.svg)
Key Takeaways
- Duplicate storage compounds across a hybrid estate, where copies cross environment boundaries and pay egress on the way.
- Zero-copy data sharing lets consumers read data where it lives instead of receiving a copy.
- Forrester puts 35% of enterprise data in duplicate copies, each one adding cost, lag, and risk.
- The test of a zero-copy approach is whether it spans on-premises and every cloud, not one vendor's platform.
- Governance gets harder, not easier, since access must now be enforced at the source.
Zero-copy data sharing exists because most data platform teams cannot answer a simple question: how many copies of the customer table exist right now? Three is a guess. The real answer involves an on-premises warehouse, two cloud accounts, a feature store, and a BI extract somebody built in 2023.
Each copy was justified when it was created, and none has an owner now. The expensive copies are the ones that crossed an environment boundary, which is exactly where the sprawl is hardest to see.
What is Zero-Copy Data Sharing and Why Does it Matter Beyond a Single Lakehouse?
Zero-copy data sharing is a data architecture in which multiple systems or applications access the same underlying data directly, without creating separate copies or moving the data between them.
Within a single platform, sharing without copying is a solved feature: most lakehouse vendors ship a sharing protocol that works between accounts on their own platform. The costs that accumulate sit outside that boundary.
The bottom three rows are where hybrid data sharing earns its keep. A sharing protocol native to one cloud platform does nothing about a copy that had to leave an on-premises system to be useful, which is the copy most likely to be large, regulated, and expensive to keep synchronized.
Why Does Duplicate Storage Quietly Become One of the Largest Costs in a Hybrid Data Estate?
Duplicate storage becomes one of the highest costs in a hybrid estate because every copy is a small, reasonable decision that nothing in the process ever retires, so those small costs compound silently into the biggest line nobody is watching.
The sequence repeats in every large estate:
- A consuming team needs a dataset and cannot get adequate performance or access against the source, so a copy is provisioned.
- The copy needs a pipeline to stay current, which becomes a permanent maintenance obligation.
- The copy drifts because the pipeline lags or a transformation was applied on the way in.
- The original need passes, the team moves on, and nobody deletes anything.
- A second team finds the copy, decides it is close enough, and builds on it.
By year three, the estate holds several versions of the same dataset with different freshness and slightly different numbers. The storage bill is the visible cost and the smallest one.
Reconciling two dashboards that disagree consumes analyst time, and the copies also bind compute to storage in ways that make either one hard to change, which is coupled compute and storage architecture debt that teams pay down for years.
Attribution makes it worse. A duplicate dataset rarely belongs to anyone in the cost model, which is precisely why coupled compute and storage is a FinOps problem before it is a storage problem. Add the transfer charges each cross-environment copy incurs, and cloud egress costs are the hidden tax that never appears on a project's budget line.
The European Commission confirms the EU Data Act removes switching charges, including data egress charges, from 12 January 2027. That applies to leaving a provider. Routine transfer between environments as part of normal operation is a different charge and is not covered, so the egress cost of a synchronization pipeline running every hour remains exactly where it is.
How Does Zero-Copy Sharing Work Across On-Premises and Cloud Environments?
Hybrid zero-copy sharing requires a metadata and access layer that spans every environment, so a consumer in one place can resolve, authorize, and read data sitting in another.
Forrester's report names four capabilities that make the pattern work, and each one has to hold across environments, not just inside one.
1. Decoupled storage and compute
Consumers bring their own engine to data they do not own. This falls apart wherever storage and compute are bound together inside a single vendor's runtime, since bringing an outside engine to that data is not an option at all.
2. Open table formats
Any engine can read the shared tables without a conversion step. Support for this works only as far as the vendor's own catalog is willing to admit outside engines, which is often narrower than the format itself would allow.
3. In-memory sharing
Data passes between processes without a materialized copy ever touching disk. This capability is typically confined to one runtime and one environment, so it disappears the moment a consumer sits outside that boundary.
4. Direct object storage access
Consumers read the underlying objects under governed access rather than through an intermediary service. This rarely extends to on-premises object stores, which is exactly where the hybrid case gets hard.
The pattern has a close relative in integration. Reading data where it lives instead of staging it through an intermediate copy is the same instinct behind zero ETL data integration, and the two share a failure mode: both work cleanly inside one vendor's estate and get difficult at the boundary between two.
Latency is the honest constraint. A consumer reading across a WAN link will not match local-copy performance for every workload, which is why practical hybrid implementations cache selectively instead of promising no data ever moves. The goal is removing the standing duplicate, not eliminating every byte of transfer.
