The bottler's dilemma
Episode 6 — SAP's €500m Dremio acquisition read through the Coca-Cola bottler bargain: the platform now owns more of the layer beneath the partner engines,
10x Signal: Data & AI Trends — Episode 6
On 6 July 2026, SAP closed its acquisition of Dremio.
The press release was short. The strategic message was not.
SAP had spent roughly €500 million in cash to buy an Iceberg-native lakehouse, a federated query engine and an open catalog designed to sit across SAP and non-SAP data.
In other words, SAP did not merely add another bottle to the shelf.
It bought part of the bottling plant.
Five episodes ago, I argued that SAP was running Coca-Cola’s playbook: keep the syrup, own the formula and let an ecosystem take the product the final mile.
Since then, this newsletter has followed the control points.
The $1 Trillion Handshake showed the wall around SAP data beginning to come down—zero copy, semantics intact, partners brought closer to the source. Episode 4 asked where the new toll booth would be built. In July, I argued that the near-$100 billion valuation gap between Databricks and Snowflake was not mainly a technology story. It was a position story: one had been provisioned into SAP’s architecture; the other had been connected to it. Episode 5 argued that Clean Core does not remove lock-in. It relocates it—from custom code to business context.
Underneath all five episodes sat one unanswered question:
What happens when the platform decides that the partner’s layer is too important to remain outside?
Bernhard Frieß asked the practical version under the July piece: how does SAP’s Dremio acquisition affect the Databricks partnership, and how long does Databricks keep its premium position?
I did not have a good answer then.
The facts have since arrived.
The 90-second truth
The event: SAP announced the Dremio acquisition on 4 May 2026, closed it on 6 July and disclosed cash consideration of approximately €500 million.
The asset: Dremio brings an Iceberg-native lakehouse, federated access to distributed data and an open catalog built on Apache Polaris. SAP says that catalog will become a discovery and semantic layer for Business Data Cloud and a foundation for the SAP Knowledge Graph.
The non-event: SAP Databricks has not disappeared. SAP still presents it as a native, SAP-managed component of Business Data Cloud for data engineering, data science, AI and machine learning.
The shift: SAP now owns more of the layer underneath the partner engines—the place where data is found, governed, related and given business meaning.
The remaining gap: semantic context is not operational intent. A catalog can describe the business; it cannot automatically define what an agent may decide, under whose authority and within which limits.
The lesson: platform partnerships are not territories. They are negotiated positions inside a product boundary the platform owner can redraw.
The partner question: do not ask only how to get embedded. Ask what remains uniquely yours after the platform absorbs what has become strategic.
That is the bottler’s dilemma.
The $1 contract that built more than 1,000 plants
The Coca-Cola analogy still works. But it works for a more interesting reason than syrup.
In 1899, Benjamin Thomas and Joseph Whitehead secured the exclusive rights to bottle and sell Coca-Cola across most of the United States. The often-told price was one dollar. One bottler’s history says Asa Candler never collected it.
The valuable part was not the dollar. It was the territory.
Thomas and Whitehead could not finance a national production and distribution network alone, so they divided the map and brought in local operators. Those operators built plants, bought equipment, hired people, acquired trucks and developed local customers.
Within twenty years, the network had grown from two plants to more than 1,000. Coca-Cola says 95% were locally owned and operated.
That capital arrived because the bargain was unusually clear:
Coca-Cola owned the formula. The bottler owned the right to develop a defined market.
Some later US bottling agreements were perpetual, although perpetual did not mean untouchable. They could still be terminated for default and other defined events. Outside the US, rights often had expiry dates. But the economic point remains: the rights were durable enough, exclusive enough and specific enough for bottlers to invest as owners rather than behave as temporary resellers.
SAP partners receive no equivalent deed.
They may receive APIs, certification, provisioning, marketplace access, joint selling, product integration, account introductions and a place in the architecture diagram.
They do not receive a perpetual territory around a capability.
No partner owns “enterprise AI for SAP data.”
No partner owns “the lakehouse beside SAP.”
