Artificial intelligence is not merely another software feature. It is a new computing layer that can understand information, generate content, make predictions, interact with software, and increasingly execute work. The previous major enterprise technology transition, from on-premises software to cloud infrastructure and software-as-a-service, provides several important lessons for the AI era. First, major platform shifts tend to expand the overall technology market rather than simply transfer revenue from incumbents to startups. During the cloud transition, established software companies continued growing, while new infrastructure providers and SaaS companies captured large new markets. Second, infrastructure can produce some of the largest and most durable winners. Cloud computing created enormous businesses around compute, storage, networking, databases, security, developer tooling, observability, and data platforms. AI is now creating a parallel infrastructure stack around chips, data centers, foundation models, inference, orchestration, vector databases, evaluation, governance, agent infrastructure, and specialized developer platforms. Third, application-layer competition will be intense because it is becoming easier to build basic AI products. A simple interface connected to a third-party model is unlikely to remain defensible. Sustainable AI businesses will need proprietary workflows, trusted distribution, valuable data, embedded customer relationships, superior economics, regulatory advantages, network effects, or the ability to complete meaningful work rather than merely generate content.
Fourth, incumbent software companies may adapt faster to AI than legacy software companies adapted to cloud computing. They already operate subscription services, continuously deploy software, possess large customer bases, and control valuable systems of record. However, adding a chatbot to an existing product is not the same as reinventing the workflow. Fifth, the most valuable AI applications may evolve from systems of record into systems of prediction and systems of execution. Instead of merely storing what happened, AI software can recommend what should happen next and eventually perform the action itself. Sixth, value capture will depend less on access to a model and more on defensibility. Models will continue improving and, in many cases, becoming less expensive. Companies that depend entirely on one external model or a temporary capability advantage may face rapid commoditization. Finally, customers may become the greatest long-term beneficiaries. Competition across chips, models, cloud providers, open-source ecosystems, and applications should improve capability while reducing the cost of intelligence. The result could be a dramatic expansion in software consumption, productivity, scientific discovery, automation, and new business formation.
Introduction: Every Platform Shift Creates a New Map of Power Technology history is filled with moments when the foundations of computing changed. Mainframes gave way to personal computers. Desktop software was connected through the internet. Web browsers became distribution platforms. Mobile devices changed how people interacted with digital services. Cloud computing transformed software delivery, infrastructure ownership, and enterprise purchasing. Software-as-a-service changed how products were developed, sold, updated, and monetized. Each transition created new market leaders. Microsoft rose during the personal computer era. Google and Amazon became dominant during the internet era. Apple and Google gained extraordinary strategic power through mobile operating systems and app distribution. Amazon Web Services, Microsoft Azure, and Google Cloud became essential infrastructure providers during the cloud transition. Artificial intelligence represents another platform shift, but it may be broader than those that preceded it. Cloud computing primarily changed where software ran, how it was delivered, and how customers paid for it. AI changes what software can do. Traditional software follows instructions written explicitly by developers. AI software can interpret unstructured information, understand natural language, identify patterns, generate new outputs, make probabilistic decisions, interact with people, and increasingly coordinate actions across multiple systems. This changes the competitive landscape across the entire technology stack.
The contest is no longer only between application companies. It includes semiconductor manufacturers, cloud providers, foundation-model developers, data-center operators, networking companies, data platforms, security vendors, orchestration tools, enterprise software incumbents, specialized startups, consulting firms, and organizations building autonomous agents. The original Andreessen Horowitz analysis examined 262 public business-to-business software companies and used the cloud transition to understand how value might be distributed during the AI era. Its central argument was that platform shifts can be positive-sum expansions and that newly created infrastructure layers may produce some of the largest winners. That framework remains useful, but the AI market has moved considerably since the article was published in August 2023. Enterprise adoption has accelerated. Model performance has improved. Open-source alternatives have expanded. AI infrastructure investment has risen sharply. Major cloud providers have integrated models throughout their platforms. Software companies are moving from experimental copilots toward agents capable of completing multistep work. The race to capture value is now fully underway.
