If intelligence gets cheap, what stays scarce?

Cheap answers leave expensive work around them

Start with a customer refund. Generating a plausible recommendation is one task. Establishing the customer’s identity, finding the purchase, checking the policy, authorizing payment and resolving an exception are others. The delivered outcome depends on the whole sequence.

The economics therefore extend beyond inference. They include information access, integration, verification and the cost of mistakes. As generation becomes cheaper, those surrounding activities account for more of the remaining cost. Our view is that this favors businesses that can execute reliably within an operating system. A supplier of generic answers has a weaker claim on the resulting value.

We would not turn that argument into a comforting prediction about human expertise. In the customer-support study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, AI assistance increased problems resolved per agent-hour by roughly 14%, with particularly strong gains for less-experienced workers. Knowledge previously associated with experience became transferable through the tool. [3]

Coordination is also contestable. In an experiment involving Procter & Gamble professionals, individuals using AI performed 0.37 standard deviations better than individuals working without it. Some benefits of collaboration can be reproduced by a machine counterpart. [4]

Yet access to a capable model and adoption throughout a business remain different stages. A Census survey covering November 2025–January 2026 found AI use in a business function at 18% of firms, rising to 32% when weighted by employment. Among users, 66% relied on AI solely to augment tasks. We read this as evidence of incomplete diffusion and substantial implementation work, rather than a permanent ceiling on automation. [5]

The scarce capability is therefore conditional: delivering a useful result through the organization as it actually exists. AI may eventually remove more of that friction. Until then, the owner of the surrounding workflow has an opportunity to collect some of the savings.

Abundant content raises the stakes for customer access

A completed result still needs a customer. When producing another article, image or recommendation becomes inexpensive, access to someone willing to consider it becomes relatively more valuable. Human attention does not expand at the speed of model output.

The commercial pool is already substantial. U.S. internet-advertising revenue reached $295 billion in 2025, an increase of $36 billion, or 13.9%, from the previous year. That is advertiser expenditure on reaching audiences; its growth also reflects formats, targeting and the wider advertising market. [6]

Pew’s study of Google searches shows how AI can alter who benefits from that access. Users clicked a traditional search result in 8% of visits where an AI summary appeared, compared with 15% where one did not. The observational comparison is consistent with more of the interaction remaining at the answer interface, although the searches in the two groups differ. [7]

Source: Pew Research Center. Note: Observed Google searches in March 2025; the groups differ in whether an AI summary appeared. [7]

The economic tension is straightforward. A publisher can contribute useful information while another business controls the customer encounter. More valuable information need not translate into more valuable distribution for its original producer.

Our view is that established audience owners have demonstrated substantial monetization, while the durability of any particular interface remains open. Google Search & other generated $63.3 billion in Q2 2026 revenue, against $54.2 billion a year earlier. The franchise continued to grow as AI search expanded. [8]

But finite attention alone offers no guarantee of rising profits. Meta’s Q2 advertising revenue increased from $46.6 billion to $59.4 billion, or 27.5%, while consolidated operating income fell from $20.4 billion to $18.8 billion, an 8.2% decline. Those growth rates come directly from the year-over-year reported figures. Distribution can remain valuable while the cost of defending and expanding it absorbs the gains. [9]

Agents could change the allocation again. If customers delegate discovery and purchasing to an independent assistant, that assistant may acquire the relationship on which today’s platforms depend. We have greater confidence that customer access remains scarce than that its current owners keep it.

Authority can survive the automation of judgment

Attention determines who gets considered. Authority determines who can act. A recommendation does not itself confer permission to transfer money, change a business record or submit a legal filing. Those rights arise from contracts, institutions and customer delegation.

Two cases define the boundary. Lemonade reported that roughly 55% of its claims were automated from start to finish at the end of 2025. That is substantial automation of a consequential process inside an insurer. [10]

In Mata v. Avianca, the court’s sanctions order put responsibility for checking legal filings on the attorneys: “existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings.” Generating the material did not transfer that responsibility to the tool. [11]

Together, these cases support a narrower and more durable proposition than the claim that every important decision requires human review. Individual interventions can disappear while institutional responsibility remains. Permission can be delegated to software; the allocation of losses and authority still matters.

We therefore see selective value in systems that control authenticated access, transaction records, permissions and exceptions. Salesforce’s proposed integration with Claude illustrates the structure. Users could explore ideas through Claude while Agentforce provides the governed execution layer within Salesforce. The announced arrangement describes a strategy, with some functionality still prospective. [12]

The conversational interface can move without the underlying operating system being replaced. Whether that separation preserves pricing power depends on the difficulty of substituting the execution layer and on which party owns the customer relationship.