No partner owns “the semantic layer above the application estate.”
They occupy those positions while their technology, commercial model and strategic usefulness justify the space.
That is not betrayal. It is how platforms behave.
But it is a very different investment proposition from Chattanooga in 1899.
Dremio does not kill Databricks. It changes the boundary.
The easy headline would be: SAP partnered with Databricks, then bought Dremio to replace it.
The easy headline is wrong.
SAP Databricks remains natively embedded in Business Data Cloud. SAP describes it as the pro-code environment for data engineering, data science, AI and machine learning, with bidirectional sharing of SAP data products through Delta Sharing.
Dremio enters from a different angle. SAP’s own acquisition announcement places it at the foundation:
an Apache Iceberg-native lakehouse;
federated analytical reach across SAP and non-SAP sources;
a universal catalog built on Apache Polaris;
discovery, lineage, permissions and semantic context;
a foundation for the SAP Knowledge Graph.
Those are not minor capabilities. They determine how the rest of the estate is found, understood and accessed.
So my answer to Bernhard is this:
Dremio does not end the Databricks partnership. It reduces the amount of strategic ground SAP needs Databricks—or any other partner—to own.
Databricks can remain extremely valuable as the preferred environment for advanced engineering and AI workloads. It has customer demand, developer gravity and capabilities SAP would be expensive and slow to recreate.
But its premium position is no longer protected merely by being closest to SAP data. SAP is moving the catalog, federation and context layer closer to itself.
Databricks keeps a privileged room.
SAP increasingly owns the building.
Snowflake should read the same signal. Its position has strengthened since the earlier chapters of this story: SAP now offers both SAP Snowflake for an SAP-led motion and Business Data Cloud Connect for customers with an existing Snowflake estate. That is meaningful progress.
It is still not a territory.
Partners may win the workload, the customer and the quarter. SAP retains the right to redefine the architecture.
SAP can buy the catalog. It cannot buy your intent.
This is where Tobias Unger’s recent sapperment article, The Limits of Language and the Power of Context, adds an important distinction.
Tobias argues that enterprise AI systems are very good at producing coherent form. Humans then project meaning onto that form. The danger begins when fluency is mistaken for understanding and the system is given authority to act.
The enterprise response has been to add context: business data, semantics, process documentation and knowledge graphs. That is necessary. It is not sufficient.
Tobias separates context into five layers:
data;
semantics and knowledge;
process;
operational governance;
intent and prescription.
Most current investment concentrates on the first three. The fifth is the difficult one. It defines what correct execution means here, who may authorise it, which exceptions are acceptable and when the machine must stop and ask.
That distinction matters for Dremio.
SAP can own the catalog that tells an agent how a purchase order relates to a supplier, plant, material and contract. It can preserve lineage, permissions and business semantics across the estate.
The catalog cannot infer whether a 7% price variance should be accepted for this supplier, whether a shipment may be retendered after the second rejection, or whether an agent may make a customer commitment without human approval.
That is not a data problem.
It is a definition-of-work problem.
Tobias calls the machine-executable units of that definition Process Atoms: context-bound behavioural rules with an activation condition, a target outcome, a defined scope, an authority source and a lifecycle record. Their aggregate becomes what he calls Company Memory—not only a record of what the organisation has done, but a governed account of how work should be done.
This points to a more defensible partner position.
Do not compete only for the infrastructure layer the platform can absorb. Help the customer turn industry knowledge, operating policy and judgment into governed execution context at a resolution agents can use.
But there is a condition: the customer must be able to own and export that context. Otherwise, the partner has not solved the rented-land problem. It has recreated it one level higher—in proprietary prompts, agent logic and hidden rules.
SAP may increasingly own the building.
The enterprise still needs to own the operating manual.
The platform does not need to copy everything
There is a temptation to turn this into a gloomy rule:
Every successful partner will eventually be absorbed or replaced.
That is too simple.
Platforms do not internalise everything that succeeds. They internalise what becomes strategically central, economically attractive and feasible to own.
