The Price of Intelligence Is Falling

For most of the artificial-intelligence boom, we have been watching the wrong price.

We watch the valuations, the funding rounds, the enormous data centers being announced across the world, and the billions committed to chips, electricity and infrastructure. We watch OpenAI and Anthropic release increasingly capable models and ask which one is smarter, faster or more powerful.

But perhaps the number that matters most is moving in the opposite direction: the price of intelligence is falling.

Not intelligence in the philosophical sense, and certainly not human intelligence, but something economically important that until recently was scarce: the ability of software to read, reason, write, analyze, code, compare alternatives and increasingly take actions on our behalf.

The Decline Is Already Visible

The decline is already visible in the economics of frontier AI. OpenAI has cut developer pricing. Anthropic has pushed frontier-level capability toward lower operating costs, with models like Claude Sonnet 5 now priced at just $2 per million input tokens—down sharply from earlier generations. Inexpensive Chinese models and open-weight systems are placing pressure on proprietary providers.

This may sound like good news for AI companies. Cheaper AI should produce more AI usage, just as cheaper computing and bandwidth eventually created applications that could never have existed when those resources were expensive. That is probably true.

But it creates a second question that is much more interesting: if intelligence becomes cheap, where does the economic value go?

That question may explain much more than the competition between OpenAI and Anthropic. It may explain why AI companies are moving into workplace software, why Microsoft is building its own models while simultaneously supporting everyone else’s, why Amazon is turning its shopping assistant into an agent, why Visa and Mastercard are redesigning payment infrastructure for machines, and eventually why some of the most powerful platforms created during the internet era may have to reconsider what business they are actually in.

The AI story may be moving from the economics of intelligence to the economics of agency. And once that happens, the structure of the internet itself begins to look different.

The Commodity Problem

Imagine that the cost of producing a competent unit of machine reasoning continues to fall. The models will still matter enormously. Some will remain better than others. Frontier research will continue to require extraordinary amounts of capital, and specialized models will command premiums. Safety, latency, reliability and context will all matter.

But beneath those differences, something familiar begins to happen: a technology that once appeared magical becomes infrastructure.

Computing followed this path. Storage followed it. Bandwidth followed it. GPS followed it. Digital photography followed it. Each remained extraordinarily important while becoming progressively cheaper per unit. Importance and pricing power are not the same thing.

This creates an uncomfortable problem for the frontier AI laboratories. If customers can increasingly obtain comparable reasoning from several providers, then the model itself begins to resemble a replaceable component.

Microsoft is already designing explicitly around that possibility. Its cloud platform supports models from OpenAI, Anthropic, Mistral, xAI and Microsoft’s own MAI family, and it has been moving toward an architecture in which context, memory, action space and the surrounding software environment are separated from any individual model family. The strategic implication is significant: if the model becomes substitutable, the provider of the model risks becoming merely one component inside a much larger system controlled by someone else.

That may turn out to be one of the defining economic ideas of the AI transition. If the model is substitutable, the model provider must move somewhere else to preserve value. This is why the important competition is no longer simply between GPT and Claude. The competition is about who owns the layer above the model.

From Selling Intelligence to Selling Work

Traditional software assumes something fundamental about the relationship between humans and computers: humans do the work and software gives them tools. Word gives you an empty document. Excel gives you cells. PowerPoint gives you slides. Photoshop gives you a canvas. Salesforce gives you forms, menus and databases. SAP gives you processes. The software may be incredibly sophisticated, but the basic contract remains remarkably consistent: the human operates it.

AI changes that contract.

If I give an agent last year’s accounts and ask for a cash-flow forecast, I do not necessarily want a better spreadsheet interface. I want the forecast. If I provide customer data and ask which clients are likely to leave, I do not necessarily want another CRM dashboard. I want an answer, the reasoning behind it, and perhaps the retention campaign launched automatically. If I ask for a presentation about quarterly performance, the economic value is not really in giving me better slide-editing tools. It is in producing the presentation.

