In Episode 175 of the AI and Data Weekly podcast, Anders Hammarbäck , Co-Founder and CEO of RedpineAI, put forward a thesis that cuts against the dominant narrative of the past three years. As models race toward commoditization, he argues, the next defensible advantage in AI will not be bigger parameter counts or marginal benchmark gains. It will be proprietary, licensed, high-quality data and the ability to turn it into usable knowledge at the point of inference.
The podcast conversation between Anders, Henrik Göthberg and Anders Arpteg ranged widely, from hackathons in Stockholm to healthcare data infrastructure, from European regulation to agentic commerce. But at its core is a simple, powerful idea: if everyone builds on the same foundation models trained on similar internet-scale corpora, differentiation must come from somewhere else.
That “somewhere else” is data.
Hammarbäck’s view is shaped by experience. Before founding RedpineAI, he spent years as a venture capital investor evaluating companies across the AI stack. He saw hundreds of startups deploying similar base models, often trained on scraped web data, competing primarily on user interface and go-to-market execution. The underlying intelligence was, in many cases, shared.
“The common challenge I saw in so many of them,” he explains, was that they were “using the same or very similar data.”
That observation became the seed of a company and, more broadly, a perspective on where AI’s true competitive advantage is shifting.
This article explores that thesis, situates it within current industry developments, and examines what it means for enterprises, AI labs, and policymakers navigating the next wave of AI transformation.
From Model-Centric to Data-Centric AI
For the better part of a decade, the AI industry has been model-centric. Successive waves of innovation like deep learning, transformers, large language models, have been defined by architectural breakthroughs and scale. Larger datasets, more compute, and more parameters drove dramatic capability gains.
The release of GPT-3 in 2020 and the public launch of ChatGPT in late 2022 marked a turning point. Generative AI went mainstream. Venture funding surged. Enterprises rushed to pilot applications. Foundation models became platforms.
Yet as the ecosystem matured, a pattern emerged. Most application-layer companies were building on top of a small set of base models from major labs such as OpenAI and Anthropic. While there were differences in fine-tuning and orchestration, the underlying training data was often similar: vast swathes of public web data, books, code repositories, and licensed corpora.
At the same time, legal challenges began to surface. High-profile lawsuits, including the case brought by The New York Times against OpenAI, raised fundamental questions about copyright, fair use, and compensation for content creators. Meanwhile, settlements such as Anthropic’s agreement over the use of copyrighted books underscored the financial and legal risks embedded in large-scale pre-training strategies.
The era of “scrape first, settle later” is giving way to a more structured, licensed data economy.
In parallel, the performance delta between leading models has narrowed. While new releases still generate excitement, improvements are increasingly incremental in many enterprise contexts. As inference costs fall and open-weight models improve, the base model itself risks becoming infrastructure-critical, but not sufficient for competitive differentiation.
This is the backdrop against which Hammarbäck’s argument lands.
If models become commodities, the moat must shift up or down the stack.
..if everyone builds on the same foundation models trained on similar internet-scale corpora, differentiation must come from somewhere else. That “somewhere else” is Data.
The Untapped Frontier: Proprietary Data
When Hammarbäck talks about proprietary data, he does not mean merely internal corporate databases. He uses the term more broadly to describe non-public data or information that exists outside the scraped internet and is often underutilized for AI.
Public internet data may amount to tens of trillions of tokens used for pre-training, but beyond that lies an order of magnitude more information: clinical trial results, industrial sensor logs, insurance claims, academic research archives, enterprise transaction histories, regulatory filings, and more.
Much of this data is fragmented, siloed, and locked behind licensing constraints. It is often high-quality, domain-specific, and accuracy-critical. In fields such as healthcare, finance, and industrial manufacturing, the difference between a general answer and a grounded, context-aware answer can be material.
RedpineAI’s mission is to unlock that data for AI systems. Not only by licensing it, but by structuring and transforming it into what the company calls a “knowledge layer.”
The key insight is that raw data alone is not enough. What matters is context, structure, provenance, and fit-for-purpose retrieval.
