Oct 8, 2024, 5:00 AM ETFireside Chat, Interview

Snowflake CEO Sridhar Ramaswamy on Using Data to Create Simple, Reliable AI for Businesses

Sridhar Ramaswamy, a former Google executive (joined April 2003) and founder of the AI search startup Neva, serves as CEO of Snowflake.
Snowflake positions itself as the "AI data cloud," a platform centered on data that transforms how information is stored, moved, and accessed.
The company reported $2.6 billion in revenue last year with over 10,000 enterprise customers.
Snowflake acquired Neva, integrating its search technology expertise into its enterprise data strategy.

Enterprise adoption of AI is shifting from skepticism to high awareness, driven by the tangible capabilities demonstrated by models like GPT-4.
Key adoption patterns observed include:
- Democratizing data access for business users without requiring data analysts or BI tools (e.g., Bayer).
- Transforming unstructured data (images, transcripts) into structured insights via AI models rather than manual engineering.
- Using "Document AI" to extract structured data from contracts and other documents.
Customers are increasingly bringing new data sources into Snowflake, moving beyond traditional warehouses to cloud storage formats like Apache Iceberg for interoperability.
A significant trend is the shift from building custom AI solutions on hyperscalers (which requires extensive engineering) to using managed platforms like Snowflake Cortex for out-of-the-box governance and security.

Snowflake targets 90-99% reliability for "talk to your data" applications (Cortex Analyst), contrasting sharply with the ~45% reliability of raw GPT-4 models.
Reliability is achieved through:
- Software Engineering over Model Tuning: Treating AI as a software engineering problem by restricting domains and managing schema context rather than relying solely on model intelligence.
- Context Engineering: Leveraging knowledge of internal schemas and semantic context to prevent hallucinations in environments with millions of tables (e.g., distinguishing between thousands of "revenue" metrics).
- Specialized Sub-tasks: Using different models for different functions (e.g., distinguishing between "should we answer?" vs. "how to answer?").
Snowflake's "right to win" is defined by making AI trivial for data already within its ecosystem, turning months-long software engineering projects into two-command analyst tasks.
The company emphasizes data security and governance, guaranteeing that customer data is never used to train cross-customer models.

Snowflake has successfully inverted its product velocity, accelerating release cycles despite its size.
Key drivers of this speed include:
- Safety Nets: Implementing auto-experimentation frameworks and regression testing to detect issues before widespread rollout.
- System Extensibility: A two-year initiative to design a more extensible architecture, allowing teams to build AI features without breaking the core platform.
- Leadership Accountability: Enforcing strict clarity on deliverables and weekly calibration on promises.
Ramaswamy argues that "virtuosity" (speed of execution and adaptability) trumps static strategy in the rapidly changing AI landscape.
The company is moving toward "grounded chatbots" that cite sources and include built-in test frameworks, treating AI output with the same rigor as traditional software acceptance criteria.

GPT-5 & Reasoning: The market awaits GPT-5 for potential breakthroughs in multi-step reasoning, though significant value can still be extracted from current models through better application design and software engineering.
Search Dynamics: Ramaswamy views the search industry as a battle of distribution rather than pure product quality, noting that consumer choice is often limited by default settings and incumbent lock-in.
Incumbents vs. Startups: Historical data suggests that well-funded incumbents (like Google) are better positioned to leverage AI than startups, unless the startup creates a new category (e.g., image or video generation).
Foundation Models: The cost of training top-tier models is likely to consolidate the field, with open-source models remaining a viable niche only if backed by strong businesses or hyperscalers.
Consumer Mobile: Ramaswamy predicts significant potential for AI integration in mobile ecosystems (e.g., Apple's ChatGPT deal) due to the controlled nature of the platform and the ability to mandate APIs for language models.

Snowflake is pivoting from a partnership-focused AI strategy to an internal-first approach, building core products like Cortex Analyst to directly serve business users.
The long-term bet is that data ecosystems will move upstream, with data residing in open, interoperable cloud storage formats (Iceberg) accessible via open catalogs.
The ultimate goal for AI is to lower the barrier to software creation and consumption, effectively democratizing the ability to encode thinking and act on data for all users.
Ramaswamy envisions AI as a "new layer" between humans and software, enabling the creation and use of software by vastly more people, similar to the democratization of Google Search via Android.

Admired AI Talent: Researchers who achieve high impact under tight budget constraints (e.g., Arthur/Reka teams, Snowflake's internal "doers").
Favorite AI App: ChatGPT, specifically for its daily utility in converting text to structured data (CSVs) and aiding in language learning.
Missing Application: An "AI phone mediator" capable of seamlessly interacting between different apps (e.g., copying a calendar address directly into a rideshare app).
Optimistic Vision: The next 5-10 years will see AI vastly increase accessibility to software creation and usage, acting as a transformative layer for humanity similar to mobile search.