1. Platform Shifts Are Usually Positive-Sum Expansions
Technology disruption is often described as a winner-takes-all battle. A startup introduces a new technology. The old incumbent fails to adapt. Customers migrate. The startup wins, and the incumbent disappears. This does happen occasionally, but major platform transitions are usually more complicated. They often expand the total market. During the cloud era, new SaaS companies captured substantial market share from traditional software vendors. At the same time, many established companies continued increasing revenue by adapting their products, acquiring competitors, building cloud services, or serving customers that could not migrate immediately. According to the historical analysis presented by a16z, aggregate revenue across its sample of public B2B software companies increased from approximately $99 billion near the beginning of the studied period to $587 billion two decades later. Incumbents added revenue even while new entrants captured a significant portion of the expanding market. This matters because the AI opportunity should not be viewed as a fixed pool of software spending that startups and incumbents must divide. AI can increase the amount organizations are willing to spend on technology because it can address categories of work that traditional software could not economically automate. Traditional enterprise software helps humans organize work. AI software can participate in the work. A customer-service platform once stored tickets, routed inquiries, and displayed customer histories. An AI-enabled platform can interpret the inquiry, retrieve relevant information, draft or deliver a response, update the customer record, recommend compensation, identify churn risk, and escalate unusual cases.
A legal platform once stored documents and supported keyword searches. An AI-enabled legal system can summarize evidence, compare clauses, identify risks, prepare drafts, organize arguments, and monitor compliance obligations. A healthcare information system once recorded patient histories. An AI-enabled system can assist with documentation, triage, clinical decision support, administrative workflows, patient communication, and operational planning. The addressable market is therefore not limited to the existing software budget. AI vendors may compete for portions of labor budgets, outsourcing budgets, consulting expenditures, business-process services, call-center expenses, administrative overhead, and spending associated with mistakes, delays, fraud, and operational inefficiency. This does not mean every AI company will succeed. It means the total economic surface area available to software is expanding.
2. AI Is a Larger Change Than the Move from On-Premises Software to SaaS
The cloud transition was revolutionary, but much of the early SaaS market recreated existing software categories through a superior delivery model. Cloud customer-relationship management replaced locally installed CRM systems. Cloud human-resources software replaced on-premises HR platforms. Cloud productivity tools replaced desktop applications and internally managed collaboration systems. The new products were often easier to deploy, continuously updated, accessible through a browser, and sold through subscriptions. These changes were enormously valuable, but the fundamental purpose of many applications remained recognizable. AI can change the purpose of software itself. Software has historically been a tool that a person operates. AI allows software to become a participant that interprets intent and produces outcomes. This transition can be understood through three generations. Generation One: Systems of Record A system of record stores authoritative information.
Examples include:
Customer records Financial transactions Employee information Inventory levels Medical histories Contracts Support tickets Project plans The primary value is organizational memory. Generation Two: Systems of Prediction A system of prediction uses historical and real-time information to estimate what is likely to happen.
Examples include:
Which customer is likely to cancel Which transaction may be fraudulent Which lead is most likely to purchase Which machine is likely to fail Which employee may leave Which shipment may arrive late Which patient may require additional attention The primary value is decision support. Generation Three: Systems of Execution A system of execution does not merely recommend an action. It performs the action within defined permissions.
Examples include:
Contacting the customer Issuing a refund Rescheduling a shipment Updating a marketing campaign Negotiating an appointment Reconciling an invoice Producing and submitting a report Opening a support case Ordering replacement inventory Coordinating work across multiple applications The primary value is completed work. A company that successfully moves from record-keeping to prediction and execution can justify a much larger share of customer spending. It is no longer selling access to a database or interface. It is selling productivity, revenue growth, risk reduction, or an operational outcome.
This is why the AI shift may become much larger than the SaaS transition.
3. The Cloud Era Shows Why Infrastructure Can Capture Enormous Value
Applications receive much of the public attention because customers interact with them directly. However, some of the most powerful businesses created during platform shifts operate beneath the application layer.
The cloud era created massive value around:
Compute Storage Networking Databases Content delivery Identity Cybersecurity Monitoring Developer platforms Data integration Data warehousing DevOps
Application programming interfaces Cloud infrastructure became a foundation on which thousands of application companies were built. The same process is occurring in AI. The emerging AI stack includes several layers. Semiconductor and Hardware Infrastructure AI systems require specialized computing resources.
This layer includes:
Graphics processing units AI accelerators Custom silicon High-bandwidth memory Networking equipment Optical interconnects Cooling systems Power-management systems Servers Storage systems Companies that control scarce or highly differentiated hardware can capture substantial value, particularly when demand exceeds supply. However, hardware leadership requires constant investment. Competitors, cloud providers, and large customers are developing alternatives, including custom chips optimized for specific workloads.
Data-Center and Energy Infrastructure AI workloads consume enormous amounts of compute, electricity, cooling capacity, and network bandwidth.