This shift aligns with broader trends in enterprise AI. According to McKinsey’s 2024 State of AI report, companies that report the highest ROI from AI initiatives are those integrating domain-specific data into their systems, rather than relying solely on generic models. Similarly, Gartner projects that by 2027, more than 50 percent of generative AI models used by enterprises will be domain-specialized.
The question is not whether AI will be used. It is what it will be grounded in.
From Training to Inference: Where Value Moves
In the early phase of generative AI, most attention focused on pre-training and fine-tuning. Labs competed to build ever-larger foundation models. Data partnerships were structured primarily around training rights.
RedpineAI began in that world, working with AI labs to secure and license datasets for model development. But as the market evolved, so did its focus.
Today, the emphasis is increasingly on inference, meaning what happens when an AI system is deployed in real-world workflows.
Agentic AI systems, which can plan, retrieve information, call tools, and execute multi-step tasks, are becoming central to enterprise strategy. These systems do not merely generate text; they act.
At the point of inference, context becomes critical. An agent working in drug discovery, for example, needs access not only to general medical knowledge but to specific, licensed clinical studies, regulatory guidelines, and up-to-date research findings. The quality of its output depends on the relevance and credibility of the data it can access in real time.
This is where the “knowledge layer” concept becomes powerful.
Rather than simply exposing large document repositories via an API, RedpineAI breaks data down into smaller, structured units, enriched with metadata and ontological relationships. It evaluates quality, tracks provenance, and optimizes retrieval so that the right data is surfaced for the right query in the right context.
In practice, this may mean that a healthcare agent querying for treatment pathways in oncology is not simply retrieving generic summaries, but licensed clinical studies, anonymized patient histories, and region-specific guidelines curated and structured to reduce hallucination risk.
This approach aligns with the industry’s broader pivot toward retrieval-augmented generation and tool-augmented agents. As systems move from static chatbots to autonomous actors, the integrity of their data sources becomes a first-order concern.
The moat, then, is not just owning data. It is owning the pipeline that turns data into actionable knowledge at scale.
Healthcare as a Proving Ground
Healthcare is the first vertical RedpineAI is targeting, and for good reason.
The sector is data-rich, highly regulated, and deeply accuracy-sensitive. Errors carry real consequences. At the same time, the potential for AI-driven impact is enormous: from accelerating drug discovery to optimizing treatment pathways and reducing administrative burden.
Industry analysts estimate that generative AI could unlock between $60 billion and $110 billion annually in productivity gains across the healthcare sector globally. Yet these gains depend heavily on data quality, interoperability, and compliance.
One example discussed in the podcast is a partnership with a Swiss biochemical company working in toxicology and drug discovery. By integrating licensed clinical studies and other external datasets via RedpineAI’s API, the company can enrich its internal models and visualization tools. The goal is not to replace domain expertise, but to augment it with broader, structured context.
This pattern of combining internal proprietary data with licensed external datasets may become the dominant architecture for enterprise AI.
It also illustrates a broader principle: in regulated industries, legally sourced, properly licensed data is not just a technical requirement. It is a strategic asset.
The Business Model of the Data Economy
One of the more intriguing aspects of RedpineAI’s approach is its economic model.
Data owners, whether research institutions, publishers, or enterprises, are compensated based on usage. As AI systems query and retrieve licensed data at inference time, revenue flows back to the source.
This creates an incentive structure that differs from one-time data sales or opaque scraping practices. It begins to resemble the evolution of digital music. Platforms such as Spotify demonstrated that licensing content and distributing royalties at scale could create sustainable ecosystems.
In payments, Stripe simplified integration with a developer-friendly API, lowering friction and accelerating adoption. RedpineAI explicitly references this model: ease of integration is critical for winning the AI developer community.
If AI agents become economic actors making purchases, booking travel, and procuring data, then the infrastructure enabling secure, transparent transactions becomes foundational.
In that future, proprietary data is not merely an input. It is a tradable asset in an emerging agentic economy.
