Google Cloud CEO Thomas Kurian on AI Competition, Agents, and Tariffs
“We've been working on contingency plans for quite a while,” Kurian says on Tariffs.
For years, Google’s cloud services offering ran a distant third behind Microsoft and Amazon. Then came generative AI.
While Google Cloud CEO Thomas Kurian doesn’t give the recent AI revolution all the credit for his division turning in 30% quarterly revenue growth and closing the gap, it’s certainly an important factor. And in an interview on Big Technology Podcast on Wednesday — as Google hosts its annual Google Cloud Next event in Las Vegas — Kurian made the case for Google’s offering.
Google, Kurian said, has DeepMind in house, still offers more than 200 other models — including DeepSeek, Llama and Mistral — and is embracing open source.
“We base it on what customers want,” he told me. “We track what's on the leaderboards, what's getting developer adoption, and put them in the platform.”
In a candid conversation touching on Google’s competition, how companies are putting generative AI into practice, and, yes, how tariffs might impact his business, Kurian shared how the current moment is shaking up the competitive balance in cloud, and what might come next.
You can read the Q&A below, edited for length and clarity, and listen to the full episode on Apple Podcasts, Spotify, or your podcast app of choice.
Alex Kantrowitz: Google Cloud Platform has been surging, with growth rates of 30% per quarter recently. Is AI responsible for that?
Thomas Kurian: AI has definitely driven adoption of different parts of our platform. When people come in for AI, some of them say: I really want to do super-scaled training or inference of my own model. There's a whole range of people doing that, all the way from foundation model companies, whether that's Anthropic or Midjourney or others.
Also, traditional companies like Ford, for example, wanted to use our chips and our system called TPU, Tensor Processing Unit, to model airflow and wind tunnel simulation using computers rather than physical wind tunnels.
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. In that case, they come in and use the platform.
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. We offer our services for cybersecurity. Depending on which customers are coming in, they come in at different layers of our stack.
You're the CEO of Google Cloud Platform. So when it comes to the broad Google Cloud Platform ability to compete, how important is AI across everything? Yes, of course, it varies for individual use cases, but broadly.
It’s going to be important going forward. We've been very measured in how we brought our AI message to the market to avoid people feeling like we're overhyping things. We've always said: We're going to build the best technology in the market. Right now, we're super proud. We have over two million developers building every day, every morning, and every night using our AI platform.
You can see the strength of our models. Gemini Pro 2.5. is the world's leading model. Gemini Flash has the best price performance. Imagen and Veo are considered state of the art for media processing.
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.
You talked about a lot of models coming out of DeepMind. Here's what Amazon Web Services might say about that: Google has its own models and it wants you to use them. Amazon, however, will let you pick whichever model you want, from Anthropic on down. What would you say to that?
We offer 200 models on 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, DeepSeek—a variety of them. We base it on what customers want. We track what's on the leaderboards, what's getting developer adoption and put them in the platform. People have been super pleased that we have an open platform. We 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…
Would you welcome them on the platform?
Of course we would.
Any talks about that?
I don't want to tell you that we won't do it. We have always said we're open to doing it. It's their decision.
Okay, but your competitors like Amazon might say something like even though Google can offer everything, they might still push you to use DeepMind models. What do you think about that?
Our field is not compensated any differently. Our partner ecosystem is able to use all the models in the platform. Most importantly, we have very large Anthropic customers running on GCP. 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.
Are you saying that Amazon’s model is terrible?
No
Okay. Why don't we move to Microsoft then? Microsoft might tell that they have this partnership with OpenAI, which is going to build the best in breed technology. What do you think about that?
They've done a good job, no question. OpenAI has done a good job. How much of credit goes to Microsoft outside of providing them a bunch of GPUs, time will tell.
There is a pretty interesting difference between Google and Microsoft, and that is that Google does have DeepMind in-house. What does DeepMind being in-house provide you?
We work extraordinarily closely with DeepMind CEO Demis Hassabis and his team. When I say extraordinarily closely, our people sit in the same buildings. My team builds the infrastructure on which the models train and inference. We get models from Demis and his team every day. In fact, we're staging models out to the developer ecosystem within a matter of a few hours after they're finally built.
Then we take feedback from users and move it upstream into pre-training to optimize the models. One benefit we have at Google is all our services, whether that's search or us or YouTube, we are inferencing off the same stack and same model series. The model learns very quickly from all that reinforcement learning feedback and gets better and better. There's a lot of close collaboration.
