Data
AI
The Stack Underneath
Hello, we are BRILLIAN’S. Trying to make an effort to put the right people for you to get the best results. Just insight !
Every team hits a point where the data it's sitting on outgrows the database it started on. The queries slow down, the sources multiply, and that's usually when people start asking which Snowflake features actually matter and which ones quietly run up the bill. Keeping everything in one place that's fast, secure, and affordable stops being simple, and this is the problem the platform was built to solve.
Before the features make sense, it helps to be clear on what Snowflake is underneath the marketing. It's a cloud data platform that runs across AWS, Azure, and Google Cloud, built for storage, analytics, governance, and increasingly for AI. It connects scattered data sources and the people who query them inside one environment, whether that's a production workload or a test run nobody wants near real systems. What makes it work isn't one headline capability. It's a handful of design choices underneath, plus a newer AI layer that changed what the platform is for.
We've put Snowflake into production enough to know which of those features earn their keep and which get switched on and forgotten. This walks through the ones that change how a data team works day to day, the AI capabilities Snowflake shipped over the past year, and which of them deserve your attention.
The thing worth understanding about Snowflake isn't any single feature. It's the architecture underneath, which solves a problem most teams only notice once their data outgrows the database they began on. These are the Snowflake features that genuinely change how a data team works, ranked roughly by how often they earn their cost on the builds we run.
This is the design choice every other Snowflake characteristic depends on. In most traditional databases, storage and compute are welded together, so adding more data forces you to pay for more processing power whether you need it or not. Snowflake pulls them apart. Your data sits in cloud storage, the compute that queries it runs separately, and you scale one without touching the other. A team holding two billion rows but running light queries pays for the storage and almost nothing for idle compute. Get this one idea and the rest of the list stops feeling like a feature dump and starts looking like consequences.
A virtual warehouse is the compute cluster you spin up to run queries, and you can run several at once, each sized differently and walled off from the others. Your analytics team can hammer a large warehouse with heavy reporting while your loading jobs run on a small one, and neither slows the other down. When the work stops, the warehouse suspends itself and stops charging you. We've watched teams cut compute spend just by right-sizing warehouses they'd over-provisioned out of habit, with no other change to how they worked.
Query load is never steady. It spikes when the whole company opens a dashboard at 9 a.m. on Monday and goes quiet by Friday afternoon. Snowflake absorbs those spikes with multi-cluster warehouses that add capacity when concurrent demand climbs and pull it back when demand drops. Nobody sits in a queue waiting for a query slot during the busy hour, and you're not paying peak rates at 3 a.m. when nobody's awake to run anything.
Here's a feature you never configure and mostly never see, which is exactly why it matters. Snowflake automatically splits every table into small chunks called micro-partitions, each holding a slice of rows along with metadata about what's inside it. When you query, Snowflake reads the metadata first and skips every partition that can't contain your answer. A filter on last week's data doesn't scan two years of history. There's no index to build, no partition key to choose, no maintenance window. The pruning just happens, and on large tables it's the difference between a query that returns in seconds and one that grinds through the whole table for no reason.
These two come from the same place, and together they're the features that make people relax. Time Travel lets you query a table as it existed at a point in the past, so when someone drops the wrong table or a bad job corrupts a column, you roll it back instead of restoring from a backup and losing a day. The retention window runs up to 90 days on Enterprise. Zero-copy cloning lets you copy an entire database without duplicating the underlying data, so the clone is instant and costs almost nothing until you start changing it. Spinning up a full production copy for a test run stops being a thing you schedule and becomes a thing you do before lunch.
Sharing data the old way means copying it, emailing a file, or standing up an API, and every copy drifts stale the moment it's made. Snowflake lets you share live data with another account without moving or duplicating it. The other side queries your data directly, always current, and you control exactly what they see. The Snowflake Marketplace runs on the same mechanism for third-party datasets you pull into your own environment. For companies that exchange data with partners or clients, this quietly removes a whole category of brittle plumbing.
Snowflake runs on AWS, Azure, and Google Cloud, and it behaves the same way on all three. If your company is already committed to one provider, Snowflake fits into it rather than forcing a second vendor relationship. A client locked into Azure for compliance reasons can run Snowflake without fighting their own infrastructure policy, which is the situation where this matters most. If you operate across more than one cloud, you can run it in each and share data between them.
Snowflake is fully managed, so there's no server to patch, no index to tune by hand, no storage volume to expand before it fills up. The work that normally eats a data engineer's week, the upgrades and the partitioning and the vacuuming, mostly disappears. This is the feature that never shows up in a demo and always shows up in your team's calendar. For a lean team, the hours you don't spend on upkeep are hours that go back into real data work.
This is the part that changed what Snowflake is for. Cortex AI is a managed layer that brings large language models and machine learning into Snowflake, callable in plain SQL, with your data never leaving the platform. The pieces most teams actually touch:
Cortex AI runs sentiment analysis, summarization, and text generation on a column the same way you'd run any other SQL function.
Snowpark runs your Python data prep and ML where the data already lives, instead of shipping it out to a separate system.
The feature store manages the inputs that feed your ML models, so the same features stay consistent across training and production.
