The Long View: When Models Become Commodities

The Price That Changed Everything

In March 2023, access to GPT-4 through OpenAI’s API cost approximately $0.03 per thousand input tokens and $0.06 per thousand output tokens. The model represented the frontier of artificial intelligence, and the pricing reflected it. Running a complex analysis that required processing fifty pages of text and generating a detailed response could cost several dollars — a meaningful expense at volume.

By early 2026, equivalent capability is available from multiple providers at a fraction of that cost. Open-weight models running on commodity hardware can deliver GPT-4-class performance for pennies. API providers compete aggressively on price. The capability that was extraordinary and expensive three years ago is now ordinary and cheap.

This collapse in the price of intelligence is one of the most consequential economic developments of the decade, and its implications extend far beyond the AI industry. When a capability that was scarce becomes abundant, the entire value chain that depends on that capability restructures. Value migrates away from the commodity and toward the things that remain scarce. Understanding where value migrates when intelligence becomes cheap is essential for anyone building a business, investing capital, or planning a career in the technology industry.

The Anatomy of Commoditization

Commoditization follows a predictable pattern across technology cycles. Understanding the pattern makes the AI industry’s current trajectory legible.

How It Starts

Every technology commoditization cycle begins with a period of scarcity. The new capability is expensive, controlled by a small number of producers, and differentiated primarily by performance. Early buyers pay premium prices because no alternatives exist and the capability is genuinely transformative.

In personal computing, the early microprocessors from Intel commanded extraordinary margins because there was no other way to get computing power in a small form factor. In cloud computing, Amazon Web Services enjoyed years of pricing power because the infrastructure to run scalable internet services was genuinely difficult to build. In telecommunications, early mobile networks charged premium rates because spectrum was limited and buildout costs were immense.

The scarcity phase is real, not artificial. The early producers are not overcharging — they are selling a genuinely rare capability at a price the market will bear.

How It Progresses

Commoditization begins when the barriers to producing the capability start falling. This happens through a combination of forces: competitors learn to replicate the technology, open-source alternatives emerge, the underlying inputs (hardware, talent, data) become more accessible, and customers develop the sophistication to evaluate alternatives.

In AI, all of these forces are operating simultaneously. Meta, Mistral, Alibaba, and others have released open-weight models that approach frontier performance. The knowledge required to train large models has diffused from a handful of research labs to a broader community of practitioners. Specialized hardware from multiple vendors is reducing dependence on any single chip manufacturer. And enterprise buyers, after three years of experimenting with AI, are increasingly capable of evaluating model performance for their specific use cases rather than relying on benchmark leaderboards.

How It Ends

Commoditization reaches its mature phase when the capability is no longer a differentiator. Buyers treat it as an input — essential but interchangeable. Prices converge toward marginal cost. Producers compete on efficiency, reliability, and price rather than on the capability itself.

AI models are approaching this phase faster than most industry observers predicted. The capability gap between the most expensive frontier models and the best open-weight alternatives has narrowed to a margin that matters for a small subset of tasks but is irrelevant for the majority of commercial applications. For most business use cases — summarization, classification, code generation, customer interaction, data analysis — the difference between the best model and the fifth-best model is smaller than the difference in cost.

Where Value Migrates

When a technology layer commoditizes, value does not disappear. It migrates to adjacent layers in the stack where scarcity still exists. Identifying those layers is the central strategic question for anyone operating in the AI economy.

Downward: Infrastructure and Efficiency

One direction value migrates is downward in the stack — toward the infrastructure that makes the commodity layer run. When model capability is abundant, the ability to serve that capability efficiently, reliably, and at low cost becomes the competitive differentiator.

This pattern played out in cloud computing. When basic compute and storage became commoditized, value accrued to the companies that could deliver those commodities most efficiently (AWS, Azure, Google Cloud) and to the infrastructure software that managed them (Kubernetes, Terraform, Datadog). The commodity itself was not valuable. The infrastructure for delivering the commodity at scale was immensely valuable.

In the AI stack, this pattern is already visible. Companies building inference optimization software, model serving infrastructure, and efficient deployment tooling are capturing value that the model providers cannot. The ability to run a given model at half the cost with equivalent latency is worth more than marginal improvements to the model itself for most production use cases.