The real danger zone has three characteristics:
The capability becomes part of the control plane. It determines access, identity, governance, semantics, orchestration or commercial ownership.
Customers begin to expect it as part of the platform. What was once an extension becomes table stakes.
The dependency weakens the platform owner’s economics or roadmap control. Too much margin, data or customer influence sits outside.
Dremio touches all three. It gives SAP more control over the open data foundation, supports a product promise customers increasingly expect, and reduces dependence on any single external engine for access to distributed enterprise data.
This does not make partners irrelevant.
It changes where relevance must live.
The safest partner position is not “SAP cannot build this.” Large platforms can build many things.
The safer position is:
SAP could build this, but doing so would take too long, distract from its core, weaken customer choice or still fail to reproduce the depth we have accumulated.
That depth can come from industry data, workflow IP, specialist talent, a community, a network, proprietary distribution or trust earned in a narrow problem over many years.
Integration is useful.
Indispensability is better.
The episode I owe you
At the end of Episode 2, I promised to explain how these alliances work in practice: joint go-to-market, field enablement and the executive relationship.
Episodes 3, 4 and 5 went up the stack instead—to control points, valuations and semantics. That was the more interesting altitude.
It was also the easier one.
The operator’s version is less glamorous.
Two CEOs announce a strategic alliance. Two logos appear on a slide. A press release uses the word “transformative.” Someone takes a photograph.
By Monday morning, a sales representative still has eleven other things to sell.
That is where the strategy either becomes revenue or returns to PowerPoint.
- The field-memory test
If the SAP seller cannot explain the partner’s role in ninety seconds, the alliance does not exist in the field.
Not the architecture. Not the corporate vision. The role:
What customer problem does it solve?
Why does it belong in this deal?
Who creates the opportunity?
Who contracts it?
Who receives quota credit?
What happens next?
Enablement is not a launch event. It is a standing operating cost. Product names change, territories move, account teams rotate and new priorities arrive every quarter.
The first thing forgotten is yesterday’s strategic partnership.
- The compensation-gravity test
Revenue share can look elegant in a contract and die quietly in a compensation plan.
If the extension pays the SAP seller less than the native alternative, it will lose.
If selling it makes the deal larger but the sales cycle longer, it will lose in the final month of the quarter.
If the partner’s own field earns more by selling around SAP than through SAP, it will do exactly that.
People do not execute the press release.
They execute the incentive.
- The 2 a.m. test
Every alliance works in a demo.
The test begins when a critical process fails at two in the morning and the customer opens one incident against an architecture owned by three companies.
Who owns first response?
Who reproduces the issue?
When does partner engineering join?
Who tells the customer’s board what happened?
I spent years on the SAP side of exactly these nights. The alliances that survived had drawn the escalation map before they needed it. The ones that failed discovered the missing map while the customer was watching.
“Seamless” is a marketing word.
An agreed severity-one process is an operating model.
- The empty-chair test
Every alliance begins with executive sponsorship.
Then one sponsor changes role.
The photograph remains. The calendar invitations stop.
If the partnership does not have a joint pipeline review, product governance, field cadence, escalation path and named successors, sponsor turnover reveals what was really built.
Usually, not much.
Alliances rarely die in strategy decks.
They die in calendars, compensation plans, support queues and organisation charts.
That is why Dremio matters. When the platform boundary can move, operational relevance must be renewed every quarter.
Three readers, one control point
Maciej Trylowski asked whether SAP customers really want one control plane or are comfortable with two.
The answer is that most customers do not wake up wanting a control plane. They want fewer copies, clear accountability, governed access and an architecture that does not punish them three years later.
Two planes can work. Many large enterprises will continue to run them. But the seams must be deliberate: which catalog is authoritative, where policy is enforced, how lineage crosses the boundary, who pays for duplicate compute and what happens when the roadmaps diverge.
After Dremio, this is no longer only a preference question. SAP now owns a stronger answer of its own. Customers should ask where investment is going—and ask for roadmap commitments in writing, not in a keynote.
Philip Zhang asked what an AI connects to once SE80 is locked.
