AI gets cheaper.
What stays
valuable?

Frontier training remains concentrated among organizations with exceptional research talent, capital and cluster access.

They can charge for capabilities that customers cannot reproduce or obtain from a cheaper model. The rent is the performance gap—not intelligence in the abstract.

Open models, distillation and diminishing willingness to pay narrow that gap use case by use case.

Accelerator supply and integrated systems remain difficult to add quickly, even as model access becomes less scarce.

Training and high-volume inference are bounded by usable compute, not chip counts alone. Vendors that combine silicon, networking and software capture more of the system economics.

Custom silicon, better utilization and a shift from training to cheaper inference weaken the premium on general-purpose frontier hardware.

Leading-edge fabrication, advanced packaging and high-bandwidth memory expand on industrial rather than software timelines.

These suppliers control capacity that cannot be replicated with code or purchased at short notice. Their bottlenecks can persist after model quality begins to converge.

Committed capacity eventually arrives; when it does, volume may remain high while scarcity margins fall.

Ever-larger clusters require more bandwidth within and between racks, making network architecture part of model economics.

A costly accelerator is useful only when data can reach it. Suppliers capture value by removing communication bottlenecks that otherwise leave scarce chips idle.

Standards, merchant silicon and slower growth in frontier cluster size can transfer value back to buyers.

Grid connections, generation, cooling and power-management equipment have lead times measured in years.

The constraint is increasingly energized capacity rather than server procurement. Suppliers sell into a build-out that cannot proceed without them and is difficult to accelerate.

Their order books can outlast peak scarcity, but returns normalize when data-center construction slows or designs become more efficient.

Few operators can finance, provision and run large clusters across regions with enterprise-grade availability.

They monetize balance sheet, procurement access and operating expertise by turning scarce physical infrastructure into capacity customers can consume immediately.

Large fixed costs make this category vulnerable once capacity is interchangeable and utilization becomes the central problem.

Supply is fragmented across chips, clouds and model architectures while production demand is rising faster than internal serving expertise.

They capture the spread between raw accelerator time and reliable, model-specific throughput. Optimized kernels, batching and capacity management let them deliver more useful tokens per dollar.

That spread compresses as optimization diffuses, capacity grows and customers treat inference providers as interchangeable.

Rapid model releases create meaningful differences in price, latency, capability and availability from one request to the next.

The gateway sees the order flow and can choose the supplier. That position compounds through usage data, policy controls, failover and one integration across many models.

Routing succeeds by commoditizing what it routes. Clouds and application platforms can absorb the feature once the decision logic becomes standard.

Capable open models and smaller hardware footprints make private and on-device deployment practical for a growing set of workloads.

These projects set the deployment standard and reduce dependence on proprietary clouds. Commercial value can accrue through managed hosting, enterprise support and control of the local model workflow.

Adoption alone does not create a business; agents also reduce the labor premium customers once paid to avoid self-deployment.

Developers need discovery, testing, governance and deployment across an unusually fast-changing supplier base.

Platforms capture value by becoming the place a model is first found, evaluated and integrated. Distribution can persist after any individual model loses its premium.

The largest clouds can bundle the catalog, while open standards keep migration costs lower than in traditional platform markets.

Agents are moving from single responses to long-running jobs that use browsers, code, files and external systems.

They sell the controlled environment in which unreliable model output becomes resumable, observable work. Security isolation and state management matter before agents can receive broader permissions.

Basic sandboxing and orchestration may consolidate into clouds, operating systems or agent frameworks.

Unit economics depend heavily on compiler choices, quantization, caching and hardware-aware scheduling.

Small efficiency gains multiply across enormous token volumes. Optimizers capture value while expertise is scarce and serving stacks remain heterogeneous.

The economics are self-eroding: successful techniques enter open runtimes, chip libraries and cloud platforms.

Enterprises can access capable models before they can safely connect them to the systems where consequential work is recorded.

These incumbents already hold current state, permissions and workflow. They can place agents inside an approved operating environment rather than asking customers to reconstruct one elsewhere.

Possession of data is not enough if a new interface owns the user and treats the system of record as a replaceable back end.

Generic intelligence is abundant, but useful answers depend on fresh, permission-aware access to proprietary state.

They sit between models and changing enterprise data. Value comes from governed access, lineage and real-time retrieval—not from attaching the word ‘AI’ to stored data.

Static datasets move easily; durable value requires staying inside the live read-and-write path of the workflow.

The immediate obstacle is not reasoning but the fragmented work required to connect agents to thousands of software tools and data sources.

Connection layers can become the authorized route through which agents discover and use enterprise tools. Breadth, maintenance and policy enforcement create the commercial advantage.

Open protocols reduce lock-in, and systems of record may expose their own first-party agent interfaces.

