Who’s Winning The AI Race? + Software’s Future — With Sridhar Ramaswamy

Alex, always great to talk to you. Thank you for having me.

First of all, the AI race changes every month. We should all feel great about making predictions because one of them will come true and the world will change enough that we have to make new predictions. I think the gap between the truly great model makers of the present era — it's like OpenAI, Anthropic, and Gemini very much in that mix — and everyone else is quite staggering. And it's also a world in which no incumbent should feel comfortable about their position because things are changing so much and a great new model can sometimes end up producing a lead that's like a year long, which is an eternity in today's world. And so I would say from that perspective it's early, there's a lot of change. What is also quite profound about this moment is the things that we can get done with the models that have already been launched, where it's merely an issue of stuff like mechanics — for can you get inference capacity, it's a lot easier to solve. I think that's the part that sometimes people overlook about what is remarkable about this moment. These models, they can do amazing things. We'll get into some of the things that we at Snowflake are doing. I think it is their ability to create value, their ability to help among the most prized of professions today, software engineering. I think that's the thing that will drive so much impact. Lots more to come, but I would say it's very, very early in the AI race.

When you say web visits, you mean for things like Gemini, correct?

I mean, I think the fact that it has become — OpenAI has become the Google of choice when it comes to chat for most of us, that's actually a durable advantage. And I use it quite often for all kinds of things, including solving problems in the real world. My coffee machine not working, or I can't open my gate anymore. Like, the amount of use that you can get is pretty remarkable. I think that lead is real. On the other hand, something pretty simple — like not simple, it's hard — faster image generation or more accurate image generation, which is what Google pioneered with Nano Banana. It's actually having a profound impact on things like their usage and OpenAI was late to the game just for that one feature. You think, come on, it's a small feature. How much can it matter? It matters. People like being able to create things. It just tells you that yes, competition is actually very fierce and big companies generally have a lot of birthing issues when it comes to new things. It's just a matter of how they work. First of all, they don't often have a clear perspective of what amazing means in a new area. And what they struggle with, even if they can understand amazing, is figuring out a path to that amazing. One can argue that xAI, for example, has actually produced what is widely acknowledged to be a world-class model that is out there. But that act of sheer creation is not something that anyone should take for granted. It doesn't matter how many, how much resources you have. It's not that easy to figure out all the little things that you have to get right in order to get to a point like that. You see other companies with tons of money struggling to be at the same caliber as OpenAI and Anthropic. Google now has had a set of pretty deep advantages in this area. They kept DeepMind quite separate and DeepMind has always been at the cutting edge of AI and it's become a real weapon for them in terms of getting to the front. And once they get there, all of the other advantages that they have — of distribution, the bottomless well of money that they can borrow from investments in things like TPUs, which kind of looked crazy back then that we would invest in it — all of those become accelerants. But I think what one should take away is that that breakthrough, which is so hard to achieve, especially for big companies with specialties, Google has managed to achieve. This just means that OpenAI and Anthropic need to understand that any kind of lead that they get is not going to be a long-lived one and they really have to work hard and compete. Honestly, I think that's a good thing for all of us. Just to give you some points of comparison, GPT-4 by all accounts was ready in August 2022. Long time ago. And it took Anthropic, I would say roughly two years, summer of 2024, to have a model that was of comparable quality to GPT-4, like two whole years, which is an eternity. And then soon after, Anthropic launched a coding model that was widely acknowledged to be the state of the art and they have stayed there. It took OpenAI and Google again a year plus to catch up to that. It tells you that leads are shrinking and there's going to be more and more competition. And of course there's a pressure from things like the open-source models to just turn this into a whole other ballgame in terms of what is possible with them.

