Operationalizing AI for the Enterprise

Manasi Vartak, Chief AI Architect & Vice President, Product Management at Cloudera discusses hybrid AI-anywhere strategies at Cloudera, workload portability, AI model training, and the role of low-code platforms in accelerating enterprise adoption of AI.

What advantages does a hybrid AI-anywhere approach offer compared to traditional cloud-only AI platforms?

Cloudera serves multinational customers across oil and gas, financial services, healthcare, and other industries. If you are a multinational organization, you operate across the Americas, Europe, APAC, and other regions. Each region has different regulatory requirements around where data can live. There are also language considerations; LLMs trained in English may not work well for Arabic, Hindi, or other regional languages.

Additionally, proximity to customers affects latency. All of this means you want to run AI where your data and customers are located.

Our hybrid AI-anywhere approach allows customers to pick and choose where they run AI, based on where their data resides. This delivers the best experience and aligns with where the industry is heading, toward sovereign AI.

Is this mainly about choosing the right hyperscaler, or does it also include on-prem deployments?

It includes both hyperscalers and on-prem. We have a very large on-prem footprint. With AI, especially large language models, beyond a certain scale, it becomes cheaper to run models within your own data centers.

We have case studies showing that if you have more than 100 models, it often makes more sense to run them on-prem. Depending on the workload, customers are increasingly moving in that direction.

Does this mean data center investments will rise again, especially driven by AI and LLM workloads?

Yes. We’re already seeing this. OpenAI, for example, is acquiring data center capacity globally, in the Middle East, Japan, and elsewhere. Oracle’s partnership with OpenAI is another example.

AI and LLM inference are key reasons data centers are again in high demand. Enterprises themselves are also investing. For large-scale workloads, owning GPUs and infrastructure can be more cost-effective than renting cloud GPUs, which are priced by tokens and become expensive at scale.

How effective is it for enterprises to move workloads between on-prem and cloud?

Workload portability is a core design philosophy at Cloudera. You should be able to move workloads seamlessly. If you are running an AI workload on-prem, moving it to the cloud should be a lift-and-shift operation, like a quick click.

Your platform supports traditional ML, generative AI, and agentic AI with both low-code and full-code flexibility. How does this help enterprises accelerate AI application development?

One of the biggest changes with Gen AI is that it’s no longer just coders building AI applications. Business users now participate as well.

Business teams understand workflows deeply. For example, support teams know how they triage cases. With low-code tools and natural language interfaces, they can now build workflows themselves. Something that would have taken days or weeks can now be done in hours.

This fundamentally changes how software is built inside organizations and significantly expands who can build AI solutions.

Can you explain the role of AI Workbench in your portfolio?

AI Workbench focuses on building AI applications and models. It supports research and development using notebooks, Jupyter, Python, and other tools. You can build applications, chatbots, or do data visualization.

Workbench, combined with AI Inference, forms the foundation for our AI Studios, which are low-code and no-code tools for building AI applications.

What role do AMPs play?

AMPs are blueprints that help customers quickly get started with specific use cases. For example, for fraud detection, AMPs provide an end-to-end blueprint within the Cloudera AI platform. These are primarily directed at developers.

Could you elaborate on AI Inference Service?

AI Inference Service is developed in partnership with NVIDIA. NVIDIA provides optimized versions of open-source models for GPU infrastructure. Our service allows customers to select open-source models and run them at scale within the Cloudera AI platform.

It is already available in the cloud, and we are bringing general availability for on-premises data services in January.

How do you see the role of human workers evolving in the age of agentic AI?

I view AI as enhancing human capabilities. AI still requires human judgment and guidance. Rather than replacing people, it enables teams to do more with the resources they already have.

How do you address governance and security across the AI lifecycle?

Governance is critical. Innovation moves fast, but governance ensures responsible adoption. Data governance includes access control, lineage, and understanding how data is used. Model governance includes approved use cases, data access, and whether data is shared externally.

We’ve developed internal governance frameworks at Cloudera and are incorporating these practices into our products to help customers adopt similar standards.

Can you elaborate on the role of synthetic data, especially in the context of training AI models?

A lot of times, organizations have PII data, such as social security numbers, card numbers, or other sensitive information, that they do not want to expose to AI models. Synthetic data helps generate data that looks similar to real data without compromising user privacy.

At Cloudera, we have built something called the Synthetic Data Studio. Customers can provide typical examples of what a dataset looks like, and the system generates data that resembles it, without using actual PII. You can then scale this significantly, for example, generating millions of rows of synthetic data that can be used for training.

Even in the largest foundation model labs, synthetic data is widely used to enable training at scale, and we are seeing this expand further.

Can synthetic data be generated for most industries and use cases?

Yes, it can be generated for any industry, as long as you have a clear understanding of what your prototypical data should look like. Synthetic data uses large language models to generate additional data that resembles what you already have.

One of the technical nuances is ensuring sufficient diversity in the synthetic data so that it accurately reflects the population. You don’t want a situation where only one type of user or scenario is represented while others are missing.

How do you ensure quality and avoid hallucinations in synthetic data?

One approach is called ‘LLM as a judge,’ where one model evaluates the output of another. This allows quality control at scale. Humans can still be involved, but typically only for a small percentage of cases.

How does Cloudera deliver enterprise-grade performance while maintaining cost control?

Cost control becomes critical during inference. Cloud GPUs and token-based pricing become expensive at scale. Running models on-prem with Cloudera can significantly reduce costs.

For scale, Cloudera’s DNA is managing large datasets; we manage 25 exabytes of data. Our platform is designed for extensibility and scale. Data Services 2.0 also enables faster adoption of new technologies.

What challenges do enterprises face when moving from partial to full AI adoption?

The first challenge is data readiness from the context of governance, access, and quality. The second is having a strong business case. Experimentation is fine, but eventually the CFO will ask whether it delivers value.

The third is governance, especially managing hallucinations and risk. The fourth is enablement and talent. AI requires new skills, and organizations are still upskilling their workforce.

Highlight a few roadmap priorities?

We’re bringing general availability of AI Inference on-prem in collaboration with NVIDIA in January. We’re also expanding AI Studios with low-code and no-code tools that allow anyone to build agents quickly.

Another focus is Data Services 2.0 and Anywhere Cloud, which lets customers deploy AI and data infrastructure flexibly across on-prem and cloud environments.