This creates opportunities in:
Data-center construction Power generation Grid infrastructure Energy storage Cooling technology Data-center operations Site selection Specialized real estate Backup systems High-capacity networking AI infrastructure is becoming inseparable from energy strategy. The firms that can secure power, land, equipment, and permits may possess a meaningful advantage, particularly in regions where infrastructure development is constrained.
Cloud Compute The major cloud platforms are positioned to benefit regardless of which application becomes dominant because training and inference require infrastructure. Microsoft reported that Azure surpassed $75 billion in annual revenue during its 2025 fiscal year, rising 34 percent, while the company described AI as a transformation affecting every layer of its technology stack. The strategic position of cloud providers is unusually strong because they can participate across multiple layers: Renting compute Providing data storage Hosting models Offering proprietary models Distributing third-party models Selling databases and security services Providing agent-development tools Integrating AI into their own software applications
This creates both opportunity and strategic tension. An independent AI company may depend on the same cloud provider that is also building a competing model or application. Foundation Models Foundation-model providers create general-purpose systems that developers and enterprises can adapt for many tasks.
Potential advantages include:
Superior model performance Lower inference costs Faster response times Multimodal capabilities Developer adoption Enterprise trust Safety systems Distribution partnerships Access to proprietary data Ability to finance large training runs However, this layer also faces intense competitive pressure. Stanford’s 2025 AI Index noted rapid growth in model creation and open-source participation. It reported that 149 foundation models were released in 2023, more than double the previous year, and that 65.7 percent were open source.
As model quality converges for many commercial tasks, customers may switch providers, use several models, or route each workload to the best combination of cost and performance. This could reduce pricing power unless a model provider builds additional advantages around distribution, infrastructure, proprietary capabilities, ecosystems, or customer integration. AI Middleware and Developer Infrastructure Between models and applications, a large tooling layer is emerging.
It includes:
Model gateways Model routing Retrieval systems Vector databases Agent orchestration Prompt management Evaluation platforms Observability Fine-tuning tools Synthetic-data platforms Guardrails Security
Identity and access control Memory systems Cost management Inference optimization Human-approval workflows These products can become valuable when they solve operational problems that remain important regardless of which underlying model wins. A model-routing company, for example, may benefit from model competition because customers need a neutral layer that selects among providers. An AI evaluation platform may become more valuable as enterprises deploy more models and agents. An identity platform designed for autonomous agents may become essential when software systems begin initiating payments, accessing accounts, communicating with customers, and making decisions on behalf of organizations. The strongest middleware businesses will not merely wrap model APIs. They will become control points within production AI systems.
4. Infrastructure Is Powerful, but It Is Not Automatically Profitable
High demand does not guarantee attractive economics. AI infrastructure can require enormous capital expenditure. Companies must finance chips, data centers, networking equipment, energy capacity, model training, engineering talent, and ongoing inference. The resulting market may generate substantial revenue while still producing difficult economics for some participants. Several forces affect profitability. Price Competition Compute providers may reduce prices to attract workloads. Model providers may lower API prices as efficiency improves or competitive pressure increases. Open-source models may reduce the amount customers are willing to pay for access to general-purpose intelligence. Depreciation Risk AI hardware can become economically obsolete before it becomes physically unusable. A new generation of accelerators may provide far better performance per dollar or per watt. Infrastructure owners must therefore recover capital investments quickly enough to remain competitive.
Utilization Risk Expensive hardware creates value only when it is being used. Low utilization can destroy infrastructure economics. Providers need scheduling, workload management, customer demand, and pricing systems that keep assets productive. Customer Concentration A company that relies on a small number of large AI laboratories or cloud customers may face negotiating pressure and revenue volatility. Vertical Integration Cloud providers are designing chips. Model developers are building applications. Application companies are fine-tuning models. Hardware companies are expanding into software and cloud services. This vertical integration can compress the margins of companies occupying narrow layers without a defensible position. Therefore, founders should not assume that being closer to the infrastructure automatically creates a better business. The important question is whether the company controls a scarce, difficult-to-replace resource or an essential workflow.
5. Why AI Application Companies Face a Defensibility Crisis
The application layer offers extraordinary opportunity because almost every industry and business function can be redesigned. It also faces a fundamental problem: basic AI applications are increasingly easy to build. A developer can connect an interface to a model API and produce a functional prototype in days or even hours. This is excellent for experimentation, but it creates a crowded market.
Many early AI products have similar characteristics:
They rely on the same underlying models. They serve similar customer needs. They use comparable interfaces. They can be reproduced quickly. They have limited switching costs. They depend on another company’s pricing and policies. They may lose differentiation when model providers release new native features. A thin application can still become a successful company, but it must become thicker over time. It needs to accumulate advantages that cannot be recreated by simply calling the same model.