Google always had — and the founders definitely — they were always well calibrated for crisis. I remember back in 2005 when live.com, the precursor to Bing, first came out with what appeared to be a really good search engine. We got into what's called a code yellow. It's like meet every day, all hands on deck, drop everything else. We got to be faster, better than them. What was it called? It was called live.com. But the — it was just a — it was called a code yellow. It's basically get the teams together, show up in front of Larry, tell them what you're doing today. And then they went to code red with this OpenAI thing at a certain point. Yeah. Yeah. But the point is, and every year that I have been at Google, I can think of one or more crises that required us to operate very differently. And what looks like a placid company from outside is very motivated, very driven. They've also struggled with structural boundaries. For example, the thing that we did for a social network, which was called — I forget — remember, Google+ — that was sort of a disaster because, you know, it's first of all, it's hard for a new player to break through, especially with something like a network effect of a social network. It's just really, really hard to do. And so they struggle with new things that they do, but they've also demonstrated an ability to adapt. Google Cloud, by you know, Google Cloud is a pretty big success. Obviously a lot of credit goes to Thomas for making that happen. It is an adaptable company. It is a malleable company. So I'm not surprised. And you know, I'm not that close to Google anymore, but folks speak about how one of the really cool things about DeepMind is having Sergey in the mini kitchen just hanging out talking to people. And so that sense of time, that sense of what is a pivotal moment, that's what great leaders bring and Google's always had that in spades. I remember when Google Plus launched, I actually was supposed to go meet a friend at Facebook that weekend and they were supposed to have their barbecue, their company barbecue, and they canceled it. And I was like, what happened? And he's like, don't you realize we're at war.

One issue we should all keep in mind is that when you're seizing lots of ground, when times are early, if you're successful, people will call you a genius. If on the other hand things don't go well and a threat shows up in the main thing that you do, people will say lack of focus. For the longest time Google was criticized for being a one-trick pony in search and after a while it was criticized for having too many efforts that lacked focus and now we are back to putting Google as a hero because they succeeded in Gemini. So we should all remember that judgments are post-fact and dependent on the outcomes produced rather than the actual strategy. There's a little bit of that. Having said that, OpenAI has a lot to offer enterprises and we are excited to partner with them because many customers are joint customers of Snowflake and of OpenAI. We've created an agentic platform called Snowflake Intelligence that's been quite transformative. Over 2,000 customers, fastest growing product, over 2,000 customers are using it pretty much, you know, three months after we released the product to GA. Enterprise customers are fussy about using products only in GA and it's among our fastest growing products ever launched and it's focused on data and Snowflake. Back to your point about focus, we wanted to make sure that we created a product that could enhance the value of things that people had already done with Snowflake. We didn't want to go and pitch our enterprise customers and say, hey, we're doing something dramatically new, you know, work on it with us. We said you can get value from your data a whole lot faster. Not only that, we also said we live what we preach. And so I often show them things like our sales agent, which puts every piece of information that my sales team has about every customer at my fingertips. What meetings did this customer have yesterday? What are the outstanding use cases? All of that is available to me. But it's also programmable. I can get the information the way I want, share it the way I want. But there's a lot more in this world of agents and enterprise. How do you help people take action? How do you help people be better grounded about the consequences of their action? How do you help them analyze situations? These are the things that we are excited to be collaborating with OpenAI on. Yes, one part of it is us using their models, but I think the much more interesting thing is going to be what are areas that are very amenable to AI creating value and how do we make sure that we make it easy for enterprises to realize that value. To make this super concrete, I was visiting a big manufacturer yesterday. My eyes kind of popped out when they said, you know, listen, we have 5 million SKUs. 5 million SKUs that they sell. And part of their issue is we have trouble pricing this because it's a big dynamic marketplace. We don't know what competitors are pricing it at. We don't know what kind of — like you have to take into account the margin that we have on the product, the NPS for the product. Can you create an agentic system that can help us do pricing better? We have all our data on Snowflake. And that is a situation in which the power of agentic technology, the ability to look at a complex situation, break it down, follow best practices for how work should be done, is going to be a big multiplier for how they get their work done. There's potentially hundreds of millions of dollars of additional revenue that this company can make if they can do a better job just with this one single project. That gives you an example of the kind of things that people are looking to do together with OpenAI and Anthropic and a data platform like Snowflake.

Well, this is a great question and it goes to a topic that I'm pretty passionate about. I call it, what does work look like in the future? And today our work is pretty much we go look at our email, we go look at our to-do list and then decide what are the things that we should be doing. Or you know, if you're like me, you have meetings on calendar where work shows up. The future that we envision very much is you describe what you want systems to do. Hey, these are the kinds of things that I should be looking at every day. For example, I look at our revenue alerts every day. I go and look at the dashboard. If there is a big up or a big down, I send out questions and so on. Very automatable. And so you have an agentic system that is connected both to the past information that's typically sitting in Snowflake or what was performance-like. It also has access to things like prediction models that say if something changes, what does the future look like? Also things like ambient information, your emails, your documents, or even things like the stock market, ambient information about the world. And your work very much becomes these are the five topics that you should be paying attention to and here is a brief for these five topics and potentially even recommendations.

