Stop assembling AI from ten vendors. Consume it as one service.
Most AI projects in India stall between the demo and production. The model works in a notebook, but then someone has to find GPUs, host the model, connect it to company documents, keep personal data out of prompts, satisfy the risk team and pay five foreign invoices in dollars. XcellAI AI as a Service is XcellHost's answer: every layer of a production AI stack — compute, platform, models, applications and governance — offered as a service you switch on, from one Indian provider.
Start small with pay-per-token serverless inference on open models, add a vector database and RAG over your documents, give staff managed Gemini, OpenAI or Microsoft 365 Copilot, and move steady workloads to reserved L4, RTX PRO 6000 or H200 GPUs. Underneath, the XcellHost 24×7 NOC and SOC run the platform in Mumbai, and an ISO/IEC 42001-aligned governance layer keeps an inventory, risk register and guardrails in place. One contract, one INR invoice with GST.
Five layers. 27 services. Pick any tile to go deeper.
Every tile is a service XcellHost already runs, with its own page, plans and tools. Filter by layer, or use the stack finder to light up the tiles your use case needs.
Prices are copied from each service's own page. * Indicative — confirm with XcellHost. INR, excluding 18% GST. GPU compute is always priced on quote. Microsoft 365 Copilot is a Microsoft licence; the figure is the indicative rate on our Microsoft 365 E7 page.
A layered stack, and one path for every request.
Click a layer to see what sits there, what XcellHost runs for you, and what you still decide.
—
- Your appweb, WhatsApp, Teams or API call
- GatewayAPI key, rate limit, PII masking, logging
- Retrievevector search over permitted documents
- Modelopen, Gemini or OpenAI — routed by policy
- Guardrailoutput checks, redaction, citations
- Answerstreamed back, metered in tokens
Six jobs, and the stack behind each.
Typical starting points for Indian organisations. Each shows the problem, the XcellAI services that solve it, and what you measure.
—
—
—
Models are the easy part. We run everything around them.
The same 24×7 NOC and SOC that run our cloud and security services operate the XcellAI platform from Mumbai.
GPUs, vector stores and logs in our Mumbai Tier IV data centre and partner DCs*.
01Endpoints, GPUs and pipelines monitored around the clock; P1s worked live.
02PII masking, output checks and audit logs on every managed endpoint.
03OpenAI-compatible API, open-weight models and a documented hand-over on exit.
04Tokens, GPU hours, cost per answer and quality, with next month's actions.
05Compute, platform, licences and services on one GST invoice from an Indian company.
06Built for Indian rules from the first prompt.
Keeping prompts, documents and logs in India does not make you compliant on its own. It removes a hard question and gives your risk team evidence. We help you meet these; you stay responsible for them.
Serverless and dedicated endpoints, vector databases and retained logs run in XcellHost's Mumbai data centre and partner DCs*. Managed Gemini, OpenAI and Copilot follow each provider's own data-location terms, which we map for you.
The Rules were notified in November 2025 with phased commencement: Consent Manager provisions from November 2026 and most obligations from May 2027. A breach goes to the Data Protection Board without delay, with a detailed report within 72 hours. Gateway PII masking and logs help you meet this.
Unveiled on 5 November 2025 under the IndiaAI Mission: seven guiding sutras and recommendations across six pillars, relying on existing laws rather than a new AI act. Our governance layer maps your inventory and risk register to them.
Approved by the Union Cabinet in March 2024 with ₹10,371.92 crore over five years; its seven pillars include IndiaAI Compute, AIKosh datasets and Safe & Trusted AI. We help you evaluate these options next to private capacity.
The April 2022 directions require reportable cyber incidents to be reported within 6 hours and ICT logs kept for a rolling 180 days in India. AI gateway and platform logs can be retained in Indian data centres.
The committee report of 13 August 2025 sets out 7 sutras and 26 recommendations for regulated entities. ISO 42001-aligned inventories, AI policies and monitoring give banks and NBFCs evidence for them.