6. The Most Important Sources of AI Defensibility
Defensibility determines who creates temporary revenue and who builds durable enterprise value. Several forms of defensibility are particularly important in the AI economy.
6.1 Proprietary Workflow Integration
The strongest AI applications often become deeply integrated into how an organization operates.
They understand:
Internal approval processes Customer histories Pricing rules Product catalogs Compliance requirements Employee responsibilities Company terminology Decision thresholds Escalation procedures Existing software systems Once a product becomes embedded in a complex workflow, replacement becomes expensive and risky. The defensibility does not come only from the AI model. It comes from operational integration.
6.2 Systems of Record
Existing systems of record possess several advantages. They already contain valuable business data. They are trusted. They have permissions. They are integrated with other systems. Employees use them daily. An AI startup can challenge these incumbents by creating a better experience, but the incumbent can also use AI to make its existing data and workflow more valuable. This creates one of the most important battles in enterprise software. Will AI become a feature inside current systems of record, or will a new generation of AI-native systems replace them? The answer will vary by category. In some markets, incumbents will successfully add intelligence to existing platforms. In others, AI-native companies will change the workflow so dramatically that the old system becomes merely a database in the background.
6.3 Proprietary Data
Data can create a moat when it is:
Exclusive Difficult to collect Continuously generated Directly connected to outcomes Legally usable High quality Expensive to reproduce Essential to product performance Not all data is defensible. Public information, purchased datasets, and generic customer interactions may be available to competitors. The most valuable data often emerges as a by-product of product usage. For example, a fraud-detection system may learn from confirmed fraud outcomes. A robotics company may collect physical-world interaction data. A healthcare platform may improve from clinician feedback. An industrial system may learn from equipment failures and maintenance results.
The a16z article highlighted companies whose access to difficult-to-obtain data creates strategic advantages, including examples from mining, defense, and computer vision.
6.4 Feedback Loops
A powerful AI product improves as it is used.
A defensible feedback loop may look like this:
More customers use the product. The product receives more interaction data. The company measures outcomes. The model or workflow improves. The product becomes more accurate or useful. Better performance attracts more customers. The crucial element is outcome measurement. Data without a connection to success or failure may produce limited improvement.
6.5 Distribution
Distribution can be more defensible than technology. A company with trusted access to millions of customers can add AI capabilities rapidly.
Distribution advantages may come from:
Existing enterprise contracts Large consumer audiences App stores Cloud marketplaces Developer ecosystems Hardware installations Channel partnerships Professional communities Industry associations Embedded financial relationships A technically superior product can still lose when a slightly weaker competitor is easier to buy, deploy, trust, or integrate.
6.6 Trust, Security, and Compliance
Enterprise customers do not purchase AI capability in isolation.
They must evaluate:
Data privacy Security Reliability Auditability Regulatory compliance Intellectual-property risk Model behavior Access controls Business continuity Vendor stability Incumbents often benefit from established security certifications, procurement relationships, and compliance systems. Startups can still win, particularly when they build trust specifically for a regulated or high-risk workflow.
A generic AI assistant may struggle to enter a hospital, bank, defense organization, or government department. A specialized platform designed around that institution’s legal, operational, and security requirements may become highly defensible.
6.7 Network Effects
Network effects arise when a product becomes more valuable as more participants use it.
Possible AI-era network effects include:
Marketplaces connecting users and specialized agents Platforms where developers publish agent capabilities Products that improve from community-created workflows Collaboration systems where teams train shared organizational memory Communication networks used by humans and agents Reputation systems for autonomous software Identity networks that verify agents and permissions Transaction platforms where agents buy and sell services Network effects are powerful because they make the product more valuable while making alternatives less attractive.
6.8 Superior Unit Economics
A company can create a lasting advantage by delivering the same or better result at lower cost.
AI unit economics depend on:
Model choice Token consumption Context size Caching Model routing Fine-tuning Hardware utilization Latency requirements Human-review costs Error rates Customer-support costs Pricing design
A product that appears profitable during experimentation may become unprofitable at scale if every customer action triggers expensive model usage. Successful AI companies must treat inference economics as a core product discipline, not merely an infrastructure expense.
6.9 Outcome Ownership
The most defensible AI businesses may be those willing to own the result.
Instead of selling software seats, they may charge for:
A qualified sales meeting A resolved support request A processed invoice A completed compliance review A successful repair A recovered payment A generated engineering design A verified medical document A completed shipment A percentage of measurable savings Outcome-based pricing can align the vendor with the customer and increase the available revenue pool. It also requires confidence, operational control, and measurable performance.