Many times when we enter a new domain, for example: We built a solution for cyber intelligence using Gemini. 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're at risk. Most importantly, you want to compare it to what parts of my configuration somebody uses to try and get in. We used our Gemini system to help prioritize and also help people hunt faster. We call it threat hunting faster.
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 requires specific tuning of the model to do that. There are things there that having a close working relationship with the DeepMind team has helped enormously.
There’s similar things when you look at, for example, customer engagement, customer service. We've got a project at Wendy's to automate food ordering in the drive-thru. 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 changing their mind when they're ordering something. “I didn't mean that one. I wanted that one, changed to this one, and which one did you mean by that one and this one?”
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. All those elements and the partnership we have with Demis has been super, super productive.
I was speaking with Mustafa Suleyman, the CEO of Microsoft AI, just a few days ago. He said that without spending the billions of dollars it takes to train the new models, you basically replicate what they're doing with a lot less money and put it into action just a little bit more slowly. What do think about that argument?
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 scales, the cost you really want to care about is inference cost, because that's what's integrated into serving. 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 and 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. The proof is in our numbers. If you look at our price performance, meaning quality performance of models and the unit price of tokens, we're extraordinarily competitive. That's number one.
Number two, on the training, there's a bit of confusion that's may exist in the market. There is frontier research exploration. Frontier research exploration, for example, could be: how do I teach a model a skill like mathematics? How do I teach a model a new skill like planning? How do I teach a model a new skill in a brand new area? Those are what we call frontier research that goes on.
Many experiments like that are done. And then, after you find the recipe, you then actually train a model. Training a model means you're running the actual training. People are mixing up the total amount of money spent on research and breakthroughs as opposed to actual training. We are very confident we won't be investing in the way we are as a company without knowing the ratios between all of these. We're very confident that we know how to run very efficient model training, what we're investing in frontier research and, most importantly, how we're handling model inferencing and being world-class at all three.
Do think there are still gains to be had by scaling up the pre-training of models?
There are gains to be had. I don't think they will be at the same ratio as earlier because there's always a law of diminishing returns at some point. I don't think we are at the point where there are no more gains but we won't see the same ratio of gains we used to see.
With inference, how much of the cost is going toward reasoning and what have these new reasoning capabilities allowed your customers to do that they couldn't do previously?
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: I want to understand what's happening in financial markets, summarize the information coming off financial market indices 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 if there are conclusions to be derived.
We are starting to see people doing much more sophisticated, complicated reasoning. We have a travel company, for example, that's working on giving me a very high-level description of what you want to travel for. I want to fly to New York. I'm taking my son. We'd like to see Coney Island and the following three things; build me a plan and in that, it can have multiple choices, but it may say, if you're traveling in June, it maybe hot in the afternoon. Therefore, you should see Coney Island in the morning and go to the museum in the afternoon. Models are starting to be able to reason on those things. We are starting to see early adopter companies test in all these different dimensions.
Yes, but in the non-reasoning versions of large language models, I could say, build me a plan and it could do that. So what does reasoning do that allows customers to be able to do stuff they could not previously?
Historically, when LLMs were used, people were worried about hallucination. They gave a large language model a single-step task, meaning, do this and come back to me so that I can determine if your answer is hallucinatory or not.
Secondly, when I asked you a question, you gave me a single answer. You didn't generate a variety of different options and then reason on it or critique them to say this might be the best answer.
That is the nature of some of the differences we see in why people are using reasoning now as opposed to prior. The more you can trust that the model can actually reason across a set.
Whenever you have a multi-step chain of thought, if you have drift, meaning early in that chain of thought, you have an incorrect answer and then it stepped on that incorrect path and reasoned a lot more. Downstream, you can get way off relative to what the right path ought to be. As models have become more sophisticated, people have trusted them. Part of it is the accuracy can be higher. Part of it is that it can evaluate a set of different choices and give you an answer based on a set of choices, not just say: here's a single answer.
Third, we also allow people to understand what the steps were in how it reasoned. They can look at it and say: yeah, maybe I agree with it, maybe I don't.
Jensen at NVIDIA says reasoning costs 100 times more than non-reasoning inference. You also have your own compute, you're also facilitating that. Is that in the ballpark?