Europe’s Moment: Regulation as Opportunity
The discussion in Episode 175 also touches on Europe’s position in the global AI race. While the continent often lags the United States and China in model development and venture funding, it holds structural advantages in regulation and data governance.
Initiatives such as the European Health Data Space and the EU Data Act aim to facilitate controlled, harmonised data sharing across member states. The EU AI Act, though controversial in its implementation details, seeks transparency around training data and risk classification.
Hammarbäck takes a pragmatic view. Regulation is necessary, particularly around privacy. But legal uncertainty and fragmented standards can slow innovation. The opportunity lies in harmonization: reducing friction across markets while preserving core protections.
If Europe can operationalize data spaces effectively and move beyond philosophical commitments to practical interoperability, it could carve out a defensible position as the world’s leader in trusted, regulated AI.
In this framing, proprietary data is not antithetical to openness. It is the foundation for fair, licensed, and privacy-preserving AI systems.
The Enterprise Blind Spot
Despite the strategic importance of data, many enterprises remain overly focused on model selection rather than data readiness.
Executives often ask: Which model should we use? Should we build or buy? How do we fine-tune?
Less frequently do they ask: What is the state of our data? How well is it structured? What external datasets could materially improve our outcomes? Do we have the ontologies, metadata, and governance frameworks to turn raw information into knowledge?
The risk is clear. If base models continue to improve and become cheaper, then investments concentrated solely on model experimentation will erode in value. Meanwhile, competitors who invest in data curation, licensing, and knowledge graph construction may build compounding advantages.
In this sense, the data strategy becomes the AI strategy.
Enterprises should consider three dimensions:
- First, internal data maturity. Are decision traces, process logs, and contextual metadata being captured and structured? Much of the knowledge embedded in organizations is tacit. Converting it into machine-readable form is a long-term moat.
- Second, external data partnerships. What non-public datasets could unlock new products or dramatically improve performance? Are there licensing models that align incentives across ecosystems?
- Third, integration architecture. How easily can AI agents access relevant data at inference time? Is the system designed for dynamic retrieval, or is it limited to static embeddings?
The winners of the next AI wave may not be those with the largest models, but those with the richest, best-structured data ecosystems.
From Data to Knowledge: The Ontology Layer
A recurring theme in the podcast is the distinction between data and knowledge.
Data, in isolation, is inert. Knowledge emerges when data points are connected, contextualized, and interpreted. In AI systems, this often means building ontologies and knowledge graphs that encode relationships between entities.
For example, in healthcare, linking patient histories, clinical guidelines, research publications, and insurance records requires more than simple retrieval. It requires mapping concepts across domains, resolving ambiguities, and validating sources.
RedpineAI’s “knowledge layer” attempts to address this by decomposing documents into smaller units, tagging them with metadata, and evaluating their relevance and ground-truth alignment.
This approach reflects a broader industry realization: scaling models without scaling structure leads to hallucinations and brittleness. Structure acts as scaffolding for reliable reasoning.
As AI systems move closer to high-stakes decision-making, the demand for scientifically grounded, explainable outputs will only grow.
The Labor Market and the Rise of Agentic Systems
One of the more forward-looking threads in the conversation concerns agentic AI and the future of work.
If agents can autonomously query data, make purchasing decisions, and execute workflows within defined budgets, the nature of digital labor changes. We already see early examples in code generation, marketing automation, and travel booking. Payment providers are experimenting with agent-specific protocols.
In such a world, data access becomes transactional and dynamic. Agents may select between competing datasets based on quality, relevance, and cost.
This raises strategic questions for organizations: how do you ensure your agents are grounded in trustworthy data? How do you prevent over-reliance on generic sources? How do you manage budget constraints and compliance?
The emergence of agentic systems amplifies the importance of proprietary, high-quality data. Autonomous actors need guardrails, and those guardrails are often encoded in data governance and retrieval logic.
A Broader Implication: AGI and the Knowledge Frontier
While the podcast does not dwell extensively on artificial general intelligence, the implications are implicit.