The unit of value therefore begins moving from the tool toward the outcome. This is where an apparently simple decline in token prices becomes strategically important. If OpenAI and Anthropic remain companies that sell increasingly inexpensive intelligence to Microsoft, Salesforce, Amazon or other platforms, those platforms may capture much of the value created by that intelligence. So the laboratories have every incentive to move upward, not merely to answer questions but to perform work and eventually to become the environment in which work happens.

This does not mean Microsoft Office disappears. That would underestimate one of the strongest distribution systems in business history, along with decades of accumulated corporate data, permissions, security infrastructure, identity systems and institutional habit. Microsoft’s response demonstrates precisely why incumbency matters. It can make models interchangeable while preserving the surrounding environment. OpenAI, in turn, needs distribution and infrastructure without becoming merely an ingredient inside Microsoft products. Anthropic faces the same strategic problem. Each side is trying to ensure that it cannot be trapped underneath the other.

That struggle tells us something important about where AI economics may be headed. The model is becoming the engine. Everyone wants to own the car.

Then the Same Logic Leaves the Office

Once you see this pattern in software, it becomes difficult not to notice it elsewhere.

Consider Amazon, Booking.com, Expedia, Airbnb or Uber. These companies are very different businesses, and reducing them to “middlemen” would be misleading. They provide payments, fraud protection, logistics, customer support, insurance, reputation systems, identity, fulfillment, discovery and enormous operating infrastructure. But they also solve a problem that existed because humans have limited attention.

Suppose I want a hotel in Rome. There may be thousands of possible rooms across hotels, apartments, guesthouses and independent properties. I cannot realistically visit every website, verify every operator, compare every cancellation policy, calculate every total price, inspect hundreds of reviews and negotiate directly with twenty owners. So I use an aggregator. Booking.com compresses an impossibly fragmented supply landscape into an interface a human can navigate.

The same is true in retail. I do not want to visit 800 websites to find a pair of headphones. Amazon solves that problem. Uber solves a similar coordination problem between riders and drivers. Food-delivery applications coordinate restaurants, couriers and customers. The platform economy was built partly because human search is expensive.

Now imagine that search is delegated. You tell your personal agent to find a small independent hotel near the center of Florence, quiet, with excellent Wi-Fi, under €220 per night, free cancellation, preferably with a balcony, and to compare direct prices with platform prices while considering your previous travel preferences. A human would find that exhausting. An agent does not. It can theoretically query dozens of suppliers, compare terms, eliminate unsuitable offers, verify policies, negotiate structured discounts and return with a few options.

The economic problem that justified the centralized interface has changed. The old architecture was consumer, then platform, then supplier. The emerging architecture could increasingly become consumer agent communicating directly with supplier agent. That does not mean Booking.com disappears. It means something more interesting: Booking.com has to justify the platform independently of the search box.

The same is true for Amazon. Amazon clearly understands the direction of travel, which is why its shopping assistant has been evolving toward greater autonomy in searching, comparing and purchasing. This is not Amazon surrendering to agents. It is Amazon attempting to become the agent. And that may reveal the next major contest of the internet. The question is no longer who owns the website or even who owns search. The more important question may become: who receives the human’s intent first?

Today, I open Amazon. Tomorrow, perhaps I tell an agent what I need. Whoever owns that relationship can decide which marketplaces, merchants, services and databases are consulted afterward. That is an extraordinary amount of economic power.

The Internet Could Unbundle and Then Re-Aggregate

At first glance, this sounds like decentralization. Millions of independent businesses could become directly addressable. A boutique hotel would no longer need the consumer to discover its website manually. It could publish availability, pricing, amenities and policies in a format agents understand. A small manufacturer could expose inventory through an API. A restaurant could expose tables. A plumber could expose appointment slots. A driver could expose location and availability. A consultant could expose expertise, schedule and price.

Instead of every supplier fighting for visibility on a giant platform, the supplier becomes something closer to a machine-readable economic node. Agents query the network, businesses answer, and transactions occur. The web becomes less like a collection of pages designed to attract human eyes and more like a network of services designed to respond to machine requests.