Useful context has a stronger claim than a generic data moat

Permission is more valuable when it is paired with information needed to make a good decision. But “proprietary data” is too broad a category to carry an investment thesis. The relevant questions are whether the information is current, legally usable, difficult to substitute and connected to something customers pay to accomplish.

RELX provides evidence of a functioning information-and-workflow franchise. In the first half of 2026, it reported 7% underlying revenue growth and an adjusted operating margin rising from 34.8% to 35.5%. Thomson Reuters’ Legal, Corporates, and Tax & Accounting businesses generated $1.62 billion in Q2 revenue, against $1.46 billion a year earlier. These businesses were growing while AI capability became more widely available. [13] [14]

Legal rights can reinforce the position, but their scope matters. In the Ross Intelligence litigation, the district court found infringement of 2,243 Thomson Reuters editorial headnotes. The protected asset was authored editorial material; the judgment leaves the underlying judicial opinions and legal facts as a separate question. [15]

Workflow vendors also have tangible adoption evidence. ServiceNow reported $3.88 billion in Q2 2026 subscription revenue, up 24.5%, and AI annual contract value above $1 billion. Microsoft disclosed more than 30 million paid Microsoft 365 Copilot seats in its fiscal fourth quarter. Those are different commercial measures, but both show AI being sold through established customer relationships. [16] [17]

Microsoft chief executive Satya Nadella put the competitive argument plainly: “And the models are an input, not some extraction of the knowledge of the enterprise.” The strongest version of that argument is that the application retains customer context and can purchase improving intelligence from competing suppliers. Our view is that strong workflow owners can retain some of those savings, provided customers continue to value the surrounding product. [17]

The unit economics explain the opportunity. Salesforce’s published Agentforce tariff charges $0.10 per action. At GPT-5’s published token prices, a range of assumed input and output volumes produces model expense of $0.0075–$0.075 per action, leaving $0.025–$0.0925 before the other costs of delivery. The assumptions below span a tenfold difference in token use because action complexity is central to the economics. [18] [19]

Lower model expense leaves more room to pay for integration, exceptions and service—if the customer price holds.

Sources: Salesforce and OpenAI. Note: Assumed token volumes; GPT-5 list prices of $1.25 per million input tokens and $10 per million output tokens. Expense equals input plus output charges; spread equals $0.10 less model expense. Retrieval, retries and support are additional costs. [18] [19]

A further tenfold reduction in token prices would leave almost all of the action tariff available for other costs and profit. The crucial commercial assumption is that the action price holds. Procurement pressure, rival applications and agents that bypass the product could give the savings to customers instead.

That is why our preference is selective. We favor control of valuable workflows over a blanket claim that enterprise software benefits. Seat-based pricing can weaken, interfaces can migrate, and information can become substitutable. The business must continue to earn its place in the transaction.

Connected power is the physical constraint that travels slowly

Software can rearrange a workflow quickly. Delivering electricity to a particular site requires a different sequence of decisions and construction. Generation, transmission, substations, equipment and operating permission must line up at the same place and time.

The IEA’s central projection puts global electricity consumption by all data centers at 950 TWh in 2030, against 485 TWh in 2025. Dividing annual energy by 8,760 hours implies average draw rising from about 55 GW to 108 GW: 53 GW of additional average demand, with electricity use growing approximately 14.4% annually. The scope includes non-AI computing and cooling. [20]

The practical constraint is local delivery. JLL reports average grid-connection waits exceeding four years in primary markets; the IEA puts new transmission development in advanced economies at four to eight years. A data center itself can become operational in two to three years. The construction clocks do not naturally align. [2] [21]

CoreWeave’s disclosures make the distinction visible. At the end of Q2 2026, it reported 1.5 GW of active power and approximately 3.7 GW contracted. Dividing the two gives an active share of roughly 41%. The remaining 2.2 GW was the gap between its contracted pipeline and active capacity at that date. [22]

Source: CoreWeave. Note: Company-reported power measures at June 30, 2026; contracted capacity includes the active portion. [22]

Permission can bind even beside an existing power plant. In the Talen–Amazon Susquehanna case, FERC rejected an interconnection-agreement amendment expanding the arrangement from 300 MW to 480 MW. The original arrangement remained distinct from the proposed expansion. A nearby generator did not automatically supply the right to increase the connection. [23]

Our view is that delivery certainty has value: an already powered, permitted site can offer something that ordinary land and a future equipment order cannot. Qualified electrical equipment and difficult-to-replace infrastructure also have a stronger claim than undifferentiated construction capacity. Eaton’s companywide backlog was up 33% at June 2026, supporting the evidence of equipment demand. [24]

We regard this as a strong constraint through the current buildout, with a finite investment horizon. GE Vernova targets annual turbine manufacturing output of 20 GW in the third quarter of 2026 and 24 GW in 2028, with actions toward 30 GW in 2030. More equipment should eventually contest today’s scarcity, even as connection and construction requirements continue to delay particular projects. [25]

The owner funding expansion may keep less than the supplier

Scarcity identifies a potential source of bargaining power. It does not settle the return on capital. The customer buying a scarce input must still earn enough from the finished service to pay for construction, financing, depreciation and replacement.

