Snowflake's CEO Tells Customers to Spend AI a Thousand Dollars at a Time

Sarah Guo welcomes Sridhar Ramaswamy back to No Priors eighteen months into the Snowflake CEO job. He had been SVP of Google Ads and founder of the subscription search engine Neeva, which Snowflake acquired in 2023, and he took the helm from Frank Slootman in early 2024. The original product, conceived more than ten years earlier by Benoit Dageville and Thierry Cruanes, was years ahead of its time and produced, at the IPO, the biggest software listing of that moment. What happened next, Ramaswamy says, is that the company was a little slow to react to machine learning and AI.

That lag was honestly part of why Slootman voluntarily pushed for a change. He felt, presciently, that Snowflake was heading into a more tumultuous time from a product perspective and wanted someone product-first in charge. The last eighteen months have been about embracing that wave. AI became everyday and commonplace; the speed of change kept rising. Snowflake, in Ramaswamy's telling, not only shifted the product, which one might have expected, but also retooled marketing and go-to-market. Last year there were a lot of doubters. He took the job because he talked to a whole lot of customers first, and they all loved Snowflake. That was the motivation. He thinks they have ridden through the doubt and are now at the cutting edge of data and AI for enterprises.

The first six months were mostly tactical, and the core word was accountability. Snowflake had gone through a rocket-ship phase, growing more than 100 percent year on year, and had specialized at every layer. The distance between the engineer who shipped a feature and the customer who used it ran through seven to ten teams. That works fine, he says, when you have perfect product-market fit and you are optimizing every function, when you are simply the winning cloud data warehouse. You can drive a truck through that. In AI, where you can barely tell what comes out next month, forget next year, it is the wrong structure.

They reorganized by product area, with accountable owners in product and engineering for AI and for the core warehousing and analytics product. They also wanted a straight line to go-to-market, so they created specialized teams, later described as war rooms or pods, that sit with product, engineering, and marketing to take new products to market. The early emphasis was faster iteration. Speed wins, Ramaswamy has believed all his life. The ability to iterate always trumps a carefully laid-out strategy. You should not do dumb things, but realizing any kind of gain requires a lot of iteration. They changed how quickly they created products and how quickly they iterated with customers.

They also had to find their spot in AI. They are not a cloud service provider and not a foundation lab, so what are they? That search produced a rename: the AI data cloud, as opposed to the data cloud. Early last year they actually went down the path of creating foundation models and trained a credible mixture-of-experts model. They also quickly realized they could not compete with OpenAI or Anthropic. They simply did not have the capital to invest meaningfully. So they pivoted to a humbler question: how does AI massively accelerate what can be done with data already in Snowflake, and over time become a reason to bring more data in. About half of the qualifying Fortune 2000 companies on the planet are Snowflake customers. They have their most valuable data on Snowflake. What does AI mean for that?

The AI product strategy, he says, did not claim they would rethink everything. They systematically invested in components, search and text-to-SQL, that added value to what people were already doing. Snowflake Intelligence is an agentic platform, but an opinionated one. Cloud providers' platforms tend to say you can bring data from anywhere, imagine any workflow, and one agent will rule them all. Nice in theory; in practice an infinity of things makes it hard to know what you should actually do. Snowflake Intelligence is focused on creating value from structured or unstructured data a whole lot faster. Internally, the use case that got them excited was taking all the dashboards sales used and putting them in one interface. Two or three versions later that became Raven, the sales data assistant. Early external customers included Cisco, Fanatics, and the USA bobsled team.

The theme is getting away from the inflexibility of dashboards. A dashboard is a 2D view of a complex surface. It has no easy answers to the many questions any reasonable person will have off it. They wanted something that freed people from that 2D style of thinking, was more flexible about access, and also knew its place. This is not a general-purpose agent to do it all. It is a foundation for getting value from data quickly. How does a user consume it? An interactive interface: ask questions, pick from canned questions so you are not blocked, or ask what data sets it can see and what kinds of questions it can answer. The aspiration is every employee in the company, not people who can write SQL. Ramaswamy does not go to a customer meeting without quickly checking the latest. Raven covers the relationship, the contract, consumption, recent conversations, and outstanding issues. There is breadth and value for many people. They do not pretend it is a BI dashboard. Tableau or Sigma can do more. That is not the goal. The product also lets you do things a dashboard cannot, and it is meant for any business user.

