The Great AI Paradox: A Falling Price Can Produce a Rising Bill
The great AI paradox is that a rapidly falling price can produce a rapidly rising bill.
At first glance, lower computing costs might appear negative for the companies supplying artificial-intelligence infrastructure. If the cost of producing an AI response keeps declining, shouldn't the industry eventually need less capital?
History often suggests the opposite.
When the price of a useful technology falls, people usually discover more ways to use it. Demand expands, new applications become economical, and total consumption can grow much faster than the unit price declines. Artificial intelligence appears to be following this pattern.
The result is an important investment paradox: AI can become dramatically cheaper to use while the world's aggregate spending on AI continues to rise.
FROM EXPENSIVE EXPERIMENT TO EVERYDAY UTILITY
The cost of running AI models has been falling as chips improve, software becomes more efficient, and competition increases. Tasks that were prohibitively expensive a few years ago can now be performed at a fraction of the cost.
That decline makes AI accessible to more companies, more employees, and more applications. A business that once used AI for a limited pilot may deploy it across customer service, software development, sales, compliance, research, and operations. A consumer may interact with AI throughout the day without consciously opening a dedicated AI product.
Lower costs also make it economical to apply AI to smaller decisions. Instead of reserving intelligence for a few high-value tasks, companies can use it continuously—to review transactions, monitor equipment, summarize communications, personalize services, or optimize schedules.
Each individual interaction becomes less expensive. The number of interactions can rise exponentially.
AGENTS CHANGE THE CONSUMPTION EQUATION
AI agents intensify this dynamic. A conventional chatbot normally waits for a person to ask a question. An agent may complete a multi-step assignment, call several tools, consult different data sources, evaluate its own output, and communicate with other agents.
One human request can therefore generate many machine interactions behind the scenes. As businesses deploy fleets of agents, machines may increasingly become major consumers of AI resources in their own right.
This resembles the evolution of the internet. Early network traffic was driven primarily by people browsing pages. Later, software services began communicating constantly with one another. The same transition may be beginning in AI: from people occasionally prompting models to machines continuously using intelligence as part of ordinary operations.
If that happens, demand for computing, storage, memory, and networking may be driven less by the number of human users and more by the amount of work delegated to software.
THE PHYSICAL FOUNDATION BEHIND DIGITAL INTELLIGENCE
AI may feel weightless, but its infrastructure is intensely physical.
Every AI workload ultimately depends on chips housed in a data center, connected through networks, supported by cooling systems, and supplied with reliable electricity. Expanding that capacity requires land, buildings, engineering, construction labor, power equipment, water or alternative cooling technologies, and access to the electrical grid.
This is why the AI investment cycle can benefit a much broader group of industries than software and semiconductors alone. Potential beneficiaries include utilities and independent power producers, electrical-equipment manufacturers, data-center operators, cooling specialists, networking companies, memory and storage providers, and the construction and engineering firms that build the facilities.
Semiconductors remain essential, but they sit inside a much larger system. The most advanced processor has limited value if a data center cannot secure enough power, connect to the grid, manage heat, obtain memory, or move data quickly.
The constraints can therefore migrate. At one stage, the limiting factor may be advanced chips. At another, it may be memory, transformers, skilled labor, available land, or electricity. Understanding where the bottleneck is moving can be as important as forecasting overall AI demand.
CYBERSECURITY BECOMES PART OF THE INFRASTRUCTURE
As AI systems gain the ability to take action, cybersecurity becomes a central part of the build-out rather than an optional software layer.
An agent that can access corporate information, initiate a payment, change a customer account, or control equipment presents a different risk from a chatbot that simply drafts text. Businesses must verify the agent's identity, define its permissions, monitor its behavior, and prevent it from being manipulated.
That creates demand for new forms of identity management, data protection, observability, governance, and threat detection. It also creates opportunities for established security providers to extend their platforms into the agentic world.
In an AI economy, trust may become as valuable as intelligence.
MORE SPENDING DOES NOT GUARANTEE BETTER RETURNS
The strength of the infrastructure cycle should not obscure the risks.
Building data centers requires enormous upfront capital, and a growing portion of that investment may be financed with debt. The long-term economics depend on sustained demand, pricing, utilization rates, energy availability, and the willingness of enterprise customers to pay for AI services.
At the corporate level, many organizations are still learning how to convert AI usage into measurable financial results. Lower prices can encourage experimentation, but experimentation is not the same as durable profitability. Companies will increasingly scrutinize AI budgets and demand evidence of higher productivity, stronger revenue, better customer outcomes, or lower operating costs.
This should create a widening gap between providers that solve a valuable business problem and those offering a generic feature without a defensible advantage.
Investors should also watch concentration, regulation, geopolitics, supply constraints, and the possibility that infrastructure is built faster than revenue develops. A powerful secular trend can still contain periods of excess enthusiasm and sharp market corrections.
LOOKING BEYOND THE OBVIOUS WINNERS
The central investment question is no longer simply, “Which company has the best AI model?” It is increasingly, “Which businesses will capture the spending required to make AI useful everywhere?”
That question leads naturally to a broader approach. Model developers and semiconductor leaders remain important, but so do the companies providing power, cooling, data centers, communications, cybersecurity, and enterprise integration. Industries using AI to improve productivity may also become beneficiaries, even if they are not traditionally considered technology companies.
During volatile markets, this breadth creates room for sector and industry rotation. Exposure can shift among the builders of AI, the suppliers supporting them, and the businesses applying AI successfully. It also helps avoid treating a complex, multi-year transformation as a narrow momentum trade.
The price of intelligence may continue to fall. But if lower prices turn AI into an everyday utility—used constantly by people, businesses, and machines—the total market surrounding it may become much larger.
