Google Cloud CEO Thomas Kurian on AI Competition, Agents, And Tariffs

Thank you for having me.

You know, AI has definitely driven adoption of different parts of our platform. And so people typically when they come in for AI, depending on the type of company, they come in at different parts of our portfolio. Some of them say, I really want to do super scaled training or inference of my own model. And so there's a whole range of people doing that all the way from foundation model companies whether that's Anthropic or Midjourney or others and also traditional companies Ford motor company for example when they brought their they wanted to use our chips and our system called TPU tensor processing unit to model air flow and wind tunnel simulation using computers rather than physical wind tunnels. So they're doing that as an example. So one set comes and says I'll use your AI infrastructure. A second set comes in and says I want to use your AI models and that could be somebody building an advertising campaign using our image processing model, somebody wanting to write code using Gemini, somebody wanting to build an application using Gemini or one of our newer models like Veo which is our video processing model. So in that case they come in and use the platform but along with that they may say I want to put my data so that the model can access it quickly and they start with one of our database offerings for example. So it certainly draws more pieces of our portfolio as part of it. And then the third is people coming in and saying I want to use a packaged agent that you have. For example we offer something for customer service. We offer something for food ordering. We offer something to help you in your vehicle like in car. We offer stuff for cyber security. So there's a whole portfolio of these and so depending on which customers coming in, they come in at different layers of our stack.

It's a good question. It depends on the country. It depends on the industry. It depends on the segment. Let me explain what I mean. If you're an AI unicorn, meaning you're funded to build a foundation model or you're building an application based on AI, that's really the central part of your decision. If you are in an industry that for example in retail where we have a product called retail search and conversational shopping where you can take Google like search using text, images, video and put it on your catalog and you can also put conversational shopping where I can ask a question. I'd like to return this dress and have the system handle that transaction for you. It's a super important thing. For example, for people in commerce, whether that's retail or telecommunication. On the other hand, if you look at a utility or an industrial manufacturer, it applies to part of their organization, but it may not be the central thing. And so, it really depends by industry and by customer segment. And so, but we part of our value proposition is that we offer all of these different capabilities. And so, AI is helping us. It's not the sole reason for our growth.

We've been very measured in how we brought our AI message to the market to avoid people feel like we're overhyping things. And we've always said we're going to build the best technology in the market. Right now we're super proud. We have over 2 million developers building every day, every morning, every night using our AI platform. And you can see the strength of our models. You know, Gemini Pro 2.5 is the world's leading model. Gemini Flash is the most price performant model. Imagen and Veo are considered state-of-the-art for media processing. And we've got tons of new stuff that we're introducing at our event next week from audio, speech, etc. So, we've been very very thoughtful about how we've introduced stuff and I'm not a marketer. So, I will tell you it's an important factor. It will be an increasingly important factor and our strength in it helps bring other products along with it.

I would say we offer 200 models in our platform. In fact, we look every quarter at what's driving popularity in the developer community and we offer them. We offer a variety of third-party models and partners, not just Anthropic, AI21 Labs, Allen Institute, there's a variety of models there. We offer all the popular open-source models. Llama, Mistral, Deep Seek, a variety of them and we base it on what customers want. So we track what's on the leaderboards and what's getting developer adoption and put them in the platform and people have been super pleased that we have an open platform. Companies always feel companies want to choose the best model for their needs and there's a range of them we're offering a platform you can choose the model you want. The only model we don't offer today is OpenAI and that's not because we don't want to offer their model it's because you welcome them on the platform.

Of course, we would.

I don't want to tell you that we won't do it. We have always said we're open to doing it. I think it's their decision.

Well, our field is not compensated any differently. Our partner ecosystem is able to use all the models in the platform and most importantly, we have very large Anthropic customers running on GCP. So if you don't have your own model or you have a model of your own but it's terrible naturally you're going to say something like that.

They've done a good job. No question. I would say OpenAI has done a good job with that. Whether that's how much of credit goes to Microsoft outside of providing them a bunch of GPUs, time will tell.

