Snowflake: CEO Says AI Is Collapsing Data Migrations From Years to Quarters, Reimagining Applications as Bundles of "Skills"
Goldman Sachs Communacopia + Technology Conference, September 8, 2026
Snowflake CEO Sridhar Ramaswamy and CFO Brian Robins used their appearance at Goldman Sachs' Communacopia + Technology Conference to lay out a more aggressive picture of how generative AI is compressing the traditional data migration timeline and reshaping what an enterprise application even is. The most concrete data point from the session: a large manufacturing client is targeting completion of a Teradata migration in under three quarters, a timeline Ramaswamy said would have been unthinkable in 2024. "Migration strikes terror in the heart of pretty much every CIO or CDO," he said, but AI coding agents are now allowing systems integrators to guarantee outcomes on fixed-cost contracts rather than the traditional time-and-materials model that has defined the services industry for decades.
Migrations Shift From Variable Cost to Fixed Cost
Ramaswamy framed this as part of a broader industrialization of software driven by AI, one he said poses a "profound threat" to every technology and software company, Snowflake included. The practical effect for Snowflake's business is that systems integrators are increasingly willing to price migrations as fixed-cost, outcome-guaranteed engagements rather than open-ended billable-hours projects, because coding agents have both accelerated the core migration work and made it dramatically easier to resolve the unpredictable edge cases that used to blow up project timelines. Robins confirmed that customer ramp curves are visibly steepening: "We do track internally how long it takes for them to get up to their consumption run rate, and we're seeing those curves get steeper and steeper and steeper."
Applications Reimagined as "Skills" on a Data Substrate
The most conceptually significant portion of the discussion centered on how Snowflake sees the application layer evolving. Ramaswamy argued that building software has become easy enough that the harder, more time-consuming part of shipping a traditional app, dealing with app stores and packaging, has become the bottleneck rather than the build itself. His proposed alternative is what he called a "handful of self-evolving skills running on top of a data substrate." Using an internal survey tool as an example, he described how a UI, database tables, notification logic, and access controls can all be assembled on demand within Snowflake rather than procured as a discrete SaaS product, with a React interface "conjured up on demand" when a user clicks a link. Crucially, these skills are not static: Ramaswamy said Snowflake's own internal support systems have evolved from manually reviewed workflows into agent-driven automation as usage patterns revealed which tasks could be handed off. This reframing, if it takes hold with customers, positions Snowflake to capture value that would otherwise flow to third-party application vendors, though Ramaswamy was careful to note this is not intended as a wholesale replacement for packaged software.
CoCo Deepens Through New Sales Motions and Job Functions
On Snowflake's CoCo agent product, both executives emphasized that the story is shifting from adoption breadth to usage depth. Ramaswamy described using internal usage data to recommend specific "skills" to customers based on observed repetitive behavior, alongside hands-on labs, two-to-three-hour technical tutorials run with customers, that the company has found highly effective at driving deeper engagement. Notably, Snowflake has created an entirely new role, the "activation engineer," specifically tasked with getting new logos live on the platform faster. Robins added color on how CoCo has broadened the buyer conversation itself: "A year ago, when I joined Snowflake, I hardly had any customer conversations. Today, every week, I'm meeting with three to five CFOs." Outcome-based pricing, he said, is the credibility mechanism unlocking this: customers who can be guaranteed a fixed price for a defined outcome are "all in" in ways they weren't under legacy time-and-materials arrangements.
Inference Strategy: A Value Hierarchy, Not a Reseller Play
Asked directly about the margin trade-off of selling pass-through inference into its installed base, Ramaswamy was blunt about where he draws the line. "If it is merely reselling undifferentiated capacity from a large supplier, you're not creating any value. That's fake news on the part of people that are doing this trying to pretend that they have a business," he said. Snowflake's preferred order of engagement is to sell CoCo or CoWork directly first; if a customer insists on raw model capacity to run its own harness, Snowflake will provide it, and if a customer wants only the data platform serving as a backend to a cloud provider, Snowflake will do that too, "somewhat more reluctantly." He pointed to scale economics, including the company's $6 billion AWS contract, as the mechanism for improving inference margins over time, alongside growing use of open-weight models run internally.
Gross Margin Guardrails on CoCo
Robins was explicit that any gross margin pressure from CoCo's success would be communicated proactively rather than surprising the Street. "If we were seeing that massive CoCo adoption beyond what we're already seeing, we would have the ability to communicate that to you within a quarter or two to manage that," he said, framing Snowflake's consumption-based model as inherently slower to ramp but more predictable once observed behavior accumulates. The company is still guiding to operating leverage for the full year despite the new product's margin profile.
An Honest Admission on Agent Latency
Ramaswamy did not dodge a pointed question about whether Snowflake's architecture, built originally for human-driven analytics rather than high-volume, low-latency agent queries, is fit for purpose. "The overall criticism that we have not addressed super low latency data well is very legit," he said, adding that Snowflake's streaming solution has brought data freshness down to two-to-three seconds but that getting to the sub-500-millisecond range required for some real-time agent use cases remains active, unfinished work. It is a rare instance of a CEO conceding a specific technical gap on stage rather than reframing the criticism, and it suggests investors should watch for concrete latency benchmarks in coming quarters as a gauge of competitive risk from newer, agent-native data infrastructure players.
Switching Costs Cut Both Ways
Ramaswamy also acknowledged the double-edged nature of faster migrations: if AI can accelerate migrations into Snowflake, it can just as easily accelerate migrations out. "If migration into Snowflake can be made a whole lot faster, migration out of Snowflake can be made a whole lot faster. That's the world we live in," he said. His response is to lean into open formats rather than fight them, pointing to Snowflake-managed Iceberg tables that let customers store data with Snowflake while keeping it queryable by other engines. Snowflake's defensibility, in his framing, now has to come from governance, disaster recovery, and agent tooling built on top of open data rather than proprietary lock-in, a notable philosophical shift from the company's earlier posture. "I think the Snowflake of old, which used to hang on to a set of what it thought were inviolable things that could never change, that company has also changed," he said.
Pricing Philosophy Amid Deflationary Pressure
On the mechanics of pricing new compute generations like Gen 2 instances, Robins said Snowflake manages the natural price-performance deflation of new hardware generations by offsetting it with volume growth and new workloads, avoiding step-downs in revenue while still passing performance gains to customers. Ramaswamy closed the pricing discussion by quoting a former boss: "Revenue solves all known problems," a line that underscores management's overall bias toward driving usage and adoption first, with margin optimization treated as a downstream consequence rather than a primary constraint.