Regulatory summaries reused from our AI Adoption Framework page (sources checked October 2026) for orientation only — not legal advice. Confirm current requirements with MeitY, CERT-In, the Data Protection Board or your sector regulator.
Pay as you go, reserve, or let us run it.
Mix all three on one contract. Most customers start pay-as-you-go, reserve GPUs once traffic is steady, and hand day-to-day running to us.
Try and grow
- Serverless open models, scale to zero
- OpenAI-compatible API
- Packs of ₹5,000, ₹25,000 or ₹1,00,000*
- Usage by key, project and model
- Runs on GPUs in Mumbai
Steady traffic
- Dedicated GPU endpoints or bare metal
- Private network, your own models
- Predictable latency, no cold starts
- Sized with our GPU estimator
- Monthly or longer terms
We run it for you
- Vector Database from ₹9,999/mo*
- RAG as a Service from ₹14,999/mo*
- AIOps from ₹24,999/estate/mo*
- Managed Gemini from ₹199/user/mo*
- 24×7 NOC & SOC, monthly review
| Engagement model | You pay for | Commitment | GPUs | Best when |
|---|---|---|---|---|
| Pay as you go | Tokens, audio minutes | None — prepaid credit | Shared pool, autoscaled | Pilots, spiky or low traffic |
| Reserved | GPU capacity per month | Monthly or longer | Yours alone | Steady volume, private models |
| Managed | Service plan per month | Monthly | Whatever the service needs | You want outcomes, not operations |
* Indicative — confirm with XcellHost. INR, excluding 18% GST. Credit packs and token rates are as on our Serverless Inference page; managed plan prices are from each service's page.
Build your AI stack. See the rupee total.
Tick the services you want, set quantities, and add fixed-price items to the cart. GPUs and anything scoped go on a quotation.
3 services selected
Tell us the job. Get the stack.
Four questions map your use case to services on each layer, an engagement model and a starting quote.
—
How many tokens, how many GPU hours?
Enter your own model rate per million tokens. GPU usage is counted in hours only — GPU prices are on quote.
requests = users × requests/day × days
input tokens = requests × input tokens per request (output likewise)
token cost = input ÷ 1,000,000 × your input rate + output ÷ 1,000,000 × your output rate
GPU-hours = GPUs × hours/day × days/month
Build it yourself, call global APIs, or consume it as a service.
| What matters | Build in-house | Global AI APIs direct | XcellAI as a Service |
|---|---|---|---|
| Where data is processed | Wherever you build it | The provider's chosen regions — check each service | Mumbai DC and partner DCs* for hosted models |
| Getting GPUs | Buy hardware, wait, own the capex | Not needed, but no control over capacity | L4, RTX PRO 6000 or H200 — shared or reserved |
| Model choice | What your team can host | That provider's models | Open models plus managed Gemini, OpenAI & Copilot |
| Billing | Capex plus many vendor bills | Card billing, often in USD | One INR invoice with GST |
| Governance & guardrails | Designed and built by you | Provider defaults; the rest is yours | ISO 42001-aligned layer, PII masking, audit logs |
| Running it 24×7 | Your team, every night | You still run the app around it | XcellHost NOC & SOC, Mumbai |
| First working pilot | Months, once hardware lands* | Days for a demo, longer for production | Proof of concept in up to 4 weeks* |
| Leaving later | Yours to keep | Prompts and code tied to one API | OpenAI-compatible API, open weights, hand-over |
From idea to production, in six steps.
Typical timings for one or two use cases*. Larger programmes run several tracks in parallel.
- 01DiscoverUse-case workshop: score ideas by value, data and risk; pick one or two.Week 1*
- 02DesignChoose layers, models and engagement model; map data flows for DPDP.Week 1–2*
- 03PilotProof of concept on your data with real users, measured against a baseline.Up to 4 weeks*
- 04HardenSecurity assessment, guardrails, SSO and logging before go-live.Week 5–7*
- 05ScaleReserve GPUs where traffic is steady; add users, sources and languages.Week 7–10*
- 06Run & improve24×7 operations, monthly usage and quality review, governance updates.ongoing