7. Incumbents Are Adapting Faster Than They Did During the Cloud Transition
Many legacy software companies were slow to adopt cloud computing. They faced technical and organizational obstacles. Their products had been designed for local installation. Their revenue depended on licenses and maintenance contracts. Their sales teams were rewarded for large upfront transactions. Their release processes were slow. Their infrastructure assumptions were different. Moving to SaaS required more than placing the existing product on a remote server. It demanded changes to architecture, pricing, sales compensation, customer support, security, deployment, and product development. Modern SaaS companies do not face the same transformation.
They already:
Host products in the cloud Operate subscription models Release updates continuously Collect usage data Maintain APIs Employ machine-learning teams Serve customers remotely Integrate with cloud infrastructure As a result, established software companies can add AI functionality faster than earlier incumbents adopted SaaS. This does not mean they will automatically win. There is a large difference between adding AI to the existing interface and redesigning the product around intelligence. A chatbot that helps users find existing features may improve usability.
A truly AI-native product may eliminate the need to operate those features manually. For example, an incumbent project-management platform might add an assistant that summarizes project updates. An AI-native competitor might automatically create plans, assign work, follow up with participants, detect delays, negotiate scheduling conflicts, update stakeholders, and adjust the project based on changing conditions. The first product improves the existing workflow. The second changes the workflow. Incumbents must therefore avoid confusing feature velocity with strategic adaptation.
8. Startups Must Attack Workflows, Not Interfaces
A common startup strategy is to create a more attractive or conversational interface around an established category. This can generate early adoption, but interfaces are easy to copy. A stronger strategy is to identify a workflow that traditional software handles poorly and redesign it from beginning to end.
Founders should ask:
What work remains manual despite widespread software adoption? Where do employees copy information between systems? Which decisions require reviewing large amounts of unstructured information? Where do delays create financial losses? Which processes depend on email, spreadsheets, or repetitive coordination? Which tasks require expensive professional labor? Where are outcomes measurable? Which workflows contain enough recurring volume to support automation? Which actions can safely be delegated to an AI system? Where can a product acquire proprietary feedback data? The best AI opportunities may not resemble traditional software categories.
An AI company may combine elements of:
Software Outsourcing Consulting Labor Data services Payments Communications Marketplaces Managed operations This creates a new competitive category sometimes described as service-as-software. Instead of providing tools for a worker, the company delivers some portion of the worker’s output.
9. The Model Layer May Become Both Strategic and Commoditized
The model layer presents a paradox. Foundation models are among the most technically advanced and capital-intensive products in the economy. They are strategically important because applications depend on their intelligence. At the same time, many model capabilities may become increasingly commoditized.
Several forces drive commoditization:
Multiple well-funded providers Open-source models Falling inference costs Better optimization Model distillation Specialized smaller models Customer ability to switch providers Multimodel routing Rapid replication of popular features Increasingly capable local models Stanford’s AI Index reported that private investment in generative AI reached $33.9 billion in 2024, representing more than one-fifth of private AI investment. Large capital inflows can accelerate innovation, but they also create intense competition.
Model providers may pursue several strategies to preserve value. Become the Intelligence Platform A provider can become the default environment where developers build, test, deploy, monitor, and manage AI applications. Own Distribution Models embedded in widely used productivity software, search engines, operating systems, devices, cloud platforms, or social networks gain immediate access to customers. Build an Ecosystem Developers, agents, plugins, integrations, marketplaces, and enterprise partners can create switching costs. Specialize A model may dominate a particular field such as medicine, law, science, coding, robotics, defense, or finance. Control Infrastructure A company that combines models with highly efficient compute may deliver better economics. Build Consumer Relationships
A model provider may establish a direct relationship with users through assistants, creative products, education platforms, or personalized agents. The likely outcome is not one universal model controlling every use case. The market may include a small number of major general-purpose platforms alongside many specialized, open, private, and on-device models.
10. AI Agents Could Create an Entirely New Infrastructure Economy
The next stage of AI moves beyond chatbots and copilots toward agents. An agent can receive a goal, plan steps, use software tools, access data, communicate with people, and perform actions within defined permissions. As agents become more capable, they will require infrastructure similar to that used by human workers and businesses.
