model and its licence, sign-in, permissions and logging.
- Step 4
Build
We set up the stack as repeatable code, so each new tenant is a configuration, not a project.
- Step 5
Test and document
We check answers on sample tenant data and write the evidence your buyers ask for.
- Step 6
Hand over
Runbooks and training, with everything in your accounts.
Related services.
Own Your AI
The overview of private AI that runs on infrastructure you control.
See Own Your AIAI security questionnaires
Answers and evidence for the AI sections of enterprise security reviews.
See questionnaire supportISO 42001 readiness
A management system for AI that buyers recognise. Mono prepares you; it does not certify.
See ISO 42001 readinessPrivate AI for SaaS FAQs.
When should a SaaS company offer single-tenant AI?
When a customer’s contract or risk team needs the AI kept apart from other tenants or kept in India, and the deal can carry the running cost. If the customer only wants a promise not to train on their data, a clear contract and strong tenant scoping on a shared stack may be enough.
Does DPDP make us host customer data in India?
No. The Act does not require personal data to stay in India. Section 16 lets the government restrict transfers to notified countries, and as of November 2026 none are notified. Sector rules can be stricter, and customer contracts often ask for India hosting, so it is common to offer it.
Is a dedicated model per customer always needed?
No. There are four levels, from a shared model with tenant-scoped data up to running inside the customer’s own cloud account. We help you pick the lightest one that meets each contract, because every step up adds running cost and operations.
Can we use any open model in a commercial SaaS product?
Not automatically. Licences are set per model and version. Some are Apache 2.0, some have user or revenue thresholds, and some treat selling model access differently from embedding a model in a feature. Read the licence on each model card and ask your counsel.
Do you certify our AI setup or give us a compliance certificate?
No. Mono is not a certification body, law firm or auditor, and there is no government DPDP certificate. We build the engineering, collect the evidence and prepare you for buyer reviews and audits that others carry out.