Say, I want you to look at the spread between how I price, how the market is pricing, identify the top 10 opportunities I should be paying attention to in my department today, generate a report for me. My job is, okay, I'm going to go through this, go through the recommendation, and figure out what do I change, and if I want to make a change, what approvals do I need to get within the company?

And the magical thing about this, by the way, we are living this with our support team. We have changed our support team from 50 people writing software, 300 people using this software to help debug support cases, to much more of a builder-user model where there are a set of tools available within our coding agent, Cortex Code. And whenever a support case comes, they use these tools to analyze what is happening. And then they tell the customer what to do. And sometimes they decide, you know, these tools are not enough. I need to build a new tool. And they add that tool itself to the suite of tools that everyone else can use. So this is work self-correcting, getting itself better over time. And the goal is just things get done a whole lot faster. Already we are seeing 10x — not 10%, 10x — reductions in the amount of time that it takes to debug complex cases that come in.

You got to walk the walk. We were in Davos together. Yes. And you know, two weeks ago, and I probably met 20-odd CEOs, CIOs, lots of partners. And my sort of SOP, standard operating procedure for each of these meetings would be I would ask our sales agent for information about the customer. What's the state of our relationship with — take your pick — and it generates a report. I would turn it on and show my phone to them. And they would go, holy cow. But uniformly, not one of these CEOs has the same tools that I do. I see. That's the difference between actually getting the work done, making AI serve meaningful needs, and yes, the hype that you're describing. All of the people that are in the camp that you're describing have never had useful products built for them that deliver meaningful value. I speak as somebody that lives this. The amount of feedback that my poor team gets about how difficult the mobile experience is, how to make it better. We just launched like Face ID authentication. That's a big deal because I don't have to log in all the time. It's taking care of all of those kinds of nuances, making enterprise data come alive, available for you, and then helping you with decisioning. That's the magic and that's why you are hearing people say it's hype. But it's companies like Snowflake that are actually living what we are preaching. And I give that same feedback to my exec team, which is, hey, all of you need to be demanding tools that are as good as the one that we have for the sales agent and our team should be providing them to you and you should be using them day-to-day in how you can work better. I agree that there is work to be done but the sheer potential of something like this is magical. I'll give you one more small example of something that is cooking this very week. I'm working with our ops team, our operations team that helps manage Snowflake, the software running in the cloud, about how to get on a more agentic bandwagon. Like, you know, super grizzled infrastructure engineers that are like, what is this? You know, we know better. But we're walking through this journey of, no, no, let's create tools that our coding agent can use and you will genuinely find that is a lot easier. And so someone created a tool that will help detect things like, oh, are there problems with warehouses resuming. Warehouse is a basic unit of work that gets stuff done for our customers. And when our customer says start this, we want it to start quickly. In like 10 seconds, I had generated a histogram of resume times, put a nice graph, and I sent it to the team with one prompt, all English, on top of a tool that somebody had built to look at resume times in warehouses. And the team is like, holy cow, that's the magic of agentic platforms. But yes, you have to do the legwork to put them into place, put the guardrails, things like that. But there's real magic here.

Well, that's also where magic can happen. Right. And you know, I told you that we released a new product called Cortex Code, which is our data coding agent. We launched it GA yesterday and it dramatically lowers the amount of time that it takes to get stuff done on Snowflake. We all get carried away with how does AI make it easier for a business user like me to get access to my data. That's great. But on the other hand, everything from how do you set up a database to how do you move data from a production-like transaction database over to Snowflake for analysis. How do you build a machine learning model? How do you build an agent that you can then give to the business user? Cortex Code is meant to address all of that again in natural language. And part of what we have built there are what we call a series of skills that help automate this work. And this is a theme that's going to come up again and again, which is how do you use AI to make launching AI products go faster, right? That's the feedback loop that one needs to be on. It's a little bit of a red pill moment where you're like, wait, you mean I can release new software products pretty much every day? Because releasing a new piece of functionality is as simple as writing a recipe in English, which all of us are very capable of doing. I think using AI to make AI go a lot faster is something that we are excited about and this product is among the best in terms of how do you get it from Snowflake.

