Krutrim was meant to be India's OpenAI. Founded in 2024 by Bhavish Aggarwal — the entrepreneur behind Ola and Ola Electric — the company entered the market with an ambitious frontier model agenda and rapidly achieved unicorn status, becoming India's first generative AI unicorn at a $1 billion-plus valuation. The declared goal was to build an Indian large language model capable of competing at the frontier: trained on Indian data, optimised for Indian languages, and sovereign in the sense that mattered to both government mission partners and enterprise customers wary of foreign model dependency.
Two years later, Krutrim has posted ₹3 billion (~$35 million at ₹85/USD) in FY2026 revenue — a 3× year-on-year increase — alongside its first annual net profit and operating margins above 10%. These are not the metrics of a company winning a frontier model benchmarking race. They are the metrics of a company that made a deliberate strategic choice: stop competing with OpenAI, Anthropic, and the Chinese open-weight labs on MMLU and context window size, and instead build a profitable cloud AI services business serving Indian enterprises that need managed AI outcomes, not raw model access.
This deep dive examines what Krutrim actually sells now, why the pivot worked, what the IndiaAI Mission's GPU subsidy programme had to do with it, how Krutrim's trajectory compares to Sarvam AI's, and what the playbook means for builders and investors in both India and the UK.
Three takeaways for Indian AI builders
Before unpacking the details, three conclusions stand out immediately for any builder or founder working in the Indian AI ecosystem.
First: profitability is achievable in generative AI if you stop trying to beat OpenAI. The frontier model race is capital-intensive, compute-intensive, and dominated by organisations with research budgets that dwarf anything an Indian startup can deploy. Competing on that axis is a choice to lose slowly while burning investor capital. Krutrim's 10%+ margins demonstrate that competing on enterprise service delivery — where the value lies in integration depth, domain expertise, and managed outcomes — is a viable path to sustainable economics in the two-to-three-year timeframe that most venture-backed companies must work within.
Second: cloud services at enterprise scale beat model-building margins in Years 2–3. Building a frontier model is a long-horizon investment with uncertain commercial payoff. Building managed cloud AI services for enterprises that will pay for measurable operational outcomes — reduced fraud rates, faster credit decisions, better customer service automation — is a near-term revenue opportunity with high switching costs once the integration is live. Krutrim reached 3× revenue growth by selling outcomes, not capabilities.
Third: the IndiaAI Mission GPU subsidy at ₹115–150 per hour is the structural moat, not the model. Access to 34,000 GPUs at 42% below market rate is a cost advantage that global competitors cannot replicate. This subsidy has an expiry — it is a government programme, not a permanent market structure — but for the window it operates, it enables Indian AI companies to build at a cost structure that makes margins achievable even at relatively modest revenue scale. Krutrim's profitability is partly a function of disciplined enterprise sales, and partly a function of compute economics that no non-Indian AI startup can access on the same terms.
Krutrim's journey — from frontier models to cloud
Krutrim's founding narrative was explicitly about the frontier. Bhavish Aggarwal launched the company in 2024 with a vision of building India's answer to GPT-4: a large language model trained primarily on Indian data, capable of handling Indian languages at a level that imported models did not, and sovereign in ownership and governance. The unicorn round came quickly, reflecting investor appetite for Indian AI plays and the credibility that Aggarwal's track record with Ola brought to the venture.
The early challenges were structural rather than execution-related. Compute costs for frontier model training are measured in hundreds of millions of dollars for truly competitive models — a scale at which even a well-capitalised Indian unicorn is fighting at a significant disadvantage relative to OpenAI, Google DeepMind, Anthropic, or xAI. Training data quality for Indian languages, while improving, remained a genuine constraint: the volume of high-quality, diverse text in Hindi, Tamil, Telugu, Bengali, and other Indian languages is substantially lower than English-language training data, creating a ceiling on Indian-language model quality relative to English-language performance. And the competitive landscape shifted dramatically with the release of high-quality open-weight models — Meta's Llama series, Alibaba's Qwen series, and DeepSeek's R1 — which provided capable base models for free, dramatically changing the economics of building on top of existing model weights versus training from scratch.