Trust sits at the center of the AI products. He tells people to think of AI the same way they think about software engineering: there is a right and a wrong. It cannot be YOLO AI, some good answers, some terrible answers, your problem. Every new thing that launches needs an eval. If you change the underlying model, you need to verify quickly that you did not blow up what you were already doing. Making it a trustworthy product for every employee is actually new for Snowflake. For pretty much all of its history the data team used it, slapped a dashboard on top, and exposed that to end users. This is a different motion, which is why they are working on identity-provider integration so companies do not have to stand up a Snowflake account for every user. There is also subscription fatigue, so Snowflake Intelligence is a consumption product: people pay for what they use, with experiments around driving adoption without runaway costs.

Guo asks how he draws the line between data, agent system, and app, given that Raven sounds like the applications she gets pitched. He tends to be emotionless about where the strongest current is. It is absurd, he says, if they think they are SAP or Salesforce. They are not. Somebody managing a hundred-billion-dollar supply-chain ecosystem is not going to say they will use Snowflake Intelligence and drop the incumbent. That is not the goal. The line between an agentic system and pure software will be blurry. Easy cases: the sales team is forced to update Salesforce on every use-case transition; that can be done with APIs. Filing vacation on top of Workday through an HR agent is a reasonable thing. Be opportunistic, but operate from value and strength, not naked ambition. Focus on what value creation means and what users really want. That tends to be more durable.

The organizational changes feel entirely too slow to Ramaswamy, which Guo notes matches the impatience she has always seen in him, even if at Snowflake's scale they look fast. Change is hard, he says. Driving behavioral change from lots of people is incredibly difficult. They were measured. Leadership alignment and clearer accountability came first, within a few quarters, because you are not dealing with as many people. The war-room or pod model of product, engineering, and go-to-market working together was an early change too, done with small groups without disrupting lots of people. Rolling out coding agents to engineers was a different kind of project. Not everybody wants to do it. Some people are skeptical.

He is a fan of combining bottom-up with top-down. The example is Benoit Dageville, the founder, who fell in love with coding agents and did more to drive adoption with engineers than any number of words from the CEO. You have to find the right people, the champions. Every large organization has forward-thinking, curious people who will work over the weekend to figure something out. Find them, encourage them, elevate them, and use that to drive change. Top-down can help, but it really needs to come from the bottom. They have now rolled coding agents out to all of the solutions engineers, who are excited because demo creation got dramatically faster. Demos used to be canned and not always customizable. Now they can build synthetic data sets for the kind of conversation Guo and Elad Gil are about to have, and run the demo live.

Guo asks whether being an investor and then an entrepreneur changed how he leads at scale. He thinks those roles are accretive in ways you do not always appreciate. He still complains to his family about the ten years of a PhD, then admits it was not a waste: it teaches you to focus on ideas and to convey them crisply. You spend enormous amounts of time writing four-line abstracts, and that turns out to be incredibly powerful. Neeva was among the hardest and most heartbreaking experiences of his life. Sometimes you are too early. He learned more about hustling, took success far less for granted, and learned social and marketing skills you take for granted at Google. His first launch at Google, three months of work by one person, was covered by the *New York Times*. Distribution is immediate. A startup makes you realize that is actually special. Google and Neeva together make him a lot more grateful. After Summit, Guo had told him it was like being invited to his rock concert: a line more than two blocks long waiting to get into the Javits Center in New York. Nothing ordinary about it.