NVIDIA’s fiscal Q2 2027 results show substantial supplier capture: $96.2 billion in revenue and a 75% GAAP gross margin. That is a strong observed position in the capacity buildout. Its customers face a separate economic test. [26]

Amazon illustrates the financing burden. AWS sales grew 37% to $42.2 billion in Q2 2026, while Amazon’s consolidated trailing free cash flow moved from positive $18.2 billion to negative $7.6 billion. Net purchases of property and equipment increased by $66.1 billion over the corresponding trailing periods, primarily reflecting AI investment. [27]

The cash bridge is revealing. A $25.8 billion deterioration in free cash flow alongside $66.1 billion of additional investment implies a $40.3 billion improvement in cash before that investment. Amazon generated substantially more cash at that stage and still spent more than the increase. The bridge applies to the consolidated company. [27]

Source: Amazon. Note: Trailing years ended June 30; cash gain is the implied change before net property-and-equipment investment: −$25.8bn + $66.1bn. [27]

This is why we are more cautious about the returns of infrastructure buyers than about the existence of demand for suppliers. Utilization, realized service prices and replacement spending determine whether the buyer eventually retains attractive economics.

The demand pipeline itself also needs scrutiny. Dominion’s documentation describes 47 GW of contract capacity as of July 2025 alongside a forecast of 16.6 GW of demand by 2046. Contract face values and expected load represent different stages of the process. Capitalizing the former as delivered demand would overstate the opportunity. [28]

Regulated utilities, equipment manufacturers and owners of powered sites face different mechanisms for collecting value. A longer queue can increase the usefulness of their assets without giving each owner the same ability to raise prices or returns. Our preference is for demonstrable bargaining power and delivery capability, assessed after the capital required to sustain them.

Human presence can remain valuable without earning a premium

The same distinction applies beyond listed technology businesses. Time with a particular person, a desirable location and attendance at a particular event cannot be reproduced simply by generating more content. Relationships, taste and deciding what matters remain central to the value people seek.

But persistence of demand is a weaker claim than pricing power. Live Nation hosted 159 million fans in 2025, up 5%, while reporting flat entry-level U.S. ticket prices. Embodied experiences remained in demand without every ticket becoming more expensive. [29]

Care work makes the point more sharply. The Bureau of Labor Statistics projects 18% employment growth for home-health and personal-care aides from 2025 to 2035, alongside median annual pay of $35,800 in 2025. Work can be needed, situated and relational while workers retain limited bargaining power. [30]

We are therefore skeptical of a broad investment thesis built on an escalating “human premium.” Particular relationships and experiences may retain enormous value. That does not promise uniformly higher wages or margins for whoever provides them.

The strongest threat is that AI removes the surrounding friction too

Our thesis is most vulnerable where the complement is itself a cognitive task. Agents may build replacement applications, migrate information, compare suppliers, coordinate teams and verify work. If those activities become reliable and inexpensive, incumbents can lose the advantages that currently let them charge for integration and context.

That objection deserves substantial weight. The productivity experiments already show expertise and coordination becoming partly substitutable, while Lemonade’s claims automation shows how far an institution can delegate execution. We cannot assume that the remaining steps stay expensive merely because they are expensive today. [3] [4] [10]

The labor evidence is also uneven. A Danish study found average chatbot time savings of 3% and no significant effect on earnings or recorded hours during its study period. Separately, U.S. payroll research associated AI exposure with a 16% relative employment decline among workers aged 22–25, controlling for firm-level shocks. Different populations, methods and periods explain why modest aggregate effects can coexist with pressure on particular entrants; the latter result remains observational. [31] [32]

The physical countercase is equally consequential. Efficiency gains, additional capacity and disappointing customer returns could turn shortages into oversupply. The distinction between contract capacity and expected demand makes that risk concrete. [28]

There is also a boundary to the premise. Familiar capabilities can become inexpensive while frontier reasoning and reliable long-horizon execution continue to command high prices. We would not extrapolate a fixed-benchmark price decline into universal abundance of useful intelligence.

Our confidence is consequently higher in the enduring categories—time, rights, access and physical delivery—than in today’s owners retaining their rents. Competition may pass much of the benefit to users.