The pivot that followed was pragmatic. Rather than continuing to compete on benchmark performance against global open-weight releases, Krutrim reoriented its technical work toward building AI-powered cloud services that Indian enterprises actually needed: models fine-tuned for specific industry domains, deployed as managed services with Indian data residency, integrated into existing enterprise workflows rather than exposed as raw API endpoints. The result, in FY2026, was ₹3 billion in revenue, 3× year-on-year growth, a first annual net profit, and margins above 10%.
| Year | Focus | Revenue (est.) | Key milestone |
|---|---|---|---|
| 2024 | Frontier model building | — | Founded; unicorn round achieved |
| 2025 | Transition: model to cloud services | ₹1B (est.) | First enterprise contracts signed |
| 2026 | Cloud AI services at scale | ₹3B | First net profit; 10%+ margins; 25+ enterprise customers |
What Krutrim actually sells now
Krutrim's commercial product is not a model. It is not an API. It is an enterprise cloud AI service — a managed offering in which Krutrim deploys, maintains, and continuously improves AI-powered workflows within a customer's existing operations. The distinction matters enormously for understanding both the economics and the competitive moat.
The sectors Krutrim serves reflect where AI ROI is highest in India's enterprise market. In telecom, Krutrim's services address network optimisation and customer service automation — two areas where Indian operators face significant operational cost pressure and where AI can demonstrably reduce costs and improve customer satisfaction scores. In financial services, the focus is credit decisioning and fraud detection: AI models trained on Indian financial behaviour patterns, integrated into lenders' underwriting workflows and real-time transaction monitoring systems. In healthcare, the services span diagnostic assist tools and medical records processing — areas where the combination of AI capability and Indian data localisation is particularly valuable, given that patient data subject to India's Digital Personal Data Protection Act must remain within Indian jurisdiction.
Krutrim has disclosed 25+ enterprise customers across these sectors, though individual customer names have not been made public. The deployment model is consistently managed service rather than self-serve API: Krutrim's team handles model deployment, fine-tuning on customer data, integration with existing systems, and ongoing performance monitoring. Customers pay for outcomes and service continuity, not for raw compute or model weights. This is why the margins are achievable — managed services command a premium over raw infrastructure, and the switching costs of a deeply integrated workflow tool are substantially higher than switching a REST API endpoint.
Enterprise customers in India pay for managed outcomes, not raw model access. If you are building for Indian enterprise and pricing by the API call, you are leaving significant margin on the table. Build the workflow integration, own the outcome metric, and price on the value delivered to the business — not on the compute consumed to deliver it. Krutrim's 10%+ margins are the proof that this approach works at scale in the Indian market.
The IndiaAI Mission factor
Krutrim's profitability story cannot be told without accounting for the structural cost advantage that the IndiaAI Mission's GPU subsidy provides. The Mission has procured 34,000 GPUs — available to Indian AI companies and research institutions at ₹115–150 per hour, approximately 42% below the prevailing commercial market rate for equivalent compute. Krutrim is one of the 12 IndiaAI Mission LLM partner organisations, meaning it has had access to this subsidised compute pool for its model training and development work.
The arithmetic of this advantage is significant. If Krutrim's training and inference compute requirements run to, say, several thousand GPU-hours per month, a 42% reduction in compute cost relative to commercial cloud rates is a direct contribution to the margin profile. At ₹3 billion in annual revenue and 10%+ net margins, Krutrim is generating approximately ₹300 million or more in annual net profit. The compute subsidy may not be the only factor explaining this margin, but it is a genuine structural advantage that does not appear on a standard income statement — it is baked into cost of goods sold as a below-market input cost.