Hardware specialization is another dimension of this downward migration. As standard GPU computing becomes the commodity layer, value moves toward specialized silicon designed for specific AI workloads. Custom inference chips, efficient edge-deployment hardware, and workload-specific accelerators represent scarce capabilities that command premium pricing even as the models running on them become interchangeable.

Upward: Applications and Workflows

Value also migrates upward — toward the applications, workflows, and user experiences that are built on top of the commodity layer. When the underlying intelligence is cheap and abundant, the scarce resource becomes the knowledge of how to apply that intelligence to solve specific problems for specific users.

This is the most significant value migration for the broader economy. The companies that will capture the most value from AI are not, in most cases, the companies that build the models. They are the companies that understand a specific domain deeply enough to apply commodity AI capability in ways that create measurable value for customers.

A legal technology company that uses commodity language models to automate contract review is selling domain expertise packaged as a product, not AI capability. A healthcare company that applies AI to clinical workflow optimization is selling healthcare knowledge, not model access. A financial services firm that uses AI for regulatory compliance analysis is selling regulatory expertise, not intelligence.

In each case, the model is an input — like electricity or internet bandwidth — not the product. The product is the domain-specific application, and its value comes from the understanding of the problem domain, the quality of the data pipeline, the design of the user experience, and the integration with existing workflows.

Laterally: Data and Distribution

The third direction of value migration is lateral — toward the assets that cannot be commoditized because they are inherently unique or accumulated through processes that cannot be easily replicated.

Data is the most important of these assets. Models can be replicated. Training techniques can be published. But proprietary datasets — customer interaction histories, transaction records, sensor data, domain-specific corpora — are unique to the organizations that collected them. As model capability becomes commodity, the data used to customize, fine-tune, and ground those models becomes the primary source of differentiation.

Distribution is the other critical lateral asset. The ability to reach users, embed in existing workflows, and occupy a position of habitual use is not affected by model commoditization. If anything, commoditization makes distribution more valuable, because when the underlying technology is interchangeable, the relationship with the customer becomes the decisive competitive factor.

This is why Microsoft’s integration of AI across Office, Teams, and Windows is strategically powerful regardless of which model it runs. The distribution channel — hundreds of millions of daily users in productivity applications — is the scarce asset. The model is the input.

The Trust Premium

One form of value that deserves special attention in the context of AI commoditization is trust. As model capability becomes commodity, the ability to provide trusted, reliable, accountable AI becomes a significant differentiator — and one that is not easily commoditized.

Trust in AI is a function of several factors: reliability (does the system produce correct outputs consistently?), transparency (can users understand how outputs were generated?), accountability (is there a responsible party when things go wrong?), and institutional validation (has the system been certified or endorsed by recognized authorities?).

These trust characteristics are expensive to build and maintain. They require sustained investment in testing, monitoring, documentation, support, and institutional relationships. They cannot be replicated by downloading open-weight models and standing up an inference endpoint. And they are increasingly valued by enterprise buyers who have moved past the experimentation phase and are deploying AI in production environments where reliability matters.

The trust premium explains why enterprise AI pricing has not collapsed to commodity levels even as the underlying model costs have plummeted. Companies are paying not for the intelligence but for the guarantee that the intelligence will work correctly, that failures will be handled, and that a responsible vendor will stand behind the system.

This dynamic is familiar from other commoditized markets. Generic pharmaceuticals contain the same active ingredients as branded drugs, but branded drugs command higher prices because of the trust built by the brand, the regulatory track record, and the institutional relationships with prescribing physicians. Enterprise software based on open-source components commands premium pricing because of the support, reliability guarantees, and accountability that the vendor provides.

In AI, the trust premium is likely to be one of the most durable sources of value as the model layer commoditizes. The companies that build the strongest trust infrastructure — through reliability engineering, transparent performance reporting, robust support, and institutional credibility — will command pricing power that pure model capability cannot sustain.

What This Means for the Major Players

The commoditization of AI models has different implications for different players in the ecosystem, depending on where their current value proposition sits in the stack.

