Private cloud is stopping innovation inside large organisations: Utho Cloud CEO

As Indian startups graduate from free cloud credits to full billing cycles, a pattern of cost shock is prompting a rethink of how infrastructure decisions are made. In a conversation with TechCircle, Manoj Dhanda, Founder and CEO of Utho Cloud, a cloud provider focused on the Indian market, discussed rising cloud costs, the GPU supply challenge, and how he sees enterprise and startup cloud strategies evolving over the next five years.

Edited Excerpts: 

How widespread is the cloud cost problem among Indian startups, and where do costs typically escalate?

Our user base is more than 95% India-based, and we are focused on the Indian market because that is where we see the biggest problem to solve. In terms of cost, three components drive it the most. The first is compute, CPU or GPU. The second is storage. The third, which can be harder to predict, is internet bandwidth, meaning egress charges. These three together are what result in high costs.

What we have done is focus on mastering those three areas. We did not try to build everything that AWS offers. We built roughly 20% of what hyperscalers offer, but that 20% addresses about 80% of what most users actually need. Areas like analytics, databases, and big data processing are not where we have focused. We focus on that 80% of the market.

Have you personally observed this cost concern becoming a real trend among Indian startups?

Absolutely. If you speak to any startup, you will see the same pattern. When a startup is deciding where to host its application, it typically looks at two things: how others have done it and the experience of its community. Since almost every SaaS company in their network is hosted on one of the three major hyperscalers, AWS, Google Cloud, or Azure, they naturally assume those are the only options.

That works fine for a year or two when they are running on free credits. But once the credits run out, the bills start arriving, and those bills can be quite high. We see many startups that are locked in, either by contract terms or because they have built their architecture around a specific provider's services, and cannot easily migrate out.

Part of this is a governance issue that starts early. When founders tell their teams to go build on AWS, and there is no one reviewing deployments or controlling resource usage, teams tend to over-provision because the credits make it feel free. A year later, when the real bills come in, the concern surfaces.

Beyond CPU costs, the situation has changed further with GPU demand. Since AI became a priority, the cost of running GPU workloads on hyperscalers is almost three times what local providers charge for the same.

How is the rise of AI workloads changing what cloud providers need to offer?

Before AI, it was essentially a CPU market. Any CPU could run most applications, and the challenge was largely about the availability of different CPU configurations. Now, with GPU-specific requirements, it is a different situation entirely.

A large language model might require an H100 or H200 GPU. A mid-sized or smaller model might run adequately on an L4 or L40S. Providers now need to maintain inventory across several GPU models, and making the right model available to a customer exactly when they need it is genuinely difficult. When we operated in the CPU world, any CPU broadly fit any workload. Today, we need to procure and stock multiple distinct GPU models and ensure availability by type. That shift, from one standard compute tier to multiple specialised GPU models, is one of the most significant operational changes in the provider world right now.

Are you seeing a shift among startups from a "scale first" mindset to more cost-aware infrastructure planning?

Yes, cost-awareness is the dominant posture now. People do not want to scale fast anymore. They want to have control. I think the industry has learned from the mistakes of the past decade with hyperscalers, scaling rapidly because credits were available or because CPU costs were not a concern. Now decisions are being made more deliberately, with cost as a primary factor.

How common are multi-cloud or hybrid setups today, particularly combining hyperscalers with regional providers like yours?

Out of every ten customers we speak with, all ten are thinking about or already pursuing a multi-cloud strategy. What we have not seen much of is a genuine hybrid setup, meaning a combination of private and public cloud.

The reason multi-cloud is becoming common is largely driven by GPU availability. Some GPU models are better available or better priced on AWS, some on Google Cloud, and some with local providers. Because the right GPU for a given AI workload may only be available on a specific provider, customers are distributing their workloads accordingly. The architecture that supports this is a distributed model, where models can be hosted anywhere, and inference or training APIs are wrapped around them so they can be consumed from a central layer. It is not quite a single-provider AI infrastructure anymore.

For large enterprises, what are the main barriers to switching cloud providers, given that cost is less of a pressure for them?

For large enterprises, the two main barriers are compliance and trust, in that order. Compliance is often addressable on our end because we operate out of Tier 3 and Tier 4 data centres that meet the same standards as hyperscalers. Trust takes more time to build.

What we do is offer free proof-of-concept environments to enterprises for three to six months. That gives their teams time to get comfortable with the platform, the environment, and our team, before they commit to moving any production infrastructure. The effort required to bring in a large enterprise is significantly more than for a mid-size customer.

There is also a segment of enterprises still running private clouds hosted directly in data centres. They carry concerns about data risk when moving to the public cloud. That said, things are changing. Engineers and DevOps teams within these organisations increasingly want to work with cloud-native technologies, and that is harder to replicate in a private cloud. Private clouds tend to slow down the pace of adoption of newer capabilities, which in turn limits innovation within the organisation.

Looking further out, how do you expect the industry to change in terms of how Indian companies choose cloud providers?

I think people will become more open and independent in how they choose where to host their applications. The industry will look quite different in three to five years.

Companies will increasingly run agentic AI systems and maintain their own model stacks, whether they are using public APIs or running their own inferencing. In terms of infrastructure, I expect the model to shift from enterprises provisioning dedicated GPUs or managing their own servers, to a model where large shared GPU clusters provide private, isolated inferencing environments for each customer. Companies will interact with their AI infrastructure through APIs or a user interface rather than managing hardware. Applications themselves will shift toward AI at their core — less about pulling data from a database and displaying it, and more about AI-driven actions and decisions.

I expect that by five years from now, more than 30% of users will have adopted this inferencing API model, similar to how managed databases work today. You host your database with a provider and just use it. You will do the same with AI models. That shift will also make users freer to move between providers, because the barrier to switching will be lower.

Things that used to take a year to change are now changing every month. But adoption in the enterprise world will lag behind the technology curve — many CIOs still feel the pace of change is risky. The technology will move fast; the adoption will take somewhat longer.