David Cushman — HFS Research[00:21]
Hello and welcome to this conversation about generative AI. I know you’ve heard a few, but I think we’re going to go in a slightly different direction to those mostly discussed today. I’m with Brijesh Singh, and he’s a Senior Vice President and AI Head at Wipro. Hi Brijesh.
Brijesh Singh — Senior Vice President and AI Head, Wipro[00:42]
Hey David, always a pleasure to meet you, and looking forward to a lively and productive conversation.
David Cushman — HFS Research[00:42]
Let’s hope so. So the reason I frame this as a slightly different conversation about GenAI is we’re going to talk about the cost of the thing. I think it’s something that is a little bit swept under the carpet. One, there’s this idea of, you know, what are the true costs of generative AI, and then there’s this larger, maybe scarier thought: what might be the environmental cost of generative AI, AI in general? So, Brijesh, first of all, from an enterprise’s point of view, what should they really be thinking about when they’re looking at putting the cost base together for this?
Brijesh Singh — Senior Vice President and AI Head, Wipro[01:24]
What is the advisory cost? What is the execution cost? What is the pre-preparation cost? What is the build cost? What is the ongoing maintenance and support cost? What is the cost related to evaluating the various models that are out there? What is the cost of governance? Are there any redundant costs that are baked in? And when we are doing all of these things — evaluating all the models, preparing models, training models, tuning models — then what is the cost to the environment? So many models are getting created, almost on a weekly basis, so what impact does that have on the environment — the adverse impact we’re talking about — because all these models, when they’re getting created, they’re getting trained and tuned with billions, if not trillions, of tokens. And when they’re getting tuned with these billions of tokens, then they are running those graphical processors, the neural processors, tensor processors, and so on, which are consuming calorie at a very high rate. So burning that energy — today we are burning fossil fuels — that is having an adverse carbon footprint on our planet. So all those things need to be considered. And I just talked about the build cycle. Imagine the run cycle: when we have employed and deployed these models, then LLMOps kicks in. And when LLMOps kicks in, then these models require continuous training, continuous improvement, human in the loop, machine in the loop, continuous alignment with data, new data, additional data, and all those things are burning calories. So we really need to evaluate what are the various costs that are coming into play. And then, which use cases make more sense for us against this cost.
David Cushman — HFS Research[03:39]
The focus away from productivity and towards value is going to prove really important. I think it’ll become even more important when we start understanding that total cost of ownership, because there are certain ways in which you can reduce those costs and they’ll be consistent, but it’s always going to be ultimately about how you build value on top of that. But I’m just wondering whether — so let’s say you’re an enterprise leader — do you really care how much it’s cost in energy terms to build the model in the first place and to train the model? You’re more interested maybe in the run cost of it, for yourself. So you’re buying someone else’s model; yes, it will have taken a great deal of energy to build in the first place. My understanding is that most — if I look at OpenAI, for example, Microsoft behind that — there’s a real focus on: by 2025, Microsoft wants to make all of its data centers carbon neutral. So that will have a knock-on effect in terms of the impact on the environment of their generative AI offerings, and I’m sure there are a lot of companies that are working towards that. And it’s very easy to become very scared very quickly by some of the figures we see associated with, let’s say, the energy use of Nvidia’s latest chip — the H100, is it? It actually runs about twice as energy intense as, let’s say, the best available competitive AMD chip, but it’s 500 times better at delivering against AI and ML tasks, at least. So you’d have to say the direction of travel, whilst it might look quite scary initially in terms of impact on the environment, perhaps it’s not quite as terrifying as we think. Well, what do you think, Brijesh?