It connects to whatever SAP exposes and governs as the interface: APIs, events, data products, knowledge graphs and approved extension services.
That solves one class of problem and creates another.
If your agents, rules and exception logic live entirely inside someone else’s platform, you have not eliminated lock-in. You have changed landlords.
This is why Clean Core requires clean context: decision rules documented outside individual agents, knowledge exportable, lineage visible and outcomes traceable. Otherwise, the custom code disappears and the dependency remains.
Andreas Reuther asked what the core competence of a Solution Extension partner should be.
My answer was: engineer the leverage, or commercialise it at scale. Trying to do both rarely compounds.
I would sharpen that now.
Engineering the leverage means owning the difficult part: domain depth, proprietary workflow, trusted data, specialist capability or a network effect that the platform cannot reproduce cheaply.
Commercialising at scale means converting platform access into distribution before the window moves.
Both can be good businesses.
Neither creates a permanent territory!
How to survive the moving boundary
If you are a customer: run four tests before approving the architecture.
Ask SAP and the partner separately who owns a severity-one incident. The answers should match.
Ask which component is strategic to each roadmap three years from now. “We are committed” is not a roadmap.
Ask what can be exported—data, metadata, rules, prompts, lineage and business logic—if the commercial relationship changes.
Ask where duplicate governance, compute or licensing can appear. Zero copy does not automatically mean zero duplication of cost or control.
If you are a partner: assume the boundary will move.
Do not build your moat from proximity to SAP alone. Build it from what proximity allows you to learn, own and compound.
Own a problem narrow enough to understand deeply and valuable enough that customers insist you remain. Keep your IP portable. Build direct customer trust. Make the joint motion easy for the field. And keep moving up the stack before the layer beneath you becomes standard.
Your objective is not to become impossible to copy.
It is to become slower, riskier and more expensive to replace than to keep.
If you are SAP: make the ecosystem bargain explicit.
Partners understand that product boundaries change. What they cannot price rationally is surprise. Clear roadmaps, stable interfaces, fair migration paths, workable field incentives and honest signals about which layers SAP intends to own will attract better partners and better investment.
Dremio makes the ecosystem stronger if it creates a more open, capable foundation.
It makes the ecosystem weaker if partners conclude that every successful adjacency is merely SAP’s future backlog.
That perception will be priced in—through lower investment, shorter horizons and more defensive architectures—whether anyone says it out loud or not.
The real bargain
Coca-Cola’s early bottlers poured concrete because the territory was clear.
SAP’s partners pour engineering, capital and field capacity into a boundary that can move.
That does not make the model unfair. It makes the bargain different.
The partner does not receive permanence.
It receives access, distribution and time.
What it builds during that time determines whether it becomes a feature, an acquisition, a durable specialist—or a logo removed from the next version of the slide.
Strategy gets the photograph.
Execution gets the customer.
Only indispensability buys time.
Next episode: the customer’s seat. How to evaluate a solution when the vendor, the extension and the implementer are three companies with three different incentives—and none of them owns the outcome.
Sources
SAP News — SAP to Acquire Dremio to Unify SAP and Non-SAP Data to Power Agentic AI, 4 May 2026.
SAP News — SAP Completes Acquisition of Dremio, 6 July 2026.
SAP Q2 2026 disclosure filed with the SEC — Dremio acquisition and approximately €0.5 billion cash consideration.
SAP — SAP Databricks in SAP Business Data Cloud.
SAP — SAP Snowflake.
Tobias Unger — The Limits of Language and the Power of Context, sapperment.
The Coca-Cola Company — The Asa Candler Era: Bottling Begins.
Coca-Cola UNITED — Savannah bottling history and the 1899 one-dollar rights story.
Coca-Cola Enterprises, Form 10-K — North American bottling rights and termination provisions.
10x Signal Episodes 1–5, and reader comments by Bernhard Frieß, Maciej Trylowski, Philip Zhang and Andreas Reuther.
First published on LinkedIn. Subscribe to 10x Signal: Data & AI Trends to get new editions as they appear — or browse the sapperment library.