Sarah runs analytics at a mid-sized ecommerce company. Last quarter her Segment bill came in at $11,400. This quarter it hit $17,800. Same product, same team, just more users. Her CEO asked a fair question: why does the customer data platform cost more than half the engineering team?
She didn’t have a good answer. She had a common one.
Companies of every size are quietly pulling apart their data stacks. Segment, mParticle, and other all-in-one customer data platforms used to be the default choice. Now the trend runs the other way. Snowflake plus dbt plus Hightouch or Census keeps showing up as the replacement pattern. It looks messier on paper. It costs less. And it gives data teams something they thought they had already lost, control.
2026 is the year composable data stacks stopped being a Reddit thread and started being a boardroom decision. Reverse ETL moved from fringe concept to core infrastructure. Warehouse-native architecture graduated from experiment to standard practice.
The obvious question is why. Breaking a single vendor into four or five moving parts sounds like inviting complexity for its own sake. The answer sits in a three-way tension between cost, flexibility, and control.
The Trap of the All-in-One Promise
Traditional CDPs sold a clean story: hand us your data and we handle everything. Ingestion, identity resolution, profile building, activation. One vendor, one contract, one throat to choke.
The story worked for a few years. Then reality set in.
Vendor lock-in is the first crack. Customer data goes into Segment’s proprietary format. Want to switch tools? First export everything, then map it to a new schema, then rebuild all your downstream connections. Companies discovered that CDPs behave less like tools and more like fly traps. Getting in is easy. Getting out takes six months and a consultant.
The second crack is the black box. What actually happens to data once it enters the platform? The identity resolution logic is opaque. Profile merging follows rules the vendor won’t fully document. When something looks wrong in a downstream system, tracing the cause means opening a support ticket and waiting.
Then there’s the pricing model. Segment charges by monthly tracked users, or MTU. Your user base grows fifty percent? Your bill grows fifty percent. Traffic spikes because of a launch or a viral moment? Congratulations, next month’s bill spikes too. For a company scaling from one million to five million users, CDP costs can outrun almost every other software line item on the balance sheet.
A 2026 industry survey put the median annual CDP spend for mid-market SaaS companies at around $120,000. The equivalent composable stack, warehouse plus tooling included, runs roughly $60,000 to $80,000. The savings pay for a data engineer, with room left over for training and coffee.
What Composable Actually Means
The composable data stack works from one core idea. The data warehouse is the single source of truth. Every other tool is a replaceable component that reads from or writes to the warehouse.
A typical setup looks like this. Snowflake, BigQuery, or Databricks stores raw and cleaned data. Fivetran or Airbyte pipes data from source systems into the warehouse. dbt handles transformation, defining customer profiles, metrics, and segments through SQL models that live in version control. Hightouch or Census pushes segment data back out to marketing tools, CRMs, and ad platforms. Looker or Metabase queries the warehouse directly for reporting.
Data lives in one place. Change your ETL tool? The data stays put. Swap your reverse ETL vendor? Your SQL models still work. Bring in a new BI tool? Just point it at the warehouse.
That’s what composable means in practice. Each layer can be swapped without disturbing the others. No vendor holds your data hostage. No single contract renewal defines your architecture for the next three years.
Reverse ETL Is the Missing Piece
For years the ETL story flowed one direction. Data came from apps and databases into the warehouse for analysis. What happened next was mostly a spreadsheet emailed to marketing.
Reverse ETL flipped the arrow. Now the warehouse pushes data back to operational tools. It sounds counterintuitive until you see it work.
Say a data analyst writes a dbt model that defines high-value customers as anyone who spent more than $1,000 in the last thirty days. That model runs as scheduled. A reverse ETL tool like Hightouch reads the output and syncs it in near-real-time to four places. Salesforce, where the sales team sees updated lead priority. Braze, which fires a VIP offer email. Google Ads, which builds a lookalike audience from the segment. Zendesk, where support agents see a VIP flag on incoming tickets.
The frequency is your choice. Real-time, hourly, or daily. The logic is a SQL query that anyone with warehouse access can read and modify. No dragging blocks around in a proprietary UI. No filing tickets to a customer success manager who takes two weeks to respond.
Hightouch and Census run neck and neck as the two big names in this space. Independent estimates put the reverse ETL market at roughly $850 million in 2026, growing around 67% year over year. That growth curve is what happens when a category solves a real problem that no incumbent solved cleanly.
Why Enterprises Keep Choosing Warehouse-Native
Cost is the loudest reason. Snowflake charges by query volume. Hightouch charges by rows synced. When your user base grows, warehouse costs move less than proportionally because query patterns don’t scale one-to-one with users. Reverse ETL costs track only the data you actually push. A CDP charging per MTU will grow its bill in lockstep with your business. A warehouse-native stack tends to grow slower, sometimes much slower.
Flexibility is the second reason. Need a new customer segment? Write a few lines of SQL in dbt, five minutes of work. Same task in a traditional CDP means either fighting a drag-and-drop UI with real limitations or waiting weeks for the vendor to build custom logic. Need to connect a new marketing tool? Modern reverse ETL platforms support over a hundred destinations. If the CDP doesn’t support what you need, you’re stuck.
Data sovereignty is the third reason. Your data lives in your warehouse, under your access controls, in your cloud region. Want to back it up? Export it. Want to audit access? Query the logs. Want to train a model on your customer data? It’s already sitting where your ML pipeline can read it. When the CDP holds your data, all you have is API access and vendor goodwill. If the vendor gets acquired or shuts down, you inherit a hard problem overnight.
The Catch
Composable stacks aren’t a free lunch. Anyone who tells you they are hasn’t tried to run one at scale.