Guo calls him perhaps the world's expert on game theory with the tech elephants: he has led the elephant, fought the elephant, and now built on top of it. How do you create durable value on cloud providers and, by analogy, on foundation models? Product-market fit continues to be magical, he says. It is why Snowflake exists. The three hyperscalers would love to own the data space the way they own other spaces, and yet Snowflake and Databricks are still here. That redeeming value is quite unique. Everyone should have humility about what it takes to create lightning in a bottle. Model companies, especially OpenAI, are in a phase where they literally do not think they can *not* do anything. He jokes that these are empires that have not met their oceans yet. You have to pay attention to what is likely to be in their immediate path. Coding agents are the obvious example: Anthropic and OpenAI will be laser-set on having the best one. Do not collide there.

The Google analogy is information. He spent enough time trying to get into physical things, shopping, airline purchases, hotels, and they did not really succeed, because they did not have core competence beyond the world of information. Discovering that kind of boundary for OpenAI or Anthropic will be fascinating, and some of it can be reasonably guessed. A set of prompts on top of one of these models is a problematic space. You need to add value. He thinks hard about what differentiates Snowflake from the model companies in areas they are likely to be interested in, and about keeping distance. Urgency, product, and collaboration with those labs all push toward the data platform as a durable category. This is also a time when literally no software company can feel secure about its place in the sun. That, he says, is perhaps as important as anything else he just said. You have to continue to earn it. If there is a lesson from the CSPs, it is infinite budgets and infinite patience. Unless you innovate and stay ahead, not just be ahead, you will be Intel. Guo has a founder-CEO of a public company she thought was unassailable who feels threatened for the first time in years. Ramaswamy thinks that is a pretty common software-CEO experience right now. When the technical environment is this fluid, defensibility is built, not strategized. Built every single day. You have to keep moving.

Looking out three to five years, core strength is the data-platform layer. Internally he talks about being there from inception to insight: from when data is first conceived to when somebody gets an insight that feeds back into the system. The pitch to CEOs and chief data officers is that the great companies of this century, Google and Meta, were more data-first than purely product-first. In the 1990s you built Photoshop, pressed CDs, and waited for feedback; data was a slow afterthought. Search ads were magical because behavior flowed back into the system in a few minutes, and his data teams were as large as the product teams. Snowflake wants to be that companion for all kinds of data, so customers have the expertise the Googles and Metas have. AI is a massive accelerant because CEOs now realize this is not just data modernization or more queries. It could change how operations work and what efficiency means. The value of data just got vastly elevated. Versus the CSPs, a data-first company with an emphasis on simplicity and integration, sharing inside the company and with partners, and AI that works with the governance you already put in place, is a higher level of abstraction. Raw compute and storage never get easier. Pre-AI, his thesis for joining was that a data platform that spans CSPs has the right, if you earn it, to be as large as a CSP itself. AI is an accelerant on getting value from data faster.

Partnerships have matured past a Snowflake-centric view of the world. SAP, Workday, and Salesforce data was always coming in; those companies now understand that this customer data is valuable and that the line between software, services, and data is blurry. From Google Payments he learned a partnership mentality: pick a set of folks and figure out how to create value together. Microsoft was an early test. The relationship was okay but not great, because Microsoft had a one-of-a-kind partnership with Databricks and was conflicted about whether Fabric or Snowflake was the answer. Satya Nadella, he says, is the master of winning partnerships. After a couple of years they are in a much better place than eighteen months ago: compete for some customers, collaborate where Azure plus Snowflake is a strict positive. The same attitude with AWS, and a similar arrangement in the works with GCP. With SAP they are thinking harder about one-plus-one-equals-three: bidirectional data share, plus analytics, AI, and agents on top of SAP data, using SAP's global presence as a leverage point to reach more companies. You cannot do this with every company. Create value together, and hopefully it is also profitable.