We work extraordinarily closely with Demis and his team. When I say extraordinary closely our people sit in the same buildings we work extraordinarily closely. My team builds the infrastructure on which the models train and inference. We get models from Demis and team every day. In fact, we're staging models out to the developer ecosystem within a matter of a few hours after they have finally built. And then we take also feedback from users and move it upstream into pre-training to optimize the models. And one benefit we have at Google is all our services whether that's search or us or YouTube this inferencing of the same stack and same model series. So the model learns very quickly from all that reinforcement learning feedback and gets better and better. So there's a lot of close collaboration. Many times if I can be frank when we enter a new domain like I'll give you an example we built a solution for cyber intelligence using Gemini so there's a lot of threats happening in the world you want to collect all that threat feed we do that using a team we have called Mandiant and also from other intelligence signals we're getting on what are the threats emerging you then want to compare it to your environment to see if you've been, you know, you're at risk. And most importantly, you want to compare it to what parts of my configuration will somebody use to try and get in. And so we used our Gemini system to help prioritize and also help people hunt faster. We call it threat hunting faster. Now in that environment, the model has to learn how to find patterns in a large number of log files that people are ingesting and that required specific tuning of the model to do that. And so there are things there that having a close working relationship with the Deep Mind team has helped enormously. Similar things when you look at for example customer engagement, customer service. We've got a project on at Wendy's to automate food ordering in the drive-thru. You know, if you actually think of a drive-thru, it's an extraordinarily complicated scenario because there's a lot of background noise. Kids screaming in a car. People change their mind when they're ordering something. I didn't mean that one. I wanted that one change to this one. And which one did you mean by that one?

So, there's a lot of things that we needed the model to do to have ultra low latency in being able to have that conversational interaction with the user. So, all those elements, the partnership we have with Demis has been super super, you know, productive and it's also, you know, most importantly, it's people working together. We're all close personal relationships that helps us get through a lot of design changes and other things and we're all rowing towards the same goal.

I you know, I don't want to comment on what he said. I can just tell you there's a lot of debate on cost of training and inference. First and foremost, in the long run, if AI really scales, the cost you really want to care about is inference cost because that's what's integrated into serving. And any company that wants to recover the cost of training has to have a large scale inference footprint. There are lots of things we've done with our Gemini Flash, Gemini Pro models that you can see and also other people using TPU for inferencing. For example, large companies are using it to allow them to optimize the cost of inference. Cost of inference can be on the efficiency with which you handle your serving fleet, how you go disaggregated serving, what you do with caching and key value stores. There's a hundred different variants of that. The proof I think is in our numbers. You know, if you look at our price performance, meaning quality performance of models and the unit price of tokens, we're extraordinarily competitive. So that's number one. Number two on the training, I think there's a bit of confusion that's you know may exist in the market. There is so there is research frontier research exploration. Frontier research exploration for example could be how do I think about teaching a model a skill like mathematics? How do I teach a model for example a new skill like planning? How do I teach a model a new skill in a brand new area? So those are what we call you know frontier research that goes on and many many experiments like that are done and then after you find the recipe you then actually train a model and train a model is actually you do the model run where you're running the actual training. I think people are mixing up the total amount of money spent on research and breakthroughs as opposed to actual training. And we are very confident we wouldn't be investing in the way we are as a company without knowing the ratios between all of these. And so we're very confident that we know how to run very efficient model training, what we're investing in frontier research and then most importantly how we're handling model inferencing and being world class at all three.

There are gains to be had. I don't think they will be at the same ratio as earlier because just you know there's always a lot of diminishing returns at some point. I don't think we are at the point where there are no more gains but I think we won't see the same ratio of gains we used to see. With inference. So that will be the new cost to basically taking the models and putting them into production and using them.

It's a really good question. I mean reasoning is something we are starting to see customers using in different parts of our enterprise customer base. For example, in financial services, we've had people say, "Hey, I want to understand what's happening in financial markets." Summarize the information coming off whether that's video feeds like CNBC, financial market indexes and other financial information and tell me what's happening and the model can not only build a plan for how it collects the information but summarize it and then reason on the summary to say are there you know conclusions to be derived right so we are starting to see people starting to do that. How much of that will be versus other scenarios time will tell but we are starting to see people doing much more sophisticated complicated reasoning even in areas we have a travel company for example that's working on give me a very high level description of what you want to travel for I want to fly to New York I'm taking you know my son we'd like to see Coney Island and the following three things build me a plan and in that it and have multiple choices. But it may say, you know, if you're traveling in June, may be hot in the afternoon. Therefore, I think we should have you see Coney in the morning and go to the museum in the afternoon. And models are starting to be able to reason on those things. And we are starting to see early adopter companies test in all these different dimensions.

There are feeds when I mentioned CNBC I'm using an example there's they have personal feeds from their broker and dealer networks which are private of their own that they're feeding into this because when they have a broker or an equity analyst make a broadcast to their internal you know teams they want to feed that as an example that I was using that just as an example to say what kind of a feed given your audience to explain what a video feed would look like right.