What Does Recent Forrester Research Say About the Shift Away From Copying Data for AI?
Forrester puts the scale of the problem at 35% of enterprise data being duplicated, with every copy adding cost, lag, and risk. That finding comes from the August 2026 Trend Report by Noel Yuhanna, Stop Copying Data: The Zero-Copy Shift Powering Enterprise Agentic AI, which maps the shift toward real-time governed access to data where it already sits.
A third of the estate is a number worth sitting with. It is not the pathological case; it is the baseline, and it means roughly one dollar in three of storage spend is carrying a copy of something the organization already owns.
AI is what turns that baseline from wasteful into unworkable. Analytical consumers were tolerant of stale copies, since a dashboard refreshed nightly is acceptable for most reporting.
Agentic systems act on what they read, and they read continuously from more sources than any dashboard ever did. Multiply the standing copy problem by the number of agents an enterprise plans to run and the arithmetic stops working, which is the shift the report is describing.
What Should Enterprises Evaluate Before Adopting a Zero-Copy Sharing Approach Across a Hybrid Estate?
Enterprises should evaluate coverage, governance consistency, and failure behavior, in that order, because an approach that fails any of the three will produce copies again within a year.
- Does it cover every environment you run? On-premises systems, each cloud, and any sovereign or regional deployment. Partial coverage means the uncovered environments keep copying, and those are usually the expensive ones.
- Is governance defined once and enforced at the source? A rule applied in one environment and reimplemented in another is two rules that will diverge.
- What happens when the source is unavailable? Consumers lose access because there is no local fallback. Decide which workloads can tolerate that before committing them.
- Can you see cross-environment access and cost? Reads that cross a boundary carry a charge, and unattributed charges become somebody's surprise.
- How is latency handled for interactive consumers? Caching policy, and who controls it.
Answering the last two requires instrumentation across environments, which is what hybrid data observability platforms built for real estates provide and single-cloud tooling does not.
See how Acceldata's hybrid data observability platform supports zero-copy access across on-premises and cloud before committing to an architecture that depends on it.
Making Zero-Copy the Default Across the Whole Estate
Zero-copy sharing pays off when it is a property of the whole estate. Implemented inside one platform, it removes the cheapest copies and leaves the expensive ones untouched.
What to hold any approach to:
- Coverage should extend to on-premises and every cloud, including sovereign deployments.
- Governance should be one model, defined once and enforced at the source.
- Cross-environment reads should be visible, with cost attributed to a consumer.
- Source unavailability should have a documented answer, per workload class.
- Caching policy should sit under your control, not the vendor's default.
Forrester's own research suggests the shift is accelerating as AI makes duplicate, ungoverned data untenable, and the enterprises moving first are the ones that can already see what their copies cost.
Acceldata's xLake platform observes data, pipelines, and access across on-premises and cloud environments from one control plane, which is what makes zero-copy access measurable instead of aspirational.
Find out how many copies your estate is really carrying. Book a demo with Acceldata and see the sprawl your own dashboards can't show you.
FAQs: Zero-Copy Data Sharing
Does zero-copy data sharing eliminate the need for data governance across systems?
No, zero-copy data sharing does not eliminate the need for governance because shared data still requires consistent controls for access, security, quality, and compliance. In fact, shared access can make centralized policies and clear ownership more important across systems.
How does zero-copy sharing affect query performance compared to a local copy of the data?
Zero-copy sharing can deliver similar performance to local data when systems have efficient access to the underlying storage, but network latency and remote access overhead can affect performance. Local copies may be faster for frequently accessed workloads, at the cost of additional storage and data synchronization.
Can zero-copy sharing work between an on-premises system and a cloud data platform?
Yes, zero-copy sharing can work between on-premises systems and cloud data platforms when they can securely access compatible shared storage or data-sharing interfaces. Performance and feasibility depend on network connectivity, data format compatibility, security controls, and the capabilities of the platforms involved.
What happens to zero-copy shared data if the source system goes offline?
If the source system or underlying storage goes offline, consumers may lose access to the shared data unless they have cached copies or another replicated access path. Zero-copy sharing reduces duplication, but it does not by itself provide high availability or disaster recovery.
Does zero-copy sharing reduce cloud egress costs, or just storage costs?
Zero-copy sharing primarily reduces storage and data duplication costs, but it can also reduce egress costs when data does not need to be copied across regions or platforms. However, queries that repeatedly move large volumes of data across network boundaries can still incur significant egress charges.
