Satyam Santosh explains why AI startups must look beyond cloud credits and build flexible, cost-transparent infrastructure using open architectures, hybrid cloud, and workload-driven strategies to scale sustainably.
As India’s AI startup ecosystem matures, founders are moving beyond the race to build models and focusing on a more fundamental challenge—building infrastructure that can scale sustainably. Rising GPU demand, unpredictable cloud costs, concerns over vendor lock-in, and growing regulatory expectations are pushing startups to rethink how they design, deploy, and manage their AI infrastructure from the outset.
In this interview, Satyam Santosh, Startup Program Lead APAC at OVHcloud, argues that infrastructure should be treated as a long-term business strategy rather than a short-term operational decision. He explains why startups should embrace open architectures, workload-based infrastructure, and cost transparency, while preserving the flexibility to adopt hybrid and multi-cloud environments as they scale. He also discusses the growing demand for bare metal infrastructure, democratizing access to high-performance compute, and how infrastructure providers must evolve their commercial models to better support AI-native businesses.
CIO&Leader: We are seeing an increasing number of growth-stage AI startups hit a financial wall once their introductory cloud credits dry up. From your vantage point, how should founders structure their multi-cloud or hybrid strategy early on to ensure architectural flexibility and avoid vendor lock-in down the line?
Satyam Santosh: Introductory cloud credits are helpful, especially when a startup is moving from idea to product. One mistake we often see founders make is treating credits as an infrastructure strategy. Credits should buy time to test, learn and ship, but the architecture should be designed for what comes after those credits run out.
Founders should think about portability much earlier than they typically do. Core data, model artifacts, deployment pipelines, identity, monitoring and orchestration should be built around open standards wherever possible. Containers, Kubernetes where relevant, infrastructure-as-code, open APIs, S3-compatible storage and well-documented data formats can give teams greater flexibility as infrastructure requirements evolve. It is equally important to recognise that not every workload needs the same infrastructure environment. AI companies have very different infrastructure needs across experimentation, training, inference, customer-facing applications and sensitive datasets. A founder can use public cloud for speed, bare metal for sustained compute, local environments for latency-sensitive workloads and hybrid models for regulated data. The goal is to match workloads to the environment that best supports them, rather than placing every system under one provider by default.
Multi-cloud does not necessarily mean using multiple providers from day one. It means preserving the freedom to make infrastructure choices as the business evolves.
As startups grow, it is worth periodically testing how easily critical workloads can be moved or replicated elsewhere. Understanding how readily workloads can be redeployed outside the primary cloud, how data can be moved without disrupting existing pipelines, and what the associated costs might be can help identify dependencies long before they become expensive or difficult to address. Multi-cloud does not necessarily mean using multiple providers from day one. It means preserving the freedom to make infrastructure choices as the business evolves. For AI startups, that flexibility becomes increasingly important as workloads scale and infrastructure costs become a larger part of the business.
CIO&Leader: Hidden fees have become a major point of discussion in cloud unit economics, sometimes accounting for unexpected spikes in monthly storage bills. For organisations constantly moving massive datasets for model fine-tuning, how critical is absolute cost transparency over premium managed services?
Satyam Santosh: For AI companies, cost transparency is critical because infrastructure is no longer just a background IT expense. It directly influences product margins, customer pricing and the pace at which teams can experiment and scale. As AI workloads mature, data is constantly moving between storage, compute and application environments for activities such as fine-tuning, model evaluation and customer-specific deployments. For many founders, the challenge is not the headline infrastructure cost but understanding how data movement, storage, retrieval, API usage and scaling behaviours affect the total bill over time. When those costs are difficult to estimate, it becomes harder to understand the true cost of a model, a customer deployment or a new product feature. Premium managed services can bring significant value by reducing operational complexity and helping smaller teams move faster. The more important question is whether founders can clearly understand the commercial impact before committing to those services. Hidden fees become particularly problematic when an AI startup moves from proof of concept to production. A demo may involve limited data movement, while a scaled deployment can generate substantial volumes of training data, embeddings, logs, evaluation datasets and customer-specific outputs. What appears manageable during the experimentation phase can quickly become an impactful cost consideration as workloads scale.
Infrastructure is no longer just a background IT expense. It directly influences product margins, customer pricing and the pace at which teams can experiment and scale.
Cost visibility should be part of infrastructure planning from the outset. Understanding how storage, network traffic, retrieval patterns, monitoring and model lifecycle management affect overall spend makes it easier to forecast growth and avoid unexpected surprises. Managed services absolutely have a role to play, particularly for teams looking to move quickly, but pricing should remain predictable and easy to understand. When founders have clear visibility into how infrastructure costs scale with usage, they can make better product, investment and growth decisions. As AI workloads grow, having a clear understanding of infrastructure costs becomes increasingly important for long-term planning and sustainable growth.
