No Priors Ep. 139 | With Snowflake CEO Sridhar Ramaswamy
Summary
Ramaswamy’s 18-month reset treated Snowflake’s AI lag as an organizational-speed problem: shorten the seven-to-10-layer distance between engineers and customers, assign accountable product leaders, and connect them closely to go-to-market teams. His operating maxim is “speed wins. The ability to iterate always trumps carefully laid-out strategies,” particularly when AI’s next month is barely predictable.
Snowflake pivoted away from foundation-model development after recognizing it lacked the capital to compete meaningfully with OpenAI or Anthropic, then repositioned from the Data Cloud to the “AI Data Cloud.” The narrower bet is to compound its installed-base advantage—“something like half” of qualifying Fortune 2000 companies—by applying search, text-to-SQL and agents to valuable customer data already in Snowflake.
Snowflake Intelligence is an opinionated enterprise-data agent, not a universal agent framework or a replacement for SAP, Salesforce, Tableau or Sigma. Its Raven sales assistant combines contracts, consumption, conversations and outstanding issues in one interface, while required evals reject “YOLO AI”: changing a model must not silently break existing answers.
The moat must be rebuilt continuously because foundation-model companies are “empires that have not met their oceans just yet,” while cloud providers possess “infinite budgets” and “infinite patience.” Thin prompt layers look exposed; Snowflake’s defense is a cross-cloud, governed data platform plus deeper Microsoft, AWS, GCP and SAP integration. As Ramaswamy warns, merely being ahead is insufficient: fail to stay ahead and “you will be Intel.”
Ramaswamy identifies coding agents, customer support and easier data access as AI’s clearest near-term enterprise returns. But he rejects giant first bets: take more “shots on goal,” iterate toward product fit, and spend with Snowflake “a thousand bucks at a time” until demonstrated value justifies scaling.
Internet advertising will survive chat interfaces, but disclosure and user agency become more important as chat narrows what is presented and commercial influence becomes harder to see. His deliberately creepy failure case is a psychiatrist biased toward one medication; the counterweight is visible sourcing, citations and easy cross-checking between systems such as Gemini and ChatGPT.
Search and other reliable external tools remain relevant even as LLMs grow more capable. Google’s advantage moved from PageRank to behavioral feedback, just as AI products can improve through eval loops; asking an LLM to internalize everything is like refusing two lines of Python for arithmetic because “you cannot be so smart that you don’t use the computer.”
Deep dive
1. Snowflake’s reset began by shortening the path to customers
Ramaswamy’s account of the CEO handoff: Snowflake’s original product was years ahead, but the company reacted slowly to machine learning and AI. Frank anticipated a more tumultuous product era and pushed for a product-first successor; extensive customer conversations—and their enthusiasm for Snowflake—convinced Ramaswamy to accept.
Hypergrowth above 100% year over year had produced specialization everywhere, leaving “seven to 10 layers of teams” between an engineer building a feature and the customer using it. That structure worked with perfect product-market fit, but not when “we can barely tell what’s going to come out next month.”
The first six months centered on accountability: distinct leaders for AI and core warehousing, paired with specialized product, engineering, marketing and go-to-market teams. Snowflake also created a credible foundation model early last year, recognized the capital disadvantage against OpenAI and Anthropic, and pivoted toward the “AI Data Cloud,” with an emphasis on faster iteration.
2. Snowflake Intelligence trades infinite flexibility for trusted answers
Ramaswamy contrasts SI with platforms promising data from anywhere, arbitrary workflows and “one agent will rule them all.” Infinite possibility makes it harder to know what to build; Snowflake instead targets faster value from structured and unstructured enterprise data through components such as search and text-to-SQL.
The product attacks dashboards’ core limitation: “A dashboard is a 2D view of a complex surface.” Snowflake’s internal Raven assistant combines customer contracts, consumption, recent conversations and unresolved issues; early work with Cisco, Fanatics and the USA Bobsled team extended that pattern beyond Snowflake.
The intended user is every employee, “not people who can write SQL.” Canned prompts prevent blank-page paralysis, while users can ask which datasets and questions are available. Ramaswamy says he would not enter a customer meeting without first checking the latest relationship context.
Trust is treated like software correctness, with “a right and there’s a wrong,” rather than “YOLO AI.” Every new capability needs an eval, and model changes must be checked against existing behavior. SI remains narrower than Tableau or Sigma, is described as a consumption product rather than another per-seat subscription, and is adding identity-provider integration for broad deployment while Snowflake experiments with ways to avoid runaway costs.