Complexity is the obvious tax. An all-in-one CDP means one vendor relationship, one dashboard, one bill. A composable stack means four or five vendors, each with their own auth, their own quirks, their own support experience. Something breaks and the debugging chain gets longer. Data didn’t sync to Salesforce? Maybe the dbt model has a bug. Maybe the reverse ETL config points to the wrong column. Maybe a warehouse permission got revoked. Maybe Salesforce’s API changed. You’ll learn to check all four.
Latency is the second issue. Traditional CDPs move data in real time or close to it. A warehouse-native pipeline has more hops. ETL syncs typically run hourly. dbt transformations run on schedules or triggers. Reverse ETL pushes data in windows measured in minutes or hours, not seconds. If you need to personalize a website recommendation the instant a user clicks something, warehouse-native won’t get you there. This is why many companies keep a CDP for real-time use cases and use the warehouse-native stack for everything else.
Talent is the third hurdle. Someone on your team needs to write SQL, understand data modeling, and configure API integrations. Small teams often can’t afford that person. All-in-one CDPs have a real advantage here. A marketing manager can build a segment without writing code. That’s not nothing, especially at companies where the marketing team runs at three times the size of engineering.
The Hybrid Answer Most Companies Land On
Actual production architectures rarely look like pure composable or pure all-in-one. Most large and mid-sized companies run some version of the hybrid stack.
A common pattern uses Segment or a similar CDP for real-time event collection, because that’s what CDPs do well. The CDP then pipes events into Snowflake through a warehouse connector. dbt handles the deep modeling inside the warehouse. Hightouch or Census sends warehouse-defined segments back out to marketing tools. A BI tool queries the warehouse for reporting.
You get the CDP’s real-time capabilities and its friendly interface for marketing users. You get the warehouse’s flexibility and cost profile for everything analytical. It’s more moving parts than either extreme, but it maps better to how real companies actually operate.
Industry surveys from 2026 suggest that around 60% of mid-market and enterprise companies now run some flavor of hybrid stack. Pure all-in-one has retreated to smaller and earlier-stage companies. Pure composable is common among data-mature engineering teams who see the CDP as extra middleware they don’t need.
How to Choose
The right architecture depends on what your team looks like and what problems you’re trying to solve.
An all-in-one CDP fits when your team has few or no data engineers, when marketing and operations people need to build segments themselves, when your data volume is small enough that per-MTU pricing doesn’t sting, and when your use cases stay in familiar territory around simple segmentation and standard automation.
A composable stack fits when you already have a data team who can write SQL and manage pipelines, when your data volume makes CDP pricing painful, when you need custom logic that off-the-shelf tools can’t easily provide, and when you already run a warehouse and want to get more value out of it.
A hybrid stack fits when you need both real-time event handling and deep analysis, when your team mixes marketing users who want a friendly interface with data users who want SQL access, and when you’re willing to trade some architectural complexity for capability.
None of these are permanent choices. Companies move between them as they grow. The interesting shift in 2026 is that the direction of travel is almost always the same, from pure CDP to hybrid, from hybrid toward more composable. Very few teams migrate back.
Where AI Fits
One reason composable stacks are getting easier to run is that AI is starting to smooth over the parts that used to require deep SQL expertise.
Hightouch and Census both ship AI features that turn natural language into SQL. Describe a segment in plain English, “users who spent over $500 in the last thirty days but haven’t logged in for a week,” and the tool generates the query. A data analyst still reviews the output, but the drafting work compresses from twenty minutes to two.
dbt is heading the same direction. AI helps with model authoring, catches quality issues, and suggests optimizations. Some of this is still marketing more than substance, but the trajectory is clear. The composable stack is becoming more accessible to people who couldn’t run one two years ago.
The counterargument comes from the CDP camp. If AI can automate segmentation, quality checks, and campaign strategy inside an all-in-one platform, the black box problem stops mattering. Users don’t need to see the plumbing if the plumbing does what they need. That argument has real weight, and the CDP vendors are betting the business on it.
Which side wins depends on questions no one has answered yet. Whether the composable stack’s flexibility can hold its lead as AI closes the ease-of-use gap. Whether the CDP’s better UX can overcome the cost and lock-in problems as AI makes the black box less scary. The next few years will settle it.
What Should You Do This Quarter
If you’re running a pure CDP and your bill is climbing faster than your revenue, price out a composable stack. The exercise alone often uncovers spending you can cut regardless of whether you actually migrate.
If you already run a warehouse but haven’t wired up reverse ETL, that’s the highest-leverage next step. It turns the warehouse from a reporting tool into an operational one. Marketing tools get better data. Sales sees better lead scoring. Support sees richer customer context. All from work you’ve already done.
If you’re at the early stages and haven’t picked a stack yet, the honest advice depends on your team. Data-heavy teams should start composable. Marketing-heavy teams should start with a CDP and plan to move data into a warehouse when volume justifies it. Either way, keep an eye on total cost of ownership, not just the headline monthly fee.
The bigger point is that composable isn’t a fad. The architectural shift is real, the economics favor it at scale, and the tooling has matured enough to make it accessible. All-in-one CDPs aren’t going away, but their role is narrowing. In 2026, the question isn’t whether to embrace warehouse-native architecture. It’s how far along that path you want to be by year end.
Sarah, from the opening scene, ended up going hybrid. Kept Segment for event collection because ripping it out would have broken twenty things. Piped everything into Snowflake. Built her segments in dbt. Wired Hightouch to push them back to Salesforce, Braze, and Google Ads. Her next quarter’s bill came in around $9,200. Better than the $17,800 she was staring at three months earlier. Not as low as the pure composable number she originally imagined, but low enough that her CEO stopped asking questions.
That’s usually how these decisions actually play out. Not clean revolutions. Just gradual, defensible cost curves that bend the right direction.
Stay updated with our latest AI insights