This may include:
Identity Authentication Email accounts Phone numbers Messaging systems Payment methods Bank access Software accounts Cloud computers Browsers API credentials Approval systems
Audit trails Reputation records Insurance Compliance controls Data permissions Secure memory Scheduling systems Monitoring and alerting Agent-to-agent communication This creates an opportunity for an “AWS for agents,” a modular infrastructure layer that allows businesses to deploy autonomous digital workers safely. The company controlling agent identity, permissions, communications, financial access, and operational monitoring could occupy an extremely valuable position. It may become the control plane through which autonomous work is authorized and governed.
The biggest opportunity may therefore not be a single agent application. It may be the infrastructure that millions of agents require to participate in the economy.
11. Where Value May Accumulate Across the AI Stack
The AI economy can be divided into several layers, each with a different path to value capture. Layer 1: Energy, Data Centers, and Physical Infrastructure Value may accumulate where supply is scarce and difficult to expand.
Defensibility can come from:
Power access Land Permits Capital Engineering expertise Supply agreements Scale Operational reliability Layer 2: Semiconductors and Computing Systems Value may accumulate around superior performance, software ecosystems, manufacturing capacity, and customer lock-in. Layer 3: Cloud and Compute Platforms Value may accumulate through scale, utilization, bundled services, enterprise relationships, and developer ecosystems.
Layer 4: Foundation Models Value may accumulate through capability, efficiency, distribution, proprietary data, trust, and platform adoption. Layer 5: AI Development and Operations Value may accumulate in tools that make models reliable, secure, measurable, governable, and economical in production. Layer 6: Horizontal Applications Products serving broad functions such as productivity, coding, search, sales, marketing, finance, and customer service can become very large but will face heavy competition from incumbents and model providers. Layer 7: Vertical Applications Industry-specific platforms may build stronger moats through domain knowledge, regulation, proprietary workflows, and specialized data. Layer 8: Autonomous Agents and Digital Labor Value may accumulate with companies that can deliver dependable outcomes and coordinate complex work. Layer 9: Agent Infrastructure and Marketplaces Identity, payments, communication, verification, discovery, reputation, security, and governance may become essential rails for the agentic economy.
No single layer is guaranteed to capture all the value. The cloud era produced major companies across infrastructure, data, cybersecurity, developer tooling, and applications. AI is likely to do the same, but with even more interaction between layers.
12. Strategic Lessons for AI Founders
Founders building in the AI era should consider several principles. Do Not Build Around a Temporary Model Limitation A startup may solve a problem that disappears when the underlying model improves. Founders must ask whether their value increases or decreases as models become more capable. The strongest products benefit from better models without becoming unnecessary. Control the Customer Relationship Dependence on another platform for distribution creates vulnerability. Direct customer relationships produce better feedback, stronger branding, and more pricing power. Become Embedded in the Workflow Usage frequency and workflow depth matter more than novelty. A product that completes a critical daily task is more defensible than one used occasionally for experimentation. Measure Outcomes
AI systems are probabilistic. Companies need evaluation frameworks that measure accuracy, business impact, failure rates, safety, and customer outcomes. Design for Multiple Models Model independence can protect margins and reduce platform risk. Products should be able to route workloads based on capability, cost, latency, privacy, and reliability. Build Trust Early Security, auditability, privacy, human oversight, and permission systems should be part of the product architecture. They are difficult to add after deployment. Understand Gross Margins Founders must track inference costs, human-review costs, customer-specific implementation, support, and infrastructure usage. Revenue growth without economic discipline can conceal an unsustainable product. Accumulate Proprietary Advantages
Every customer interaction should strengthen at least one of the following:
Data Workflow knowledge Distribution Brand Network effects Integrations Product performance Regulatory expertise Operational scale Sell Transformation, Not Artificial Intelligence Customers rarely purchase AI for its own sake.
They purchase:
Lower costs Faster service More revenue Better decisions Reduced risk Improved customer experiences Higher employee productivity Completed work The sales message should focus on the business result.
13. Strategic Lessons for Established Companies
Incumbents should not treat AI as a decorative product feature. They should examine which parts of their business could be replaced by intelligent systems. Cannibalize Before a Competitor Does A company may need to automate activities that currently generate service revenue or justify premium software seats. Protecting the old model can create an opening for an AI-native competitor. Convert Proprietary Data into Customer Value Possessing data is not enough. The company must use it to produce predictions, automate decisions, personalize experiences, or reduce risk. Redesign Pricing Seat-based pricing may become less suitable when AI reduces the number of humans required to complete a process.
Companies may need:
Consumption pricing Outcome pricing Transaction pricing Hybrid subscriptions Agent-based pricing Shared-savings arrangements Rebuild Workflows Adding a chat interface to a legacy platform may improve accessibility, but it will not protect the company if a competitor automates the complete process. Open the Platform Incumbents can strengthen their position by allowing third-party developers and agents to build on their data, workflows, and distribution. Protect Trust Existing brands may possess a major advantage in regulated and mission-critical markets.