That would be a profound shift. But we should be careful before celebrating the return of a decentralized internet, because decentralization immediately creates another problem: who decides which supplier the agent trusts?

Suppose my AI finds a hotel I have never heard of offering a room 20 percent cheaper than Booking.com. Is it real? Suppose it finds an online merchant in another country selling the same product for half the Amazon price. Will the product arrive? Is it counterfeit? What happens if the merchant disappears? What data did the agent use? Were the reviews manipulated? Was the price comparison complete? Is the agent recommending this seller because it is genuinely the best choice, because it has some commercial arrangement, or because the merchant optimized itself for whatever ranking mechanism the agent uses?

The disappearance of one intermediary does not eliminate the need for intermediation. It moves it. That is why the agentic economy may produce not simply an unbundling, but a great re-aggregation. Marketplaces may lose some control over discovery while agent ecosystems gain influence over ranking. Search engines may lose some traffic while identity providers become more important. Interfaces may shrink while trust infrastructure grows. Some commissions may fall while new verification, access or transaction fees appear elsewhere.

Economic power rarely vanishes. It migrates.

What Becomes Scarce When Intelligence Is Abundant?

This brings us to what may be the deeper question. If machine intelligence becomes abundant, what remains scarce?

Trust does. Reputation does. Identity does. Permission does. Proprietary data does. Physical infrastructure does. Energy does. Logistics do. Real estate does. Regulatory authorization does. And ultimately human intent does.

Amazon’s warehouses do not disappear because an agent can compare prices. An aircraft still has to move the parcel. A hotel still has to own or operate the room. A restaurant still has to cook dinner. Uber’s software may become easier to reproduce, but a car and driver, or an autonomous vehicle, still have to appear at the right street corner.

The agentic economy therefore does not make the physical world irrelevant. It may make the distinction between information monopolies and physical moats far more visible. Amazon is particularly interesting for this reason. Even if product discovery moves outside Amazon, its logistics network remains enormously valuable. A merchant might sell directly through an agent while still using Amazon fulfillment. In that world, Amazon could lose part of the storefront and preserve part of the infrastructure.

A similar transformation could happen elsewhere. Booking platforms might become distribution, identity, insurance or settlement infrastructure. Payment companies might become authorization networks for agents. Software companies might become systems of record operated invisibly by AI. The interface shrinks, the infrastructure remains, and the battle shifts to deciding who controls the infrastructure underneath the invisible interface.

When Software Can Spend Money

An agent becomes economically different from ordinary software when it can act. And action eventually means payment. This is where the agentic economy stops being a new user interface and becomes something closer to an economic system.

Software has always been capable of triggering transactions under predefined rules. What is different now is the possibility that a probabilistic reasoning system can decide that a transaction should happen, select the counterparty, negotiate conditions and initiate the payment within authority delegated by a human or organization. That requires new financial infrastructure.

Some of it may be built on crypto rails. Coinbase’s x402, for example, revives the HTTP 402 “Payment Required” concept and allows software, including AI agents, to make stablecoin payments directly when requesting an API or online service. A service can effectively respond that the requested information or computation costs a certain amount. The agent evaluates the price, pays and continues automatically.

That is a remarkably machine-native model: no subscription form, no sales call, no employee entering a credit-card number. Software encounters another service, evaluates the price, pays and proceeds.

But it would be a mistake to assume this necessarily means crypto replaces the traditional financial system. Visa and Mastercard are adapting just as aggressively. Both are building infrastructure designed specifically for agent-led commerce, with new concepts around machine identity, authenticated intent and agent-specific payment credentials. In 2025, Mastercard launched its Agent Pay program; Visa followed with Visa Intelligent Commerce, opening its network to AI developers through AI-Ready Cards. By 2026, both networks were actively settling agent transactions and partnering with AI labs to define authentication standards.

That convergence is revealing. The future may not be “crypto versus banks.” It may be something much larger: money itself becoming machine-readable, permissioned and programmable.