CIO&Leader: While India’s application-layer AI ecosystem is booming, founders frequently report bottlenecks when trying to access high-performance GPU compute. Is the domestic startup surge hitting a physical infrastructure ceiling, and how can the industry better democratize access to high-performance computing?
Satyam Santosh: India’s AI startup ecosystem has grown rapidly over the past few years, particularly at the application layer. We are seeing founders build across agentic AI, copilots, language technologies, healthcare, financial services, analytics and a range of industry-specific use cases. As these companies scale, demand for high-performance compute is increasing as well. From our conversations with startups, the challenge is not simply access to GPUs, but access to the right infrastructure at the right time, with predictable pricing, strong performance and enough flexibility to support different workloads. A model training company, a computer vision startup and an enterprise inference platform will all have very different requirements in terms of compute, memory, storage and networking.
The conversation is also broader than raw GPU availability. Startups need reliable access, shorter procurement cycles, flexible consumption models and support in making informed infrastructure decisions. The bigger gap is expertise in using the available infrastructure. We regularly see early teams burn through compute on training runs that were never optimised, or over-provision because nobody showed them what efficient looks like. Many early-stage teams are still learning how to optimise models, improve data quality and manage inference costs efficiently, which means access to expertise can be just as important as access to hardware. Democratising high-performance computing will require a combination of expanded infrastructure capacity, ecosystem collaboration and programmes that make advanced compute more accessible to startups. The goal should be to ensure that promising founders can access the resources they need regardless of their stage or location. When that happens, startups can spend more time building products and solving customer problems rather than navigating infrastructure constraints.
CIO&Leader: There is a noticeable shift from convenience-led cloud buying—where everything is hosted under one monolithic provider—to control-driven infrastructure setups. Are you seeing Indian deep-tech startups actively choosing bare metal or localized cloud environments to protect their core IP and manage performance latency?
Satyam Santosh: Yes, especially among deep-tech companies where the core value sits in proprietary models, datasets, inference systems or domain-specific workflows. In these businesses, infrastructure is increasingly becoming part of the product strategy because it affects performance, customer trust, compliance requirements and intellectual property protection. The shift is visible across sectors such as AI security, financial technology, healthcare, manufacturing, robotics, geospatial intelligence and video intelligence.
Infrastructure is increasingly becoming part of the product strategy because it affects performance, customer trust, compliance requirements and intellectual property protection.
Founders are asking sharper infrastructure questions than they did a few years ago. They want to know where the data resides, who can access it, how predictable is latency, and how much control do they have over the hardware stack. Organisations are also planning for ten-fold growth in usage. These considerations are driving greater interest in infrastructure models that offer more control and flexibility, particularly for companies building proprietary technology and serving regulated industries.
Bare metal has returned to the conversation because many AI workloads are resource-intensive and performance-sensitive. For sustained training, high-throughput inference and specialised workloads, it can offer more predictable performance and greater control over infrastructure. At the same time, localised cloud environments are becoming increasingly relevant for startups serving enterprises, public institutions and regulated sectors where data governance and latency are important considerations. Mature founders are adopting workload-based architectures, using different environments for different requirements rather than relying on a single approach. Public cloud, managed services, bare metal and localised environments each have a role to play depending on the workload, performance requirements and sensitivity of the data involved.
CIO&Leader: Given how quickly model architectures change, operational agility is a significant competitive advantage. How should infrastructure providers rewrite the standard contract playbook to give companies the flexibility to scale compute resources up or down without facing heavy penalties?
Satyam Santosh: AI companies build in cycles. They test a model, scale a workload, pause an experiment, change an architecture, switch GPU requirements, add customers and revise deployment patterns. Contracts should reflect that operating rhythm. Many AI startups need the flexibility to reserve capacity for predictable workloads while retaining the ability to scale up or down as requirements change. Experimental workloads often look very different from production workloads, which means long-term commitments may make sense in some cases, while shorter terms and greater flexibility are needed in others. The ability to move between different compute configurations, storage tiers and deployment models can also be important as model architectures evolve and infrastructure requirements change.
Flexibility should extend beyond compute capacity alone. Transparent pricing, straightforward scaling policies and support for workload migration can make it easier for startups to adapt without being locked into decisions made at an earlier stage of growth. Many founders also benefit from guidance around capacity planning, workload optimisation and cost forecasting, particularly as infrastructure requirements become more complex. Ultimately, infrastructure contracts should support experimentation, iteration and growth rather than assume a fixed operating model. When commercial terms align more closely with how AI products are actually built, startups can focus more of their time and resources on product development and customer outcomes.