Yeah, I think there is going to be a concentration towards platform players. But I would also be cautious about general pronouncements for the simple reason that we are all actors in this space. We all get to change the outcome. I feel very good about Snowflake as a data platform. But I honestly do not want to be in a situation where access to Snowflake is always mediated through someone else. That's always a very dangerous place to be, especially in a moment like this. This is the reason that we develop not only Snowflake Intelligence, which is the best way for a business user to get access to their business information that is trustable through the devices that they want, like their phones, rather than trudge through dashboards. But we are also investing massively in how do you make creating data products, how do you make creating applications a whole lot easier. Absolutely. It's going to be the case that there's a lot of functionality that sits in complex applications. We're actively working with all of those folks, whether it's a ServiceNow or a Salesforce or SAP, with whom we have a big partnership, in creating this agentic future together. Agent future is very much going to be what I said: past, present, future, and actions. And so we think we stand a very, very good chance of being the platform where this work happens. But as I said, it's a foot race and it's all about creating value really fast for your customers. And I would shy away from X is going to win or Y is going to win. The companies that are going to win are the ones that have great capabilities but also take the time to figure out how to create value for their customers.

It's a very real risk. I think people that were confident about their position in the world because they were essentially walled gardens for data and functionality, and are slow at providing modern ways of dealing with information, are going to struggle in this world. This is the reason that I stress us living by what we speak in terms of AI and agentic platforms and this future of work concept, precisely because unless you live it, you don't actually feel it. And unless you live it and feel it, you're not going to help your customers get there. I think niche SaaS software providers that basically benefited from lock-in — think about it. If you used a piece of SaaS software, logged into it on your browser, God help you if you want your data back. Just like, not going to happen.

What this current moment is pointing out is that that's a very dangerous place to be. And a lot of these players risk becoming dumb backends to the models, which is why Snowflake is so leaning forward on agentic AI and living by what we speak, because that's the place where value is going to get created.

There are a few things that we should take into consideration here. As you know, companies are valued not on what they're doing today, but on what they're going to do in the future. And I would actually distinguish data platforms like Snowflake from pure software providers operating on a subscription model. Not that it's a bad model, but the way they have operated is AI became another SKU for these folks, and customers have had to sign up for AI products regardless of whether they created value or not. That's sort of become the favored way of becoming AI-native. I think what the current moment points to is a real risk that that is not a winning AI strategy, meaning that work is not going to get done by interacting with a chatbot on a particular SaaS app that you used. Which is why our vision of agents operating on a data platform that has much of the analytic insights about the past, as a lot of our customers do, but with the ability to bring in integrations via MCP, via other APIs for how do you talk to other systems — I think that's the compelling vision. I think companies are going to win if they have both a convincing vision for how work gets done in the future, but are able to back it up with, 'and here is how we help you, the customer, get it done fast.' The model makers approach it from this view of the model is everything and nothing else matters. We approach it from the viewpoint of it's the entirety of the experience. It's the model — that's why we partner with all of these folks. It's the most critical data that's valuable to your company, but it's also integrations with the operational systems that really help get work done. I think that's the compelling vision for how work gets done. What the markets are in some ways pricing is the fact that AI as a bolt-on to SaaS software does not feel like a winning strategy. You know, I feel much better about the path that we are pitching. Also, our products are consumption-based, meaning that if something doesn't get used as much, there's not a penalty to just building them and using them as much as you want.

That's very much a fear that we need to operate with. That's very much the opportunity of the moment.

Through, yes. I — and you know, everything else, the world is just a dumb data pipe that feeds into that big brain. That's the vision that they would like to see come true. And the vision that I would like to see come true is, hey, we host the most important data for every company and the most important predictive models for every company, and I can create agents that can deliver substantial value. But by the way, we also follow, like others do, an interoperability strategy, because if a customer comes and says, 'I want to build a data product on Snowflake, fine, it can have an AI interface, but I really want it to be accessible somewhere else,' I don't get to say no to that. The only people that win are the ones that effectively deliver what customers want.

We stick to customer choice. What do customers want, right? If they want to access data through a Snowflake Intelligence agent, the OpenAI team doesn't say no. If on the other hand our customers want to expose, you know, data like important enterprise data that they have as an MCP endpoint into ChatGPT, we don't get to say no.

It's all about creating products and value. It's not about any one. No one has an insurmountable moat. Like OpenAI doesn't get to say the only way you get 5.2 is to come to ChatGPT, right? I don't get to say the only way you get to access data on Snowflake is to come to Snowflake Intelligence. It's pretty much may the best player win. And so it's very much about creating value.

Maybe not. It's very, very early. And remember, we are still living in a world — I don't know how many tabs you have open, Alex. Mine is 200.

That's pretty good.