Cheaper AI does not necessarily mean a smaller opportunity. It may be precisely what makes the opportunity so large.
At first glance, lower computing costs might appear negative for the companies supplying artificial-intelligence infrastructure. If the cost of producing an AI response keeps declining, shouldn't the industry eventually need less capital?
History often suggests the opposite.
When the price of a useful technology falls, people usually discover more ways to use it. Demand expands, new applications become economical, and total consumption can grow much faster than the unit price declines. Artificial intelligence appears to be following this pattern.
The result is an important investment paradox: AI can become dramatically cheaper to use while the world's aggregate spending on AI continues to rise.
FROM EXPENSIVE EXPERIMENT TO EVERYDAY UTILITY
The cost of running AI models has been falling as chips improve, software becomes more efficient, and competition increases. Tasks that were prohibitively expensive a few years ago can now be performed at a fraction of the cost.
That decline makes AI accessible to more companies, more employees, and more applications. A business that once used AI for a limited pilot may deploy it across customer service, software development, sales, compliance, research, and operations. A consumer may interact with AI throughout the day without consciously opening a dedicated AI product.
Lower costs also make it economical to apply AI to smaller decisions. Instead of reserving intelligence for a few high-value tasks, companies can use it continuously—to review transactions, monitor equipment, summarize communications, personalize services, or optimize schedules.
Each individual interaction becomes less expensive. The number of interactions can rise exponentially.
AGENTS CHANGE THE CONSUMPTION EQUATION
AI agents intensify this dynamic. A conventional chatbot normally waits for a person to ask a question. An agent may complete a multi-step assignment, call several tools, consult different data sources, evaluate its own output, and communicate with other agents.
One human request can therefore generate many machine interactions behind the scenes. As businesses deploy fleets of agents, machines may increasingly become major consumers of AI resources in their own right.
This resembles the evolution of the internet. Early network traffic was driven primarily by people browsing pages. Later, software services began communicating constantly with one another. The same transition may be beginning in AI: from people occasionally prompting models to machines continuously using intelligence as part of ordinary operations.
If that happens, demand for computing, storage, memory, and networking may be driven less by the number of human users and more by the amount of work delegated to software.
THE PHYSICAL FOUNDATION BEHIND DIGITAL INTELLIGENCE
AI may feel weightless, but its infrastructure is intensely physical.
Every AI workload ultimately depends on chips housed in a data center, connected through networks, supported by cooling systems, and supplied with reliable electricity. Expanding that capacity requires land, buildings, engineering, construction labor, power equipment, water or alternative cooling technologies, and access to the electrical grid.
This is why the AI investment cycle can benefit a much broader group of industries than software and semiconductors alone. Potential beneficiaries include utilities and independent power producers, electrical-equipment manufacturers, data-center operators, cooling specialists, networking companies, memory and storage providers, and the construction and engineering firms that build the facilities.
Semiconductors remain essential, but they sit inside a much larger system. The most advanced processor has limited value if a data center cannot secure enough power, connect to the grid, manage heat, obtain memory, or move data quickly.
The constraints can therefore migrate. At one stage, the limiting factor may be advanced chips. At another, it may be memory, transformers, skilled labor, available land, or electricity. Understanding where the bottleneck is moving can be as important as forecasting overall AI demand.
CYBERSECURITY BECOMES PART OF THE INFRASTRUCTURE
As AI systems gain the ability to take action, cybersecurity becomes a central part of the build-out rather than an optional software layer.
An agent that can access corporate information, initiate a payment, change a customer account, or control equipment presents a different risk from a chatbot that simply drafts text. Businesses must verify the agent's identity, define its permissions, monitor its behavior, and prevent it from being manipulated.
That creates demand for new forms of identity management, data protection, observability, governance, and threat detection. It also creates opportunities for established security providers to extend their platforms into the agentic world.
In an AI economy, trust may become as valuable as intelligence.
MORE SPENDING DOES NOT GUARANTEE BETTER RETURNS
The strength of the infrastructure cycle should not obscure the risks.
Building data centers requires enormous upfront capital, and a growing portion of that investment may be financed with debt. The long-term economics depend on sustained demand, pricing, utilization rates, energy availability, and the willingness of enterprise customers to pay for AI services.
At the corporate level, many organizations are still learning how to convert AI usage into measurable financial results. Lower prices can encourage experimentation, but experimentation is not the same as durable profitability. Companies will increasingly scrutinize AI budgets and demand evidence of higher productivity, stronger revenue, better customer outcomes, or lower operating costs.
This should create a widening gap between providers that solve a valuable business problem and those offering a generic feature without a defensible advantage.
Investors should also watch concentration, regulation, geopolitics, supply constraints, and the possibility that infrastructure is built faster than revenue develops. A powerful secular trend can still contain periods of excess enthusiasm and sharp market corrections.
LOOKING BEYOND THE OBVIOUS WINNERS
The central investment question is no longer simply, “Which company has the best AI model?” It is increasingly, “Which businesses will capture the spending required to make AI useful everywhere?”
That question leads naturally to a broader approach. Model developers and semiconductor leaders remain important, but so do the companies providing power, cooling, data centers, communications, cybersecurity, and enterprise integration. Industries using AI to improve productivity may also become beneficiaries, even if they are not traditionally considered technology companies.
During volatile markets, this breadth creates room for sector and industry rotation. Exposure can shift among the builders of AI, the suppliers supporting them, and the businesses applying AI successfully. It also helps avoid treating a complex, multi-year transformation as a narrow momentum trade.
The price of intelligence may continue to fall. But if lower prices turn AI into an everyday utility—used constantly by people, businesses, and machines—the total market surrounding it may become much larger.
Cheaper AI does not necessarily mean a smaller opportunity. It may be precisely what makes the opportunity so large.