Document AI pulls structured fields out of invoices, contracts, and reports through AI_EXTRACT.
Most of what shipped over the past year landed on the AI side, which tells you where Snowflake thinks the platform is going. If you last looked at it as a SQL warehouse, here's what's changed. These are the Snowflake new features worth knowing about, with dates, because anything moving this fast is worth checking against the live release notes before you build on it.
Cortex Code is a Snowflake-aware AI coding agent for writing and debugging SQL and Python against your own tables. It launched in November 2025 and is already used by more than 50% of Snowflake customers.
Snowflake Intelligence turned into a work agent for business users that learns individual preferences and automates tasks against governed data, expanded in April 2026.
AI_EXTRACT went generally available in October 2025, replacing the older Document AI workflow for pulling fields out of unstructured documents.
The AI credit overhaul in April 2026 moved Cortex to a flat $2.00 per credit globally, decoupled from your Snowflake edition, with cost reductions of up to 70% reported.
A $200M OpenAI partnership in February 2026 embedded OpenAI frontier models natively inside Cortex across AWS, Azure, and Google Cloud.
The honest read on all of this. Most teams adopting Snowflake today still do it for the warehouse, not the AI, and the AI features are a reason to stay rather than the reason to start. The exception is the credit change, which matters to anyone already running Cortex, because the cost math shifted underneath them and an estimate built six months ago is now wrong. That's the one release worth re-running your numbers on.
Here's the uncomfortable part. Most teams use a fraction of what Snowflake offers and pay for the platform anyway, which is fine, but it means the feature list above isn't a shopping list. It's a menu where three or four items carry the whole meal. After enough builds, the sort looks like this.
The features you'll use whether you notice them or not:
Separation of storage and compute, micro-partitioning, and near-zero maintenance. Nobody picks Snowflake for these, but they're the reason the bill stays sane and the queries stay fast. They run in the background and you never configure them.
The features that earn their cost once you're past the basics:
Virtual warehouses and multi-cluster scaling, the moment more than a handful of people query the same data at once.
Time Travel and zero-copy cloning, the first time someone drops a table or you need a full test environment by lunch.
Data sharing, if you exchange data with partners or clients. If you don't, it sits unused.
The features most teams switch on and forget:
A chunk of the AI layer. Cortex is genuinely useful for the teams whose roadmap leans on it, but plenty of teams turn it on, run a demo, and never build the pipeline that justifies the spend. If your center of gravity is heavy ML and model training, that's worth a hard look, because it's the exact ground where teams end up comparing Snowflake against Databricks and sometimes find the engineering-first feature set fits the work better.
So which features matter for your workload depends on one question. Are you mostly running SQL analytics that lots of people hit at once, or are you doing heavy transformation and machine learning? Snowflake is excellent at the first and competent at the second. If you're squarely in the second camp, the special features people praise Snowflake for aren't the ones you'll lean on, and that's worth knowing before you commit.
The features people sell Snowflake on aren't usually the ones that make it worth running. The architecture underneath, the storage-compute split, the micro-partitioning, the maintenance that just doesn't happen, is what keeps it fast and affordable while your data grows. The AI layer is real and getting better, but for most teams it's a reason to stay, not the reason to start.
The catch is the same one the platform keeps circling back to. Snowflake will hold your data and scale your queries without much fuss. Running it efficiently, warehouses right-sized and idle compute off, is where the actual engineering lives, and it decides whether your bill makes sense six months in. That's the part we care about most, and the bulk of our data engineering work is making sure Snowflake lands on the right side of it. Before you commit, map your real workload against the three-way sort above. If you're mostly running SQL analytics for a crowd, it's the easy call. If you're heavy on ML, ask harder questions first.
The features that matter most are the architectural ones: separation of storage and compute, automatic micro-partitioning, multi-cluster scaling for concurrency, and near-zero maintenance. On top of those sit data sharing, Time Travel, zero-copy cloning, and the Cortex AI layer. Most teams lean on three or four of these and pay for the rest without noticing.
The big one is that Snowflake splits storage from compute, so your data can grow without forcing you to pay for more processing power. Traditional databases weld the two together. That single design choice is why Snowflake scales the way it does and why you can run several isolated workloads at once without them slowing each other down.
Snowflake's AI features run through Cortex AI, a managed layer that brings large language models and machine learning into the platform, callable in plain SQL with your data never leaving Snowflake. It covers Cortex AI functions for text analysis and generation, Snowpark for Python and ML, a feature store for managing ML inputs, and Document AI for pulling structured data out of unstructured files.
The recent releases are mostly AI. Cortex Code, a Snowflake-aware coding agent, launched in November 2025 and is now used by more than half of customers. Snowflake Intelligence became a work agent for business users, AI_EXTRACT went generally available for document extraction, and an April 2026 credit change moved Cortex to a flat $2.00 per credit globally.
For a small team running SQL analytics, usually yes, because the maintenance burden is close to zero and you're not paying for idle compute. Where it gets expensive is neglect. An idle warehouse left running or an oversized one quietly runs up the bill. The platform rewards attention and punishes forgetting, so the value depends less on the features and more on whether someone keeps an eye on usage.
You might also like