They should reinforce this advantage through responsible deployment, governance, transparency, and security.
14. Strategic Lessons for Investors
Investors evaluating AI companies should look beyond rapid user growth and impressive demonstrations.
Important questions include:
Does the product solve a valuable recurring problem? Who controls the customer relationship? How dependent is the business on one model provider? What happens when model prices fall? What happens when model capability improves? Can an incumbent reproduce the product? Does usage create proprietary data? Are outcomes measurable? What are the true gross margins after inference and human review? Is the product becoming more embedded over time? Does the company own an essential workflow? Can the product expand into adjacent services?
Is there a credible path to trust and compliance? Does the company benefit from multimodel competition? Is the technology advantage durable or temporary? AI companies should not be valued solely as software companies when their cost structures resemble services, infrastructure, or labor businesses. Investors must examine the underlying economics rather than relying on the appearance of software.
15. The Consumer and the Enterprise May Capture More Value Than Any Single Vendor
The largest long-term beneficiary of a platform shift is often the customer. Competition encourages providers to deliver more capability at lower cost. Cloud computing allowed startups to access infrastructure that once required enormous capital investment. Open-source software gave developers sophisticated tools without building everything internally. Smartphones placed computing, communication, navigation, media, commerce, and photography in billions of hands. AI can extend this process by making expertise more widely available.
Individuals may gain access to:
Personalized education Administrative assistance Financial guidance Translation Creative tools Software development Research assistance Healthcare navigation Legal information Career support
Enterprises may gain:
Automated operations Faster research More productive employees Better customer service Lower software-development costs Improved fraud detection Faster product design More accurate forecasting Reduced administrative overhead The economic value created for users may exceed the revenue captured by AI vendors. This does not weaken the business opportunity. It expands adoption. When customers receive significantly more value than they pay for, demand grows, new use cases emerge, and the overall market becomes larger.
Key Takeaways
AI is a platform shift, not merely a software feature. It changes software from a passive tool into a system capable of prediction, communication, and execution. Technology transitions are often positive-sum. Incumbents can grow while startups create new categories and customers gain new capabilities. Infrastructure can create enormous winners. Chips, data centers, cloud platforms, models, middleware, security, and agent infrastructure may capture substantial value. Infrastructure revenue does not guarantee attractive economics. Capital requirements, price competition, utilization, depreciation, and vertical integration can reduce profitability. Basic AI applications are becoming easy to build. Thin wrappers around third-party models face rapid commoditization. Workflow ownership is more defensible than interface ownership. Products embedded in critical business processes are difficult to replace. Systems of record may evolve into systems of prediction and execution. The largest AI companies may perform work rather than merely organize information. Incumbents possess data, customers, integrations, trust, and distribution. These are major advantages, but they must redesign workflows rather than simply add chatbots. Startups win by changing the operating model. They should attack neglected, manual, expensive, or fragmented workflows. Proprietary data matters only when it is exclusive, useful, lawful, and connected to measurable outcomes. Model access alone is not a durable moat. Companies need distribution, workflow depth, data, trust, network effects, or superior economics.
AI agents will require a new infrastructure layer. Identity, communications, payments, permissions, memory, monitoring, and governance may become major markets. Pricing will evolve. Seat-based subscriptions may give way to usage, transactions, agents, outcomes, or shared savings. Companies must understand unit economics early. Inference and human-review costs can weaken margins as usage increases. Customers may capture the largest share of total economic benefit. Better and cheaper intelligence can increase productivity and expand access to expertise.
Frequently Asked Questions
What does “value capture” mean in artificial intelligence?
Value capture refers to the portion of the economic benefit that a company converts into revenue, profit, market power, or lasting enterprise value. An AI product may create enormous value for customers but capture little of it if competition is intense, switching is easy, or the underlying capability becomes commoditized.
Will infrastructure companies capture most of the value?
Infrastructure companies are likely to capture substantial value because AI requires compute, chips, data centers, cloud services, networking, models, and developer tooling. However, application companies that own important workflows, distribution, or customer outcomes can also become extremely valuable. The result will probably be distributed across several layers.
Are foundation models becoming commodities?
Some model capabilities are becoming more widely available and less expensive. This does not mean all models will be identical. Leading providers may retain advantages in advanced reasoning, multimodal performance, efficiency, enterprise trust, developer ecosystems, or specialized capabilities. For many ordinary tasks, however, customers may be able to choose among multiple sufficiently capable models.