Every enterprise customer asks Guo where the highest-ROI AI use cases are. Ramaswamy's stack rank starts with a fact: every company now has technologists, even if they are not software companies. Coding agents are probably the easiest ROI, making new projects faster and demystifying technology so more people can get at it. Snowflake is a large user of coding agents and is working on ones that will be part of the product, because what is good for other people is also good for Snowflake. Customer support is next. It fits a pattern: a repository of human knowledge, an easy backup when the AI cannot do something, plus models that tap voice and typed questions and generate answers. Then faster, easier, more seamless access to data, especially when you do not have to pay a $50 per-user-per-month license. That is part of the motivation for Snowflake Intelligence: democratizing data access. Those are the areas of more or less guaranteed ROI.

Then he throws cold water. Obsessing about a lot of ROI too quickly is a bad idea for many companies. You do not want your first step to be a hundred feet. You want a lot of little things that prove value. People can get plenty from ChatGPT, even the free ones, for day-to-day work. The more companies demystify using AI, the more chance they have of figuring out value, because you take the risk. He places a lot of emphasis on how many shots you take on goal, how many projects you can run very quickly so you get a feel for the landscape of change. The sales data assistant had three versions before it. They all stuck. The first was enablement, the second added customer information, and there was a Streamlit Customer 360 app; all of it culminated in the combined assistant. That journey of iteration is often as important as launching the one big thing. He prefers not to take big bets, in engineering or in these projects. Iterate and create value every step of the way. The same advice goes to customers: you should not spend a lot of money on AI with Snowflake. Do it a thousand bucks at a time, and when you have significant value you feel good about, then you can ramp it up.

That is one reason he is bullish on applied companies that know their immediate usefulness. If you have customer trust and you understand the workflows, the landscape of what you could do with AI is very large. There is land for the taking for people with velocity and the paranoia to keep expanding. It is also an argument for why that layer should exist: it reduces time to value versus building on a generic framework or raw APIs. Customers have found that creating something like Cortex Analyst, which in a simple way is text-to-SQL, is a much harder problem than they think. There is more trust in Snowflake because they did many analyst projects before Snowflake Intelligence, a step-level increase in both capability and complexity.

Guo's other recurring question is what happens to internet ads if chat interfaces replace ten blue links. Advertising is an incredibly powerful medium and business, he says, and there are good ways of doing it. Seven years out of Google, he is more proud of the search-ads team's work now than when he left. You know good advertising when you see it. It will reinvent itself in the chat world. Let us just hope it does not become more insidious in discoverability, in being able to tell what is an ad. It would be creepy to have a psychiatrist with an affinity for prescribing one medication versus another. Those are the unfortunate things we will discover. The ad model is here to stay, in different forms, and as a consumer you have to be smart about what is in it for you. There is an increased premium on preserving our agency. He is encouraged that citations and sourcing have taken off. The set of things presented to consumers has narrowed a great deal, but the fact that people want primary sources is a useful indicator. A practical trick: take a Gemini Deep Research article, paste it into ChatGPT, and ask it to verify all the links. Neeva launched citations in early 2023 and was proud of it; the idea's relevance is as strong as ever. Deep Research products are amazing. You and I can get an expert paper on any topic. We just have to have the brainpower to digest it.

Last architectural question: are traditional information retrieval and indexing less relevant as more data lives in the model, even in the enterprise? It is tempting to trivialize search as just IR. The insight that powered Google was PageRank, harnessing the whole internet to figure out what was popular. PageRank ran out of juice in about six years, around 2004 or 2005. What became more and more relevant, though Google never really liked to talk about it, was click behavior on the results. That feedback loop is what eventually gave so much value. AI systems, including Snowflake Intelligence, need the same construct: eval loops are how you launch a meaningful product, and they are also how the product gets better over time. Maybe that gets encoded into context later. For now it is like asking whether LLMs should do math. You can argue they are so powerful they should. Any reasonable person will say no: write two lines of Python and run them. He thinks of trust the same way. There are well-known solutions for what is most trustworthy for a given question. Use them as a tool the agent calls, rather than in a maximalist mode where AI solves everything. Practical people use the best tools available. Search APIs exist because models benefit from external information they cannot easily internalize yet. You cannot be so smart that you do not use the computer.