By Abhijit Ahaskar & Uday Bhaskarwar
Oracle’s AI-first strategy has helped it avoid the overall volatility faced by the broader SaaS sector. The American company is firmly established in the AI landscape with gigawatt-scale GPU clusters and dedicated cloud regions, and is incorporating GenAI models and AI agents across its software portfolio. The company’s multi-cloud strategy allows customers to continue using their applications on competing cloud platforms.
In an exclusive conversation with CXOtoday, Pradeep Vincent, Senior Vice President and Chief Technical Architect, Oracle Cloud Infrastructure (OCI), talks about the blurring lines between AI and applications, expansion opportunities in India, the importance of protecting private data on GenAI, and the transformative role of AI in software. Vincent was in Mumbai to attend the Oracle AI World event on February 10th. Edited excerpt:
Q. What are the criteria for selecting the AI models offered in Oracle Fusion Cloud applications and how do you provide them to your customers?
Our goal is to become a model Switzerland. We want to offer any model you want on our platform. The world of models is rapidly evolving and will continue to do so. There are some basic, large-scale models that are good at a variety of features, but may not be cost-effective. Additionally, there are models that focus on coding, models that focus on inference, and even models that are smaller and optimized for cost efficiency. Some models offer better performance in certain languages.
We expect customers to require a large number of models, but that will likely change as well. Our goal is essentially to enable them. Therefore, we will consider their demands and work with model providers to deliver.
Q. Are there any plans to build your own models?
There are several internal models. These are used for domain-specific purposes, such as metadata generation mechanisms and the process of populating vector databases. These will be considered part of the Fusion ecosystem. Regarding base models, our goal is to partner with other model providers rather than developing our own large scale base models.
In the long term, our goal is not to have a proprietary model. We are in a very dynamic industry, so we don’t want to be tied to any particular model or company. We want to keep our options open so that we can provide our customers with the best solution as quickly as possible.
Q. What is your take on what happened last week in terms of the impact of AI on the SaaS sector?
Regardless of what happened last week, I think AI will transform the software industry. When it comes to internal productivity tools, we can see that the way software is developed has changed over the last year. A year ago it wasn’t that good, but now it’s actually a lot better. Internally, we use this quite a bit for our own software development. Therefore, we expect many changes that will manifest themselves in the market in their own way. I believe AI will revolutionize the software world. What we should focus on as engineers is already changing, and will continue to change. Being a good programmer is no longer a big deal. We need to shift our focus to what humans are good at.
Q. We do not charge additional fees for Gen AI or Agent AI features that are part of Oracle Fusion applications. However, there is a GPU cost involved. How do you absorb it?
We don’t differentiate between apps and AI. In the early stages, we will ship specific AI capabilities. But the trajectory is that AI is an app. they are one and the same. In that world, GPUs are no different from computing, and cost models and pricing essentially include that.
Q. With Oracle Alloy, you essentially relinquish control that many hyperscalers keep. So how do you think this initiative will benefit you in the long run?
Control is not our differentiator. Our differentiators are our technology and operational capabilities. In the context of Alloy, our ability is to provide automation for third parties to operate. One of the things I get asked is how is this different from on-premises? The reason we’re able to provide Alloy with a platform like this is because we run fairly large and complex operations and we learn a lot from them.
Alloys are an important business, but they also have a very important public realm. We’re applying the lessons we learned from that to automating Alloy operations so that Alloy operators have a slightly better framework. After all, engineers aren’t the ones writing the code. Therefore, it is necessary to provide a cleaner interface for operation. This will be even more difficult to do if you do nothing. Essentially, the fact that we operate in a public region translates directly into the automation that Alloy can provide.
Q. How do you plan to democratize access to OCI superclusters for medium-sized enterprises where cloud taxes may be prohibitive?
Clusters are determined by the scale the customer wants, not how we want to sell. In smaller regions, we actually ran GPU racks without clusters. We also implemented smaller clusters based on customer requests. So most customers won’t find it prohibitive as long as they need a GPU. There are predefined choices when it comes to cluster sizes that you can choose from. Among them they can choose what they want. It wasn’t a big friction point.
Q. India is building its own sovereign AI model and Oracle is one of the companies providing the computing infrastructure. So how will joining India’s AI mission help Oracle expand?
We have operated GPU clusters around the world. There is a very large cluster in the United States. To some extent, it depends on the amount of energy available and the amount of customer demand seen in the area. However, clusters are also operating in other parts of the world, such as Europe and Asia. We are actually very excited to expand our footprint in India. Recent tax deferrals and data location regulations by the Indian government are encouraging the expansion of data center footprint in India. We look forward to working with our customers and partners to expand our footprint here.
Q. What are some of the scalability challenges that enterprises face when scaling their GenAI operations? How is OCI designed to address these challenges?
From the beginning, we have focused on both scaling up and scaling down all of our computing products. As a cloud infrastructure company, we have extensive experience in this. Very large gigawatt scale clusters are planned. At the same time, there are dedicated regions where you can scale down.
This all applies to AI as well, whether it’s GPU infrastructure, AI databases, GenAI infrastructure, or apps running on top of it. All of these can be deployed in any of these regions. We have unique experience not only in scaling up, but also in reducing the footprint and making smaller deployments effective.
Q. What is the biggest enterprise security risk from GenAI and what can be done about it?
From a security perspective, AI increases challenges for enterprises in a number of ways. AI functions like the human brain to some extent. There is no ability to forget something because there is no mechanism for forgetting something. Therefore, from a compliance and security perspective, it is very difficult to rigorously prove that you forgot something. This has implications in terms of processing corporate data, especially personal data.
If you train a model on private data that is accessed by a user who is not expected to see the data, it becomes difficult from a governance and compliance perspective to prove that the data was not accessible to that user. This is because the model is now trained using that data. To avoid this, don’t use private data to fine-tune or train your model. You need to keep them separate, and that’s a core part of your strategy.
In general, a data management mechanism is required that does not allow private data feeding or training. This is critical for enterprise adoption of AI. Writing poetry is one thing, but working with secure corporate data is quite another.