What is an AI wrapper?
An AI wrapper is an application that provides a relatively simple interface around a third-party model without adding substantial proprietary technology, workflow integration, data, or operational capability. A wrapper can still become useful, but it needs to develop deeper advantages to remain defensible.
Can incumbent software companies win the AI era?
Yes. Incumbents possess customer relationships, data, integrations, distribution, trusted brands, and systems of record. They may lose when they protect existing workflows instead of redesigning them or when their organizational structure prevents rapid experimentation.
Why are vertical AI applications attractive?
Vertical applications serve specific industries such as healthcare, law, construction, insurance, finance, manufacturing, or logistics. They can build moats through domain expertise, regulatory knowledge, specialized data, industry integrations, and workflow depth.
What is the difference between a copilot and an agent?
A copilot assists a human by generating suggestions, summaries, or content. An agent can pursue a goal, plan actions, use tools, interact with systems, and complete tasks with varying levels of autonomy. The distinction is not always precise, but agents generally possess greater ability to act.
How can an AI startup protect itself from model providers?
It can remain model-independent, control distribution, own customer workflows, accumulate proprietary data, develop specialized integrations, build a trusted brand, and deliver outcomes that require more than model access.
Will AI replace SaaS?
AI is unlikely to eliminate all SaaS, but it may change how software is designed, purchased, priced, and used. Some SaaS applications will become intelligent platforms. Others may be replaced by agents or outcome-based services.
Why might seat-based pricing decline?
If an AI system allows fewer employees to complete more work, charging according to the number of human users may no longer reflect the value delivered. Vendors may shift toward consumption, completed work, transactions, or business outcomes.
What is the biggest opportunity in AI infrastructure?
There is no single guaranteed winner. Promising areas include compute optimization, agent identity, secure execution, AI evaluation, governance, model routing, observability, data infrastructure, cybersecurity, memory systems, payments, and agent communications.
What is the greatest risk for AI application companies?
The greatest risk is building a product whose differentiating capability becomes a standard feature of an underlying model, cloud platform, or incumbent software suite.
What should enterprises prioritize when adopting AI?
Enterprises should prioritize valuable workflows, clear outcome measurement, data readiness, security, governance, integration, human oversight, and realistic unit economics. They should avoid deploying AI merely to demonstrate that they are participating in the trend.
Conclusion: The Winners Will Control More Than Intelligence Artificial intelligence will create enormous economic value, but access to intelligence will not be enough to guarantee durable success. Models will improve. Infrastructure will expand. Prices will fall. Open-source alternatives will mature. Capabilities that appear extraordinary today will become standard tomorrow. The long-term winners will control something that remains scarce. That scarce asset may be proprietary data, trusted distribution, a difficult workflow, customer relationships, physical infrastructure, regulatory permission, a developer ecosystem, a network, superior economics, or the authority to execute work. Cloud history shows that platform shifts do not create only one type of winner. They create foundational infrastructure companies, category-defining applications, powerful developer platforms, specialized vertical vendors, and revitalized incumbents. They also destroy companies that confuse temporary adoption with permanent advantage. The AI era is likely to follow the same broad pattern, but at greater speed and across a much larger portion of the economy.
The most important strategic question is therefore not:
“Where can artificial intelligence be added?”
It is:
“Which essential layer, workflow, relationship, or economic outcome can this company control as intelligence becomes abundant?” The businesses that answer that question convincingly will not merely participate in the AI economy. They will determine how value moves through it.
Relevant Articles and Resources
1. The Race to Capture Value: Cloud Lessons for the AI Era
The original Andreessen Horowitz analysis examining how the cloud and SaaS transition may help explain value creation, incumbent adaptation, infrastructure development, and defensibility in the AI market.
2. Stanford AI Index Report
A comprehensive research resource covering AI investment, model development, adoption, technical performance, policy, and industry trends.
3. Microsoft 2025 Annual Report
A useful primary source for understanding how a major cloud and enterprise-software provider is integrating AI across infrastructure, models, developer tools, productivity software, and business applications.
4. Alphabet Investor Relations and Earnings Materials
Primary-source information about Google Cloud, Alphabet’s full-stack AI strategy, AI infrastructure investments, model deployment, and integration across consumer and enterprise products.
5. Stanford AI Index: Economy
Research and statistics related to private investment, enterprise adoption, workforce effects, and the economic development of artificial intelligence.
6. Stanford AI Index: Research and Development
Data concerning foundation-model development, open-source participation, AI research output, and the changing competitive landscape at the model layer.