> ## Content Index > Fetch the complete content index at: https://www.martechtherapy.com/llms.txt > Use this file to discover other available public pages before exploring further. # “Open CDP & Local AI CDP” Oh my… - URL: https://www.martechtherapy.com/sovereign-cdp-open-cdp-local-ai/ - Published: 2026-10-22T07:23:10.000Z - Updated: 2026-10-22T07:23:10.000Z - Description: Two launches move the CDP's compute and its model inside your own boundary. Sovereignty doesn't make it cheaper, and the open question is who it's actually for. - Author: Matthew Niederberger - Tags: Customer Data Platform, Composable CDP, Data privacy, Martech readiness In a regulated company, the question that stalls an AI project on customer data rarely comes from marketing. It comes from risk and legal, and it's simple enough. > Where does the customer data get processed, and where does the model run? For a long time "in the vendor's cloud" was enough. Then composable came along and "it never leaves our warehouse" became the better reply. For banks, insurers and telcos, that reply is getting harder to approve too, because the warehouse belongs to somebody, and so does the model the agents call. I'll be upfront about where I'm standing. I've spent a lot of time with composable architectures and what they cost to run. I haven't run a CDP on bare-metal Spark or hosted a language model next to one. So this is me reasoning from the architecture and from what two vendors are saying, and you should treat it that way. ## The line keeps moving down Two launches, five weeks apart, push on the same boundary from different directions. In September, Zeotap announced [Open CDP](https://zeotap.com/blog/zeotap-launches-open-cdp-the-first-fully-composable-cdp-that-runs-on-open-source-compute-and-the-customers-own-lakehouse-sovereign-by-design/?ref=martechtherapy.com). It runs their composable CDP on open-source Apache Spark over Iceberg tables in the customer's own object storage, governed through whatever catalog the customer already uses. Their case is that warehouse-native composable still rents a proprietary warehouse to do the work, which ties you to one cloud and bills every audience refresh as consumption. Composable, in their telling, stopped a layer too high. In October, Meiro announced [Local AI CDP](https://meiro.io/blog/local-ai-cdp-for-sovereign-ai-inference/?ref=martechtherapy.com), which keeps the model, large or small, inside the same deployment as the customer data. Their case is that warehouse-native and zero-copy settle where data is stored, and leave open where inference happens. If your agents send customer context to an external model to pick the next best action, the data stayed home and the reasoning didn't. Compute in one case, the model in the other. Both describe sovereignty as something the architecture delivers rather than something a contract promises, in close to identical wording. [![Diagram: which CDP layers sit inside your own boundary across packaged, warehouse-native composable, Zeotap Open CDP and Meiro Local AI CDP, with the people who run it as the one layer nobody sells as swappable.](https://storage.ghost.io/c/57/c5/57c5ffb5-ecd9-4bf6-b870-68986d981176/content/images/2026/10/who-is-a-sovereign-cdp-actually-for-sovereign-f1-the-line-moves-down.png)](https://storage.ghost.io/c/57/c5/57c5ffb5-ecd9-4bf6-b870-68986d981176/content/images/2026/10/who-is-a-sovereign-cdp-actually-for-sovereign-f1-the-line-moves-down.svg?ref=martechtherapy.com) When I wrote about [Databricks CustomerLake](https://www.martechtherapy.com/databricks-customerlake-the-warehouse-cdp/) and [Snowflake recruiting Zeotap](https://www.martechtherapy.com/how-snowflake-recruited-zeotaps-composable-cdp/), the pull was toward wherever the data already lives. These launches push back on that pull. Zeotap is doing both at once, with a Snowflake Native App inside one platform and Open CDP as a way out of all of them. ## **Is a sovereign CDP cheaper?** Noy by default. Zeotap does lead with total cost of ownership, and the metering point is real. Continuous audience computation on a warehouse bill is expensive and hard to predict. But predictable isn't cheaper. The invoice shrinks in one place and a workload appears somewhere else. In [the inherited costs of the composable stack](https://www.martechtherapy.com/the-inherited-costs-of-the-composable-stack/) I wrote that every pricing model hides a workload. This is the next layer down, and the workload is of a different kind. Someone has to keep the Spark cluster healthy and patched. Someone has to decide whether the model you swapped in is good enough to choose which customer gets which offer, and spot when it stops being good enough. And someone has to sit across the table from the auditor. Look, that last one is the part I'd underline. Zeotap describes the benefit as the difference between a control "that can be evidenced to a regulator" and one that can't. I think it's the most honest phrase in either announcement. Evidence doesn't come from a slide. It comes from a person who understands the setup well enough to show it working, and to explain what happens when it isn't. Both vendors also offer to run the whole thing for you, and for most buyers that's probably the sensible choice. But then the vendor holds the operational knowledge again, and what you have is residency plus a contract with a capable partner. That can be enough. Just know that's what you bought. It's the same question I asked about [forward deployed engineers](https://www.martechtherapy.com/what-is-a-forward-deployed-engineer-and-who-is-paying-for-the-one-in-your-team/): when the vendor's people step back, does the capability stay with you? ## Sign up for Martech Therapy Subscribe Email sent! Check your inbox to complete your signup. New articles straight to your inbox when they're published. ## The one layer you can't swap What strikes me about both launches is how much they promise you can replace. Zeotap frames openness as its retention story, arguing that customers stay precisely because every layer can be swapped out. Meiro says you can change the model in most configurations without rebuilding the stack. The people are missing from that list. If sovereignty moves accountability inside the building, then platform engineering, model evaluation and the ability to explain all of it to compliance become running costs you carry every year. They're also where I'd put the investment. Vendors and models are getting easier to replace. Someone who understands why your identity graph is configured the way it is, and can prove it to a regulator, isn't. In readiness terms this is a capability question before it's a technology one. Buy the architecture without those skills and you'll probably end up on the managed option, paying for both. ## **Are sovereign CDPs a niche, or where the market is heading?** This is where I'm genuinely unsure. Both vendors aim at the same buyer: Zeotap names telco, retail and financial services, Meiro names banking, insurance and telco. Few buyers, large contracts, high event volumes and regulators who ask awkward questions. If it stays there, *sovereign CDPs become a well-paid niche*, much like on-premise deployments always were (Meiro has offered those for years, so this is an extension for them, not a pivot). There's another way to look at it. Meiro's drivers aren't European. They point to the UK Treasury's AI adoption plan for financial services, Australia's prudential regulator and Saudi in-Kingdom residency rules. Zeotap names DORA and the CLOUD Act. Jurisdiction keeps coming up well outside Brussels. And agentic AI turns "where does the model run" into a question every CDP will be asked eventually, since nearly all of them now ship agents that read customer profiles. My guess is both, in sequence. Regulated sectors pay for it first. Whether it spreads depends less on the vendors than on whether a retailer or a travel brand, with no regulator asking, starts putting the question into an RFP. There's a vendor side to this too. For an independent CDP, deployment flexibility is one of the few things Databricks and Snowflake can't easily copy, since their business depends on your data living in their platform. So part of this is a market signal and part of it is the vendors hedging. That doesn't make it less real. It does mean the marketing will run ahead of the demand for a while. So I'll ask, because I don't know. If you work outside banking, insurance or telco, has anyone in your organization asked where the model runs? Or is it still a question for someone else's industry? Do you have any questions after reading this article? Or need support with your Martech projects? [Contact me today >> ](https://www.linkedin.com/in/matthewniederberger/?ref=martechtherapy.com)