The broader implication for the Indian AI ecosystem is that the IndiaAI Mission subsidy is a time-limited moat. The 34,000 GPUs the Mission has procured represent finite capacity, and access to subsidised compute will eventually be reduced or restructured as the Mission's priorities evolve. For Krutrim and other Mission partners, the rational play is to use the subsidised compute window to build the enterprise relationships, integration depth, and switching costs that persist after the subsidy expires. Krutrim's 25+ enterprise contracts and managed service model are exactly the kind of durable competitive position that a temporary compute advantage should be used to build.
For more context on the IndiaAI Mission's compute programme and its 12 sovereign LLM partners, see our earlier coverage at /news/indiaai-mission-12-llm-partners-sarvam-bharatgen.
How does this compare to Sarvam AI?
Krutrim and Sarvam AI are the two most prominent Indian AI companies, and their diverging strategies in FY2026 make for an instructive comparison. Both are IndiaAI Mission LLM partners. Both were founded with ambitions around Indian-language AI. Both have achieved significant scale. But the paths they have chosen reflect fundamentally different bets about where durable value accrues in the Indian AI stack.
| Metric | Krutrim | Sarvam AI |
|---|---|---|
| Model focus | Cloud services prioritised over model development | Multilingual model stack (Sarvam-30B, Sarvam-105B) |
| Revenue model | Enterprise cloud contracts; managed service delivery | API access plus government partnerships |
| Profitability | First net profit FY2026; 10%+ margins | Raised $350M Series C; investing in model capability |
| Differentiation | Managed cloud, deep enterprise integrations | Indic language specialisation; sovereign model stack |
Sarvam AI's $350M Series C — covered in detail at /news/sarvam-350m-series-c-india-sovereign-ai — represents a different thesis: that the value in Indian AI ultimately lies in owning the model layer, particularly for Indic languages where imported models remain genuinely weaker. Sarvam is betting that its multilingual stack — trained on Indian language data at a depth no foreign model can match — will become infrastructure for a wide range of Indian AI applications, similar to the way Llama has become infrastructure for the English-language AI ecosystem globally.
Krutrim is betting differently: that in a world where capable open-weight models are freely available and frontier model performance gaps are narrowing, the durable value lies in enterprise integration and managed service delivery, not model ownership. Profitability at 10%+ margins with 3× revenue growth is a meaningful endorsement of this thesis — at least for the current market moment.
Krutrim's managed cloud services strategy faces a structural risk that becomes more acute as the business scales. AWS, Azure, and Google Cloud are all investing heavily in AI services for the Indian enterprise market, and all three have substantially greater resources for product development, sales, and customer support than Krutrim. If Krutrim cannot demonstrate a durable differentiation — whether through proprietary model fine-tuning, Indian data moats, or integration depth that the hyperscalers cannot replicate — its managed service premium is at risk of compression. The compute subsidy from the IndiaAI Mission is a temporary structural advantage. What replaces it as the durable moat when the subsidy eventually narrows is the question Krutrim's leadership needs to have answered before FY2028.
What UK builders and investors should take from this
The UK AI ecosystem has watched Indian AI development from a distance, often underestimating its maturity. Krutrim's FY2026 results are a signal that should change that assessment. A GenAI company achieving 10%+ net margins at ₹3 billion in revenue is not a story about subsidised compute and government mission support — it is a story about enterprise AI services maturing to the point where customers pay real money for real operational outcomes.
For UK-based builders and investors, three specific implications follow. First, Indian cloud AI is maturing at a pace that makes it worth monitoring as a source of enterprise AI service patterns that can be adapted for the UK market. The Indian enterprise buying behaviour — preference for managed outcomes over raw model access, high sensitivity to data localisation requirements, strong sector specificity in telecom, BFSI, and healthcare — is not unique to India. UK enterprises in the same sectors are navigating comparable buying dynamics, and playbooks that work in India often translate with relatively modest adaptation.
