AI usage isn’t the same as AI value. That’s why Snowflake introduced “intelligence efficiency” this week.
In a post on Snowflake’s blog, CEO Sridhar Ramaswamy laid out how companies are pouring money into AI, and boards want proof it’s translating into more revenue or faster execution. Cortex AI Gateway now does dynamic model routing, evaluating each task against cost, quality, and latency, picking the best model for the job from a stable of open and proprietary options, then learning from the result over time.
Model choice matters, and routing intelligently across Anthropic, Google, Mistral, OpenAI, and the rest is the freedom enterprises should demand from their AI stack. We’ve made the same case about compute for years: don’t let one vendor lock in your workloads. Snowflake extending that argument to models is a natural next step, and it’s a big part of why Snowflake chose AtScale to power CoCo and Cortex through governed Semantic Views in the first place. We’re already the layer that makes any model, in any tool, trustworthy against the same data. Model choice and compute choice were always going to need each other.
Where Routing Ends and Governance Begins
Picking the best model for a task doesn’t tell you whether it knows what your business means by “revenue.” Ask even the most capable model “what was our revenue last quarter,” and it has to guess: recognized or booked, gross or net, billed or collected? Route that question to ChatGPT, Claude, or Gemini, and each one can guess differently, and each answers with total confidence. That ambiguity was never the router’s job to solve. Governance exists to fix exactly that.
This is the layer AtScale adds underneath the routing decision. Revenue, churn, margin, active users are governed definitions that have to compute the same way every time, no matter which model or tool is asking. AtScale’s AI Computation Engine, ACE, codifies that meaning once and computes it consistently, so whichever model Cortex AI Gateway routes to gets the same trustworthy number. In benchmark testing with a Tier 1 bank, that governed computation layer cut warehouse compute by up to 21,000x and eliminated roughly $9 million a year in what we call the AI rediscovery tax, the cost of asking the same question over and over because no one agreed on the answer the first time. Anthropic’s own agent accuracy moved from 21% to past 95% once its context was governed this way.
Freedom of Choice, Extended to Compute
Snowflake’s right that enterprises shouldn’t be locked into one model. Extend that same freedom to compute: enterprises shouldn’t be locked into one warehouse or one vendor’s private definition of “revenue” either. The router decides which model answers your question. The computation layer makes sure that answer holds up, and keeps holding up as you swap models, warehouses, and agents underneath it.
That’s the synergy we see here: Snowflake solving model efficiency, AtScale solving computational accuracy. Put them together, and intelligence efficiency becomes something you can actually measure.
Reviewed by: Mark Palmer
SHARE
Power BI/Fabric Benchmarks