Sovereign Data: The Decisive Factor Beyond AI Model Selection

In discussions about AI model selection, people tend to focus on benchmark scores, inference speed, or parameter counts. But Factory co-founder and CTO EnoReyes raised a more fundamental question in his keynote at Fireworks Forge: who owns the data behind the model may matter just as much as which model you choose. This view pushes "sovereign data" to the core of the future of open intelligence.

Sovereign Data: The Overlooked Competitive Dimension

Reyes's core argument is that an enterprise's dependence on a model is essentially a dependence on data. When a model calls an external API, the input data, the corpus used for fine-tuning, and the feedback accumulated from model outputs may all flow into third-party hands. This ambiguity over data ownership is becoming a hidden risk in enterprise AI strategy, against the backdrop of tightening regulation and rising demand for protecting trade secrets.

"Sovereign data" emphasizes that organizations have control over their own data assets—including where data is stored, access permissions, scope of use, and ownership of derived data. In Reyes's view, this is not just a compliance issue but a precondition for whether open intelligence can truly serve an enterprise's long-term interests.

According to Reyes's remarks in the keynote, Factory—a company focused on AI engineering—regards data sovereignty as the cornerstone of building an open intelligence ecosystem. This stance contrasts with the current mainstream cloud vendors' "model-as-a-service" model: while the latter lowers the barrier to use, it quietly transfers control over data.

The term "sovereign data" borrows the concept of "sovereignty" from political science, emphasizing that control over data assets belongs to the organization that produces or holds them, not to the service provider hosting them. Closely related regulatory context includes the right to data portability under the EU's General Data Protection Regulation (GDPR), as well as the requirements on cloud service switching and data migration in the EU Data Act. In AI scenarios, the boundaries of sovereign data are more complex than traditional data governance: beyond raw inputs, fine-tuning corpora, prompts, feedback signals from model outputs, and even the resulting embedding vectors may all constitute derived data assets with commercial value. Therefore, determining whether data sovereignty is complete requires examining data flows across the entire chain of "collection—training—inference—feedback," rather than merely looking at which region the data is stored in.

The Symbiotic Relationship Between Open Intelligence and Data Sovereignty

Open intelligence is often understood as open-sourcing model weights or opening APIs, but Reyes's perspective goes further—true openness should include data portability and auditability. If the model is open but the data pipelines on which training and inference depend are locked in, then openness is only superficial.

This view resonates in the developer community. As enterprises deepen their AI applications, more and more teams are finding that the cost of switching models is far lower than the cost of migrating data. When data formats, annotation systems, and feedback loops are deeply tied to a specific platform, so-called "multi-model strategies" often exist in name only. Sovereign data thus becomes the underlying safeguard for maintaining flexibility in technology choices.

Reyes will further elaborate on this framework in his keynote at Fireworks Forge, scheduled to be livestreamed on November 3. Based on the information disclosed so far, the keynote will center on "why sovereign data is the future of open intelligence" rather than staying at the conceptual level.

"Model-as-a-Service" is the current mainstream form of AI delivery: vendors host model weights and inference infrastructure, and users pay per call via API. The cost is that data pipelines become deeply coupled with the platform, creating what is commonly called "vendor lock-in." It is worth noting that open-source model weights do not automatically equal data freedom—weights can be downloaded, but if the fine-tuning pipelines, evaluation sets, and feedback data built around the model cannot be exported, migration costs remain high. Industry responses to this include adopting open data formats and annotation standards, retaining exportable feedback logs, and decoupling the inference layer from the data layer at the architectural level. Together, these practices form the technical foundation of so-called "portability" and are a key indicator for assessing whether open intelligence lives up to its name.

Implications for Enterprise AI Strategy

For technology decision-makers, Reyes's reminder has practical value. When evaluating AI solutions, in addition to model capabilities, at least three data dimensions should be given equal attention: whether training data can be managed independently by the enterprise, whether inference data is retained in a controllable environment, and whether ownership of the feedback data on which model iteration depends is clear.

This perspective also explains why some enterprises are beginning to shift to private deployment or hybrid architectures—not as a compromise on performance, but as a deliberate choice for data sovereignty. As model capabilities gradually converge, control over data becomes the key to differentiated competition.

Reyes's keynote preview did not offer specific technical solutions or product details, but the question he raised is sharp enough in itself: while chasing stronger models, are enterprises handing over more valuable data assets? The answer to this question may determine the long-term success or failure of AI projects more than any model leaderboard.

Keynote info: Fireworks Forge, livestreamed on November 3, speaker EnoReyes (co-founder and CTO of Factory).