Second, the UK-India tech corridor creates concrete opportunities. Indian AI services built on data-resident Indian infrastructure can be packaged and offered to UK enterprises with Indian operations — banks, insurers, telecoms, and healthcare networks that have significant India-side workloads. Krutrim's managed service model, if it develops international distribution partnerships, could become a route to AI-powered services for UK enterprise customers who need India-side AI capability without building it internally. The IndiaAI Mission's partnership with NVIDIA, covered at /news/indiaai-mission-budget-double-nvidia-partnership-2026, provides further context on the infrastructure build-out underpinning this opportunity.
Third, the investor angle is becoming more tractable. A profitable Indian AI company growing at 3× year-on-year is no longer a speculative bet on a market that might develop. It is a demonstrated commercial model in a market with a clear growth trajectory. UK institutional investors and family offices that have been cautious about Indian AI exposure — waiting for proof of monetisation rather than funding model-building ambitions — have a clearer signal now that the category can generate real returns on a reasonable timescale.
The broader playbook that Krutrim's results illustrate is worth articulating: subsidised compute from a national mission programme, deployed against enterprise services targets, generates profitability before scale, which in turn creates the credibility and cash flow for international expansion. India is executing this playbook in AI. The UK's own sovereign AI investments — including GPU infrastructure and model development programmes — could follow a similar logic if structured with commercial sustainability as a design constraint rather than an afterthought.
The blueprint other Indian AI founders should study
Krutrim's journey from frontier model aspirant to profitable cloud services business carries lessons that apply well beyond Krutrim itself. For any Indian AI founder navigating the question of where to focus technical and commercial resources, the FY2026 results are a useful stress-test of the available strategic options.
Stop competing on MMLU. Compete on enterprise integration depth. The benchmark leaderboard is not a revenue model. The companies that win enterprise AI contracts in India are winning on the depth of their domain knowledge, the quality of their workflow integration, and the reliability of their managed service delivery — none of which appear in standard AI benchmark comparisons. If your company is spending engineering cycles on benchmark performance that customers are not measuring when they make buying decisions, you are optimising for the wrong metric.
Use the IndiaAI Mission GPU subsidies while they last — 34,000 GPUs at 42% below market is a finite resource. The application process for Mission partner status is open to Indian-registered companies working on AI for Indian use-cases. If you qualify and have not applied, the cost advantage you are leaving on the table is directly calculable: take your monthly compute spend, multiply by 0.42, and that is your annual disadvantage relative to companies like Krutrim that are inside the programme. For a growth-stage startup, that gap can be the difference between a positive and a negative margin on your cloud AI service delivery.
Target sectors with high AI ROI and low existing automation. India's BFSI sector — banking, financial services, and insurance — has enormous AI ROI potential in credit decisioning and fraud, and remains substantially less automated than equivalent sectors in the US or UK. Healthcare AI is at an earlier stage still, with large digital records infrastructure gaps that AI-powered processing can address. Tier-2 and Tier-3 telecom customers — regional operators and MVNO-adjacent businesses — have significant customer service automation needs and are often underserved by the largest enterprise AI vendors. These sectors are not glamorous from a research perspective, but they are where Indian enterprise AI contracts are being signed and renewed.
Build managed service depth rather than API surface area. Krutrim's 25+ enterprise customers are sticky precisely because the integration work required to onboard them is also the work that makes switching expensive. Each additional workflow Krutrim integrates within a customer's operations increases the switching cost. API customers can switch providers in a sprint cycle; managed service customers face months of re-integration work. At 10%+ margins, Krutrim can afford to invest in the integration depth that creates this stickiness. Early-stage founders should design for managed service delivery from the outset — even if the initial version is relatively lightweight — rather than defaulting to API-first models that optimise for developer adoption at the expense of enterprise retention.
For context on the GPU cost environment that makes all of this more achievable than it was two years ago, see our coverage of Neysa's $1.2B Series B and India's sovereign GPU cloud ambitions — which documents the compute infrastructure investment underpinning the broader Indian AI ecosystem.