When foundation models commoditize, the value moves to fine-tuning and evaluation

For most of the last three years, the AI conversation has been around which lab has the best model? Benchmark leaderboards, parameter counts and release cadences set the terms of competition and enterprise buyers treated frontier model access as the core asset worth securing. However, as the landscape shifts, AI model builder providers that compete solely on model benchmarks are fighting yesterday's war.

At our recent World Models Summit, Yann LeCun, professor of computer science and data science at New York University and executive chairman of the AI startup AMI Labs, Turing Award winner and one of the pioneers of modern deep learning, shared a perspective that frontier models are heading toward Linux. Every server in existence runs Linux for all practical purposes. Foundation models are going the same way. "It is inevitable that foundation models are going to become some kind of open-source commodity," he told the room. "You are not going to make money with the model itself. The value will come from how you fine-tune these models for specific applications, languages, cultures and value systems, not from who controls the largest pre-training checkpoint."

As frontier models commoditize and converge, value moves one layer up the stack to the layer where a model is fine-tuned, instruction-tuned, evaluated and red-teamed. Frontier models are increasingly an input, not a finished product. The true differentiator is the quality, compliance and specificity of the data used to adapt them.

The transition towards open-source commodities

A November 2025 paper, The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference, found that the cost of achieving a given benchmark performance has been falling at a median rate of five to 10 times per year for frontier capability and 40 to 900 times per year for capability that has already commoditized. Departmental AI spending has surged dramatically year over year, but it is directed more toward applications, vertical tools and integrations sitting on top of foundation models.

Hugging Face now hosts more than 2.9 million models. Over 30% of the Fortune 500 maintain verified accounts there, building on open-source frameworks rather than relying exclusively on closed, proprietary APIs.

While the absolute frontier continues to push forward on long-horizon reasoning, tool use and agentic capabilities, the "broad middle" of enterprise capabilities such as drafting, summarization, classification, code completion and customer dialogue has converged across multiple proprietary and open-weight models. When the floor rises and the transaction cost collapses, the baseline model itself ceases to be your competitive moat.

Need for world models

LLMs excel in discrete symbolic domains like language, computer code and mathematics, where it’s possible to calculate highly accurate probability distributions over a finite dictionary. However, the real world is continuous, high-dimensional and largely unpredictable at the pixel or token level. Forcing a generative system to reconstruct unpredictable, granular details — like the rustling of leaves or the ripples on a distant pond — actively degrades its structural understanding of a scene. Furthermore, text is an incredibly low-bandwidth medium. True human-level intelligence cannot be back-formed purely from language; it must be grounded in the massive, sensory-rich experience of physical reality.

Contextual intelligence, industry specificity and strict governance

Generic models trained on the open internet do not understand the nuances of your specific customers, your historical contracts or your proprietary product catalogs. Every vertical industry operates within its own distinct lexicon:

• Automotive engineering: Operations revolve around tolerance stacks, validation cycles and revision control.
• Capital markets: Reasoning requires deep context regarding risk-weighted assets, liquidity buffers and complex legal frameworks like the international swaps and derivatives association (ISDA) master agreement.
• Security operations: Critical patterns must be extracted from a chaotic, high-volume sea of telemetry signals and identity anomalies.

Vertical AI solutions understand these specialized contexts, jargon and task-specific subtleties that horizontal models miss. This industry specificity is a massive market reality; vertical AI spend has grown exponentially, producing highly successful specialized tech segments.

Because fields like healthcare, finance and legal are highly regulated, solutions must meet strict data, access and audit requirements. A general-purpose model with strong public benchmarks but no domain-specific evaluation suite remains an unfinished product from a tool adoption and reliability perspective. Training AI securely within these environments requires wrapping these models in strict data governance policies, including end-to-end data encryption, granular access controls and immutable audit logs.

Digital sovereignty, culture and the multilingual ecosystem

The requirement for localized specialization is now extending to encompass global geographic and cultural boundaries, driven by the rise of digital sovereignty. It refers to the capacity of a state or region to independently govern, control and protect its digital infrastructure, data and technologies in alignment with its own laws, values and strategic interests.

Digital sovereignty operates across three essential dimensions:

• Data sovereignty: Ensuring data generated within a territory is subject to local governance, such as the EU's general data protection regulation (GDPR). Open-source structures allow public institutions to fine-tune systems on local datasets under national legal frameworks without data leaving the jurisdiction to ensure legal adherence.
• Technological sovereignty: Pursuing self-reliance to reduce dependence on a handful of foreign tech monopolies. Open source allows academic researchers and local communities to maintain core AI capabilities, fostering regional talent and expertise.
• Regulatory sovereignty: Enacting and enforcing digital laws. When a model's architecture, data documentation and training methods are fully transparent, governments and civil society can meaningfully evaluate compliance, fairness and safety.

Sovereign investment builds durable, in-house capability that gets stronger over time and that value compounds when you invest heavily in your custom training data pipelines:

• Building proprietary, domain-expert data labeling and annotation capacity.
• Constructing custom evaluation harnesses that reflect your operational realities rather than generic public benchmarks.
• Standing up a continuous red-teaming function staffed by specialists who understand your specific industry and cultural failure modes.

At TELUS Digital, we support each phase of the post-training data layer, utilizing domain-expert reviewers, multilingual data pipelines and culturally-adapted training sets. By treating the frontier model as a highly scalable engine and injecting your organization's unique context, language, culture and operational boundaries at the post-training layer, you transform a generic commodity into a powerful, proprietary moat. Contact our AGI data experts to bring your organization’s context to the frontier.