Once machines can spend, another scarce resource becomes obvious: authority. Who authorized the agent to spend €500? Could it spend €5,000? Could it accept a non-refundable reservation? Could it renew a contract? Could it purchase from a new supplier? Could it transfer assets? What happens if its reasoning was wrong?

These are not edge cases. They become central design problems once agents move from advising humans to representing them. The economic architecture must therefore contain not only intelligence but identity, mandate, limits, auditability and recourse. The agent has to know not only what it can do. The rest of the economy must know for whom it is acting.

From the Attention Economy to the Intention Economy

For roughly two decades, much of the internet has been organized around capturing human attention. Search engines competed for queries. Social networks competed for time. Marketplaces competed to become the destination we opened first. Advertising became one of the dominant business models because controlling attention meant controlling the route between intention and purchase.

Agents may disrupt that sequence. If I tell an agent what I want and it performs the discovery itself, the number of pages I visit may collapse. I do not need ten blue links; I need the hotel. I do not need twenty product pages; I need the right washing machine delivered on Tuesday. I do not need to navigate four software applications; I need the quarterly report prepared, checked and sent to the board.

That means a substantial part of the internet may migrate from competing for attention to competing for machine interpretation of intention.

Businesses will still need to be discoverable, but discoverability may mean something different. Instead of optimizing a webpage primarily for Google or designing an advertisement to persuade a person, a business may increasingly need to expose trustworthy structured information that an agent can understand: price, availability, location, delivery time, cancellation conditions, identity, reputation, capabilities, warranty, return policy and permissions.

The web does not disappear. It changes audience. For thirty years we have designed much of the commercial internet for people looking at screens. We may now begin designing part of it for software representing people. That is a very different architecture.

The Trillion-Dollar Question

This returns us to the original problem. If intelligence becomes cheaper and frontier models become increasingly interchangeable, how can companies building those models justify enormous valuations and enormous capital expenditures?

One answer is simply scale. Lower prices could produce such a dramatic increase in consumption that total revenue continues growing despite declining unit economics. That has happened repeatedly in technology.

But there is another possibility: the laboratories do not intend to remain laboratories.

They move from tokens to subscriptions, from subscriptions to enterprise workflows, from workflows to agents, from agents to transactions, from transactions to marketplaces, and from marketplaces to infrastructure. At each step, they move closer to economic activity itself.

The prize is no longer merely being paid every time a model produces a token. The prize is becoming the system through which people and companies express intentions and convert those intentions into actions. Microsoft understands this. Amazon understands it. Visa understands it. Mastercard understands it. OpenAI and Anthropic certainly understand it.

That is why what looks today like competition over model pricing may eventually turn out to have been the opening phase of a much larger restructuring.

The Agentic Economy

We often imagine technological revolutions through the products that first reveal them. The early web looked like websites. The smartphone revolution looked like phones. Cloud computing looked like rented servers. Artificial intelligence currently looks like chatbots. But mature technologies rarely remain confined to the interface through which we first encounter them. The deeper transformation occurs when the technology becomes infrastructure and begins reorganizing the systems around it.

Cheap intelligence could do exactly that. First it enters software, then it performs work, then it coordinates with other software, then it receives authority, then it controls budgets, then it negotiates, and then it transacts. At that point, calling it a chatbot misses almost everything important about it.

We have created software that can increasingly represent intention. And an economy in which software can represent intention will not necessarily preserve the institutions created for a world in which only humans could search, compare, negotiate and decide.

Some platforms will adapt. Some may become stronger. Some may retreat into infrastructure. Entirely new intermediaries will appear. Many predictions will undoubtedly be wrong. But the direction is becoming easier to see.

The most important consequence of cheaper artificial intelligence may not be that everyone gets a more powerful assistant. It may be that intelligence itself stops being the scarce layer. When that happens, value moves upward and outward toward agency, trust, identity, distribution, payments, proprietary context and control of the physical world.

The question is therefore no longer simply who will build the smartest AI.

It is what happens when intelligence becomes cheap, and who will control what that intelligence is allowed to do.

Karim Spring