The Model Providers

Companies whose primary business model is selling model access through APIs face the most direct challenge. OpenAI, Anthropic, and similar companies have invested billions in training frontier models and need to generate revenue from that investment. When equivalent capability is available from open-weight models at a fraction of the cost, the pricing pressure is relentless.

These companies have several strategic responses available. They can compete on frontier performance — staying ahead of the commodity curve by continuously advancing the state of the art. They can build proprietary features above the model layer — developer tools, enterprise integrations, workflow products — that create value beyond raw model access. They can establish trust premiums through reliability engineering and enterprise support. And they can use their API relationships as distribution channels for higher-value products.

The frontier strategy has limits. Training costs escalate with each generation, and the performance gap between frontier and commodity narrows quickly. Staying permanently ahead of the commodity curve requires exponentially increasing investment with diminishing returns.

The more sustainable strategy is to use model capability as a wedge to build durable advantages in adjacent layers — developer ecosystems, enterprise relationships, data flywheels, and trust infrastructure. The model itself becomes the loss leader. The business is built on everything around it.

The Cloud Providers

Amazon, Microsoft, and Google are positioned to benefit from model commoditization. As models become interchangeable, the infrastructure for serving them — compute, networking, storage, orchestration tools — becomes more valuable, not less. Every organization running AI needs infrastructure, and the cloud providers control the most efficient infrastructure at scale.

Moreover, the cloud providers can be model-agnostic. They can host OpenAI’s models, Anthropic’s models, Meta’s open-weight models, and any other model, taking a margin on the inference compute regardless of which model wins. Commoditization of the model layer is commoditization of their customers’ products, not their own.

Microsoft occupies a uniquely advantaged position because it combines cloud infrastructure (Azure) with dominant distribution (Office, Windows, LinkedIn) and a strategic relationship with a frontier model provider (OpenAI). If model capability commoditizes, Microsoft retains the infrastructure business, the distribution advantage, and the option to switch model providers if the economics or capabilities warrant it.

The Application Builders

Startups and established companies building AI-powered applications are the primary beneficiaries of model commoditization. Every reduction in the cost of model access reduces their cost of goods sold, widens their margins, and makes it easier to build competitive products on top of commodity intelligence.

The risk for application builders is that commoditization is bidirectional. If the model underneath their application is a commodity, the application itself may also become a commodity if its value proposition is primarily “AI applied to this use case” without deeper differentiation through data, workflow integration, or domain expertise.

The application builders that will thrive are those that use commodity AI as an input to products whose value comes from something AI cannot easily replicate: proprietary data, deep domain knowledge, integration with complex workflows, network effects, or relationships with specific customer segments.

The Enterprise Buyers

For organizations that consume AI rather than produce it, model commoditization is unambiguously positive. It means more capability for less money, more choices among providers, less vendor lock-in, and better negotiating leverage. Enterprises can increasingly treat AI model access the way they treat cloud compute — as a utility to be procured from whichever provider offers the best price-performance for a given workload.

The strategic challenge for enterprises is not access to AI capability, which is becoming cheap and abundant. It is the organizational capacity to use that capability effectively: the data infrastructure, the workflow redesign, the change management, and the talent to bridge the gap between commodity AI capability and specific business value. These organizational capabilities are the scarce resource in a world of abundant intelligence.

Historical Parallels

The commoditization of AI models is not historically unprecedented. Several prior technology cycles followed similar patterns and offer useful analogies.

The Server Market

In the 1990s, enterprise servers were expensive, proprietary, and differentiated primarily by performance. Sun Microsystems, HP, and IBM sold hardware at premium margins because the capability was scarce and the alternatives were limited.

The arrival of commodity x86 servers running Linux destroyed this value proposition over the course of a decade. The capability — running enterprise workloads reliably — became cheap and interchangeable. Value migrated to the software layer above (enterprise applications, databases, middleware) and the service layer around the hardware (managed hosting, which evolved into cloud computing).

The companies that understood this shift early — Amazon, which built AWS on commodity hardware, and Google, which built its infrastructure on commodity servers — captured enormous value. The companies that tried to defend premium pricing on the hardware itself — Sun Microsystems most notably — did not survive.

The Bandwidth Market

In the early 2000s, internet bandwidth was scarce and expensive. Telecommunications companies invested hundreds of billions in fiber-optic networks during the dot-com boom, and bandwidth prices reflected the scarcity and the capital investment.