Agents need identities distinct from their human sponsors and permissions bounded by task, time and resource.

Autonomy cannot expand without a control plane that can authenticate an agent, limit its authority and revoke it. Existing trust relationships and integrations are difficult to reproduce quickly.

Identity suites can bundle agent controls; point solutions must own a genuinely new policy layer rather than a renamed service account.

Non-deterministic agents are reaching production before organizations have adequate evidence of how they behave.

These platforms turn traces and outcomes into tests, audit records and feedback. That evidence lets customers grant more autonomy without accepting unknowable risk.

Tracing is easy to commoditize; durable value depends on proprietary evaluation data, workflow integration or becoming part of the approval process.

Prompt injection, tool misuse and data leakage become budgeted problems once agents can take actions rather than merely generate text.

They capture value by reducing the expected cost of autonomy and by producing the controls regulators, insurers and security teams require before deployment.

Security platforms can absorb these capabilities, and fear alone will not sustain a category without measurable risk reduction.

Narrow domains can support agents earlier because the workflow, vocabulary, data and acceptable outcome are more clearly defined.

They capture value by owning the completed task and the feedback it generates. Deep integration and domain liability matter more than access to any one model.

A vertical product is vulnerable if customers can reproduce the workflow with a general assistant and a small number of connectors.

Continuous learning requires specialized human feedback, task environments and outcome data after the easy public corpus has been consumed.

They organize scarce expertise and generate the environments in which models learn to perform economically useful work. The value lies in repeatable production, not a one-off dataset.

Margins fall when data production is undifferentiated or when customers internalize the workflow; privileged supply and quality control must persist.

Capability differences become less visible to mainstream users while assistants are embedded into phones, PCs, glasses and home devices.

The operating surface receives intent before a model is selected. Defaults, hardware integration and installed base let the platform route demand and make upstream suppliers invisible.

A default is valuable only if people use it. Weak products can still cede the relationship to an application that users actively choose.

Assistants are building direct consumer habits before the underlying models become fully substitutable.

Repeated use creates memory, preference data, identity and trust. If users delegate decisions to the assistant, it becomes the buyer of downstream models and services rather than a reseller of answers.

The relationship can be displaced by the operating system, browser or specialist product where the user’s intent actually originates.

Enterprise agents first appear inside products employees already open each day and that already carry organizational identity.

These products own both attention and workflow context. They can choose the model, surface the agent at the moment of intent and retain the feedback generated by completed work.

They lose leverage if the agent becomes the primary interface and reduces the underlying application to a tool call.

Research, navigation and commercial discovery are being compressed into conversational and agentic interfaces.

Browsers and search products observe high-intent queries at the point where users choose what to read, buy or do. That position can direct traffic, transactions and model selection.

Answer interfaces can weaken the open-web supply that makes them useful, while operating systems can move discovery one layer higher.

Delegation becomes economically consequential when agents can compare, purchase and settle on a user’s behalf.

These networks combine merchant access, identity, payment credentials and dispute handling. An agent needs all four to convert a recommendation into a trusted transaction.

Standards may separate the agent relationship from fulfillment and payment, forcing each layer to compete on price.

Users delegate sooner in domains where a product already understands the task, history and acceptable boundary of action.

Domain trust and longitudinal context make the product a plausible representative rather than a generic interface. The strongest positions own a recurring decision, not a single query.

The relationship is less durable where a general assistant can import the history and reproduce the service without regulatory or workflow friction.

Capable local models make it possible to keep private context and continuous memory on hardware the user controls.

If local models become capable and economical enough, applications could use private context without a metered cloud call for every task. A personal agent could combine local inference with paid cloud services. This is one possible architecture for owning the customer relationship, not a prerequisite.

Convenience may outweigh sovereignty for most consumers, leaving the category influential technically but difficult to monetize directly.

Falling inference costs allow products to deliver completed work in markets that could not support the cost of a human service or repeated frontier-model calls.

These companies can turn labor into a repeatable software product: research completed, claims processed, customers served or code shipped. Cheap intelligence creates the supply; workflow ownership, brand and distribution capture the margin by packaging it into an outcome someone will pay for.

Cheaper intelligence can lower the barrier for competitors too. Without proprietary workflow, trust, data or a low-cost route to customers, the service becomes as interchangeable as the model beneath it.

Synthetic production becomes abundant while human attention remains fixed and verified authorship becomes easier to value by contrast.

Platforms and brands capture value by aggregating scarce audiences, identity and cultural signal. Utility content can automate; participation, status and human provenance cannot be manufactured at the same marginal cost.

The undifferentiated middle is exposed. Only products with genuine audience ownership, trusted curation or scarce human origin retain the premium.