As artificial intelligence (AI) models become cheaper and increasingly interchangeable, the industry’s most durable value may shift away from the models themselves towards companies that can turn intelligence into secure, reliable and completed work.
That is the central argument from Jonathan Curtis, Portfolio Manager at Franklin Equity, who believes the commercial AI race will not ultimately be decided by which company briefly possesses the smartest model.
Instead, competitive advantage will belong to businesses that control the infrastructure, proprietary data, workflows, governance and customer relationships surrounding those models.
“The model is becoming a commodity. The system is becoming the moat,” Curtis said, arguing that enterprises are not simply purchasing intelligence by the token. They are paying for outcomes that can be delivered reliably, securely and at scale.
Intelligence Is Getting Cheaper
The rapid improvement of open-weight AI is narrowing performance gaps and lowering the cost of accessing advanced capabilities.
Models that once required enormous development budgets can increasingly be downloaded, adapted and deployed at a fraction of the previous cost. This has fuelled concerns that model providers and the data centres supporting them could struggle to generate attractive returns as prices fall.
Curtis, however, rejects the view that cheaper intelligence will inevitably weaken the AI investment opportunity.
Businesses do not ultimately pay for individual tokens, he said. They pay for code that can be deployed, customer problems that can be resolved, claims that can be underwritten and decisions that can withstand scrutiny.
The model remains important, but it is only one component within the broader system required to produce a dependable result.
Cost Per Task Matters More Than Cost Per Token
For enterprises, the relevant economic measure is not necessarily the price of generating an answer. It is the total cost of completing a task correctly.
A cheaper model may require several attempts, additional verification, human supervision and extensive correction. A more expensive model could prove more economical if it produces the required result accurately on its first attempt.
Curtis compares the distinction to selecting a contractor solely according to hourly wages while ignoring the duration of the project, the cost of rework and the quality of the completed job.
“The more important question is not how cheaply a model can generate an answer,” he said. “It is how efficiently the full system can produce a trustworthy outcome.”
This shifts the investment debate away from headline model benchmarks and towards the systems that enable AI to function effectively inside complex organisations.
The System Becomes The Competitive Advantage
A model cannot perform valuable enterprise work in isolation.
It must be connected to proprietary information, organisational context, business applications and clearly defined workflows. It also requires permissions, security controls, output evaluations, audit trails and rules governing when human intervention is necessary.
Curtis describes this surrounding system as the “harness”. It selects the most suitable model, connects it with data and tools, tests its output, enforces permissions and determines when a person must review or approve an action.
Over time, that harness can accumulate institutional memory, workflow logic, security policies and knowledge of where failures are most likely to occur.
An enterprise may replace its underlying model several times while retaining much of this surrounding infrastructure. As models become more interchangeable, Curtis believes this persistent layer could become considerably more valuable than the model itself.
AI’s Profit Pool Moves Up The Stack
The commoditisation of model access does not mean the economics of AI will disappear. It means the value is likely to migrate.
Curtis expects the most valuable layers to include computing infrastructure, distribution, proprietary context, workflow ownership, governance and trust.
The weakest business model may be one that sells undifferentiated access to a model while depending on a temporary performance lead to maintain premium pricing. Such advantages can quickly erode as innovations spread, customers switch providers and applications begin routing work across competing models.
Frontier model developers are already responding by expanding into enterprise platforms, agent systems, security and workflow software.
But moving up the technology stack will not be straightforward. Established software providers, professional services firms and large enterprises already control customer relationships, data and critical business processes — positions they are likely to defend aggressively.
Model providers could also create conflicts with their own customers. The further they expand into legal services, software development, financial analysis, research or customer support, the more likely those customers are to view them as competitors rather than neutral technology suppliers.
That tension could strengthen demand for private deployments, model portability and platforms capable of working with several models rather than depending on a single provider.
Lower Prices Could Expand The Market
Falling AI prices may also increase, rather than reduce, the industry’s overall revenue opportunity.
Traditional software interactions may require a single request and response. An AI agent attempting to complete a business process could require dozens or even hundreds of model calls to plan its work, retrieve information, use tools, test alternatives, correct mistakes and verify the result.
As intelligence becomes cheaper, companies are likely to apply it to more problems and automate more stages of their operations.
Curtis points to the equivalent of Jevons Paradox in computing: as a resource becomes cheaper and more efficient to use, total consumption can rise faster than its unit price falls.
This suggests lower model costs could expand both demand for intelligence and the infrastructure required to produce it. The crucial variable will be whether usage grows faster than prices decline.
Winning Platforms Will Use Multiple Models
Curtis believes the strongest competitive position may belong to companies capable of orchestrating several AI models.
Such platforms could route each task to the provider offering the best combination of quality, security, reliability, speed and cost. Expensive frontier models could be reserved for difficult problems, while smaller and cheaper alternatives handle routine work.
This ability to manage what Curtis calls “disciplined abundance” could become a major enterprise advantage.
For investors, the key question is therefore not simply which company has the best-performing model. It is which companies can convert abundant intelligence into completed work at the lowest total cost — while maintaining sufficient reliability and trust for customers to redesign their operations around the system.
Trust May Become AI’s Most Valuable Product
For consequential business functions, price and raw performance will not be enough.
Customers will pay for reliability, explainability, security, auditability and clear accountability when something goes wrong. A low-cost model has limited commercial value if an organisation cannot approve it for production use.
Infrastructure providers could benefit if consumption grows faster than unit prices decline, while enterprise software platforms may create value by aggregating models and managing deployment, security, billing and governance.
Applications that own critical workflows, proprietary context and customer relationships could also build lasting advantages.
By contrast, Curtis sees the greatest vulnerability in the “undifferentiated middle” — expensive models and thin applications that lack cost advantages, unique data, distribution, workflow ownership or responsibility for the final outcome.
The Investment Opportunity Is In Trusted Work
The AI market is likely to remain noisy and volatile as investors attempt to determine where its long-term value will settle.
But Curtis’ conclusion is clear: the supply of intelligence is increasing rapidly, and simply producing more of it will not guarantee commercial success.
The greatest value is likely to be captured by companies that can translate that intelligence efficiently into trusted, explainable and completed work.
In the next stage of the AI economy, the winning business may not own the most powerful model. It will own the system that enterprises cannot afford to operate without.