Brijesh Singh — Senior Vice President and AI Head, Wipro[05:28]
All Fortune 500 companies, more or less, and global 2000 organizations across the planet, and even the boutique shops — we are all becoming environmentally conscious. So because we are doing that, I believe we are all going to be conscious about how we leverage these models, these GenAI models. How do we minimize? There are techniques, there are technical solutions that are getting created that, instead of doing multi-shot prompts to the model to get the information that I’m looking for, how do I do it in a single shot, right? We at Wipro have created solutions as a part of our GenAI Studio that address many of these areas. So if you take a step back and go to the pre-GenAI world, the performance of the model is always dependent on the data. We picked various models, various algorithms available to us, and did some experimentation with data to understand which one is yielding me a better score, better performance, better value, and then we said we are going to go forward with that model and deploy in production and run with the data. The same thing is applicable with GenAI models. The difference is, here the scope has become vast, in the sense that now an organization has a 300, 400, 500 wish list of use cases. Now, a very important point to remember: for each of these use cases one has to evaluate the existing models in the marketplace to figure out which model will work better with my particular data, because the general-purpose models may not be adequately trained with that business function and with your business data the way it needs to be, right? So you need to do that evaluation, and that evaluation can only be done if you prepare your data, you contextualize your data, and you work with the model in doing some embeddings and evaluating the models. And for every use case, for all your business functions, that’s going to cost money and that’s also going to increase the carbon footprint, as we discussed, because each time you’re working with a model, you are burning the calories at the back end because you are running those GPUs, NPUs, CPUs, right? So now we need to be prudent about it — that as an enterprise, companies like ours are taking a leap forward with our enterprise generative AI solution that we have created as our GenAI Studio. Data chunking, domain guardrails, restriction on multi-shots, reducing hallucination, increasing accuracy, and so on are the techniques that we have built, which are going to create lesser and lesser impact on the planet. So they both save money for the enterprise and reduce the carbon footprint of using the technology.
David Cushman — HFS Research[08:55]
You got it.
Brijesh Singh — Senior Vice President and AI Head, Wipro[08:55]
Improving sustainability and reducing the cost for the enterprise — not only by reducing the cycle, so we can do it faster because of the accelerators and guardrails that a company like ours is bringing to the table, but also minimizing the impact in terms of consumption of cloud resources. So compute, storage, network — all those consumptions will be reduced in the build, fine-tuning, and training cycles.
David Cushman — HFS Research[09:22]
So this is good, because we started off listing a whole range of costs that actually could be quite scary, right? You think, oh, well, hang on, I’ll leave that alone for a bit. But now, I think we’ve got — yes, we can deliver this within some constraints that mean we’re not going to burn through every spare dollar you have, and there’s a future for the planet. I guess the final part of this jigsaw is knowing that TCO, knowing that total cost of ownership of whatever the GenAI is that you apply to a particular process, or a whole function, or even the whole business. How do you think that can inspire enterprises to focus on value rather than simply cutting costs and boosting productivity, which is a lot of what we’ve seen so far?
Brijesh Singh — Senior Vice President and AI Head, Wipro[10:21]
That’s the $1 trillion question — as the market is supposed to be $1 trillion. That’s the $1 trillion question: how do organizations prioritize and evaluate the right use cases and then execute the right use cases? So every day, when I’m working with my clients, this is the number one advice that I’m coming out with: that an organization needs to look at it holistically. That organization, particularly with the AI-enabled transformation program that everybody has embarked upon, they have to think about it — how do I create the common services, common platform, common architecture, common set of APIs, and so on, that can be leveraged by all the lines of business, all the use cases that I have in my organization, to reduce my TCO, to increase my speed to market, to improve the value that I’m bringing to my organization? Once I do that, and then have a governance framework around it, where when the use cases are coming in, there is a process that I’m putting in place for which ones are going to be positively impacting, say, my increase in revenue — which ones are going to be positively impacting by bringing more revenue for the company at a higher margin, right? Which ones are going to be impacting by taking significant cost out and improving the experience of the stakeholders?
David Cushman — HFS Research[12:06]
So I think what you’re telling us, Brijesh, is that the low-hanging fruit is fine — and of course you’ve got to pick it — but move swiftly, get a bit of ambition about making use of what this thing does that is different in creating value, rather than robotically hacking away at costs and not really seeing the value. Thank you for your insights, and if we get to meet soon, I’m looking forward to so doing. Thank you very much for your time.
Brijesh Singh — Senior Vice President and AI Head, Wipro[12:35]
Thank you, David. Always a pleasure having a conversation with you. Have a wonderful afternoon.