3. Organizational change travels through internal champions
Leadership alignment and the cross-functional “war room” or pod model came first because they affected relatively small groups. Broader behavioral change was deliberately staged: “Change is hard,” particularly when skeptical employees must alter daily workflows rather than merely accept a new strategy.
Coding-agent adoption combined executive direction with grassroots credibility. Founder Benoit’s enthusiasm persuaded engineers more effectively than CEO directives, reinforcing the prescription to find champions, encourage them and elevate them. Sarah describes the relevant champions as curious people willing to experiment over weekends. Solution engineers now use agents to turn canned demos into customer-specific prototypes with synthetic data.
Each earlier role changed his leadership: a PhD taught him to compress ideas into crisp four-line abstracts; Google showed the extraordinary distribution that could put one person’s three-month project in The New York Times; Neeva’s painful, possibly-too-early outcome taught hustle, marketing and not taking success for granted.
4. The durable moat is a data platform that keeps moving
Product-market fit remains “lightning in a bottle”: all three hyperscalers would prefer to own the data space, yet Snowflake and Databricks exist. That does not confer permanent safety; it shows that a focused product can beat bundled services when its differentiated value is strong enough.
OpenAI and Anthropic are “empires that have not met their oceans just yet,” so builders must anticipate their likely expansion. Coding agents sit clearly in their path, and a product consisting mainly of prompts over one model is vulnerable; durability requires meaningful distance between the model and the value delivered.
Guo’s formulation—“defensibility is built, not strategized”—wins Ramaswamy’s agreement. Cloud providers have effectively unlimited patience and budgets, so companies must “not just be ahead but stay ahead”; otherwise, his blunt endpoint is, “you will be Intel.”
Snowflake’s aspiration runs “from inception to insight.” Google and Meta exemplified a data-first model in which behavior fed back into products quickly; Ramaswamy notes that his Google data teams were as large as the product teams. Snowflake aims to provide that capability across clouds through shared, governed data and integrated AI—a higher abstraction than buying raw compute and storage and writing everything yourself.
5. Partnerships and small experiments are the commercialization strategy
Snowflake is moving beyond a Snowflake-centric worldview. Its previously conflicted Microsoft relationship now spans Fabric integration and a more workable operating posture: the companies may compete for some customers while treating Azure plus Snowflake as a “strictly positive” combination elsewhere. Ramaswamy describes the same posture with AWS and a similar arrangement under development with GCP.
With SAP, the ambition is a “one plus one equals three”: bidirectional data sharing plus joint analytics, AI and agents over SAP data. SAP’s global footprint could also expand Snowflake’s distribution, though Ramaswamy stresses that partnerships this deep can only work with a select group.
His clearest ROI ranking starts with coding agents, then customer support—where models can access institutional knowledge across voice and text with humans as fallback—and faster, easier data access without a “$50 per user per month license.” These are “more or less guaranteed ROI” areas, not promises that every workflow should immediately become agentic.
Guo argues that trusted applied vendors compress time-to-value versus generic frameworks; Ramaswamy’s evidence is Cortex Analyst, whose apparently simple text-to-SQL problem proved much harder than customers expected. Still, he advises against 100-foot first steps: Raven followed two or three earlier versions, including enablement, customer information and a Customer 360 Streamlit app, and customers should take more “shots on goal” while spending “a thousand bucks at a time.”
6. Ads and retrieval survive because intelligence still needs accountability
Advertising is “here to stay” and will reinvent itself for chat, but Ramaswamy worries about commercial influence becoming less discoverable. A psychiatrist quietly favoring one medication is his cautionary example; preserving user agency requires consumers to understand “what’s in these things for you.”
Ramaswamy agrees that the rise of citations and sourcing is encouraging even as chat experiences narrow what is presented. A Gemini deep-research article can be pasted into ChatGPT for link verification, while expert-level papers are now available on almost any subject. Neeva’s early-2023 citation work, he argues, remains highly relevant.
Search is more than retrieval: PageRank “ran out of juice” around 2004–05, and Google’s click-feedback loop became the deeper advantage. AI systems similarly need eval loops both to launch meaningfully and improve. Like using Python for arithmetic, calling search or another proven tool is rational intelligence: “You cannot be so smart that you don’t use the computer.”