If progress toward more general capabilities continues, the limiting factor may not be architecture but grounding. General intelligence without domain-specific knowledge risks being shallow. Deep expertise requires access to curated, structured, evolving data.
In that sense, proprietary data is not merely a business moat. It is a civilizational resource. How societies structure, license, and govern their knowledge assets will shape the trajectory of AI development.
The next frontier may not be scaling parameters from billions to trillions, but scaling the integration of real-world knowledge into AI systems responsibly.
What This Means in Practice: Five Imperatives for Organizations Building AI
If proprietary data and structured knowledge become the real moat in the AI era, then organizations must shift their focus accordingly. This is not a philosophical shift. It is operational.
First, treat data strategy as core business strategy, not an IT initiative. Too many AI programs begin with model selection discussions. Which model should we use? Should we fine-tune or use RAG? Should we build our own? These are important questions, but they are downstream decisions. The upstream question is far more fundamental: what unique data assets do we have, and how can they be structured to create defensible advantage?
Organizations that win in AI will be those that identify their most valuable data domains, clean them, contextualize them, and make them usable for intelligent systems. That requires investment in metadata, ontologies, decision trace capture, and governance frameworks. It is less glamorous than deploying a chatbot, but infinitely more durable.
Second, move from document storage to knowledge structuring. Many enterprises believe they are data-rich because they have vast repositories of documents. In reality, much of that information is unstructured, unlinked, and unusable for agentic systems. The real leap comes when organizations break information down into smaller, meaningful components, connect them through explicit relationships, and encode context around them.
When AI systems operate in high-stakes domains such as healthcare, finance, or industrial engineering, surface-level retrieval is not enough. They must be able to trace sources, understand interdependencies, and retrieve contextually relevant data at inference time. Building this internal knowledge layer is not optional if trust and reliability matter.
Third, invest early in licensing and external data partnerships. Internal data alone rarely provides full competitive advantage. The most powerful AI systems combine internal proprietary knowledge with carefully selected external datasets. This might include industry research, regulatory databases, anonymized benchmark data, or domain-specific corpora that are not publicly available.
Organizations that proactively secure licensing agreements and build relationships with data providers will have a structural advantage as legal scrutiny around training data intensifies. Waiting until regulation forces compliance is a reactive strategy. Building a transparent, licensed data ecosystem is a proactive moat.
Fourth, design AI systems for inference-time intelligence, not just training-time optimization. Much of the early AI focus has been on model training and fine-tuning. Increasingly, however, value creation happens at inference. Agentic systems retrieve information dynamically, make decisions within budget constraints, and operate inside real workflows.
That means your architecture must support secure, real-time data retrieval with quality evaluation and provenance tracking. It also means designing governance models that define what agents are allowed to access, under which conditions, and with what economic boundaries. In a world where agents can transact, procure, and execute tasks autonomously, data access becomes a controlled economic interface.
Fifth, prepare for regulatory maturity rather than regulatory avoidance. Especially in Europe, organizations often frame regulation as a constraint. In reality, clarity around data rights, transparency requirements, and risk classification can become a competitive differentiator. Enterprises that build privacy-aware, explainable, licensed AI systems will be better positioned as enforcement strengthens globally.
Rather than asking how little compliance is required, forward-looking organizations ask how trust can become part of the product itself. In domains such as healthcare or finance, that trust layer may ultimately be more valuable than marginal model improvements.
Taken together, these imperatives signal a shift in mindset. The AI race is no longer just about access to the best models. It is about access to the best knowledge, structured in the most usable way, governed responsibly, and deployed intelligently.
The organizations that understand this early will not only deploy AI. They will compound advantage with every new dataset they structure, every ontology they refine, and every inference-time decision they ground in high-quality, licensed information.
You can listen and watch the entire episode here.
This article was enhanced with the help of AI tools, drawing on the podcast transcript and complementary online research. To go deeper into the source material, I encourage you to listen to the full episode and make your own learnings.