The glut that followed the buildout — massive overcapacity that drove bandwidth prices toward zero — did not destroy the internet economy. It supercharged it. When bandwidth became cheap, the applications built on bandwidth — streaming video, cloud computing, social media, e-commerce — exploded. Value migrated from the bandwidth providers (many of whom went bankrupt) to the application layer built on top of commodity bandwidth.

The parallel to AI is direct. As model intelligence becomes cheap and abundant, the applications built on intelligence will proliferate in ways that are difficult to predict. The value will accrue not to the providers of commodity intelligence but to the builders of applications that leverage it.

The Electricity Analogy

The most expansive analogy is electricity itself. In the early decades of electrification, generating electrical power was a differentiated, premium-priced capability. Companies built proprietary generating stations and competed on the reliability and cost of their power.

Within decades, electricity became a regulated utility — cheap, ubiquitous, and interchangeable. The value created by electrification accrued not to the power companies but to the industries that used electricity as an input: manufacturing, transportation, communication, entertainment. The power companies earned regulated returns. The companies that applied electrical power to transform industries earned extraordinary returns.

AI may be following this pattern. The model itself may be on its way to becoming a utility — essential, ubiquitous, and priced at a regulated or competitive margin. The extraordinary returns may accrue to the organizations that use commodity intelligence to transform how work is done, how decisions are made, and how value is created in every sector of the economy.

The Implications Nobody Is Discussing

Beyond the industry structure questions that dominate current analysis, model commoditization has several implications that deserve more attention than they are receiving.

The Democratization of Intelligence

When GPT-4-level intelligence costs pennies, every organization — and potentially every individual — can access a level of analytical and creative capability that was previously available only to those who could afford to hire expensive professionals. A small business in a developing country can access the same quality of legal analysis, financial modeling, and strategic advice as a Fortune 500 company.

This has the potential to be profoundly equalizing. It is also potentially destabilizing, because the professionals whose expertise is being commoditized — lawyers, consultants, analysts, writers — represent a significant portion of the knowledge economy’s workforce.

The Quality Ceiling Problem

When multiple models can produce adequate output for most tasks, the distinction between good and great becomes harder to identify and harder to monetize. If a commodity model can draft a competent legal brief, the incremental value of a slightly better draft from a frontier model may not justify the incremental cost.

This creates a dynamic where “good enough” becomes the standard for an expanding range of tasks, with potential consequences for quality standards across knowledge work. The risk is not that AI produces bad work. It is that AI produces acceptable work at such low cost that the incentive to produce excellent work diminishes.

The Attention Economy Becomes Everything

When producing content, analysis, and communication becomes nearly free, the binding constraint on the value of any piece of content is not the cost of production but the attention of the audience. We already live in an attention economy. AI commoditization intensifies this dynamic by orders of magnitude.

The scarcest resource in a world of commodity intelligence is not intelligence itself. It is human attention, human trust, and human judgment about what is worth attending to. The organizations and individuals that can command attention in a flood of AI-generated content will capture disproportionate value. Those that produce content without a distribution advantage will find their output lost in an ocean of commodity intelligence.

Conclusion

The commoditization of AI model capability is not a future event. It is happening now, and it is happening faster than most participants in the AI economy anticipated. GPT-4-level intelligence, which was frontier capability three years ago, is approaching commodity pricing from multiple directions — open-weight models, competitive API providers, and efficiency improvements that reduce inference costs by orders of magnitude.

The strategic implications are clear, even if the specific outcomes are uncertain. Value is migrating away from the model layer and toward the layers where scarcity persists: infrastructure efficiency, domain-specific applications, proprietary data, distribution networks, and trust. The companies that recognize this migration and position themselves accordingly will thrive. Those that cling to the model layer as a source of sustainable differentiation will face intensifying margin pressure.

The historical parallels — servers, bandwidth, electricity — all suggest the same conclusion. When a transformative capability commoditizes, the total value created by that capability increases enormously, but the value captured by the commodity producers decreases. The winners are those who build on top of the commodity, not those who produce it.

Intelligence is becoming a commodity. What you build with it is not.