I started thinking about the economics of artificial intelligence on the flight home from DEF CON. I had spent the week talking with security researchers, developers, technology companies, and vendors, nearly all of whom were discussing AI in one form or another. Some were using it to improve existing products. Others were building entirely new businesses around it. Everyone seemed to be trying to determine where the value would eventually settle.
The conversations left me thinking about a basic contradiction. AI companies want us to use their models as much as possible. They want AI to become something we depend on throughout the day, recommend to others, and eventually incorporate into nearly everything we do.
At the same time, every interaction consumes computing resources. The more successfully these companies encourage us to use their products, the more expensive those products become to operate.
That creates a strange economic problem. The companies building AI models are spending extraordinary amounts of money to make intelligence less expensive and more widely available. As that intelligence becomes cheaper, people find more ways to consume it. They also use it to create software, images, websites, research, and other products that were previously valuable because relatively few people could produce them.
AI is simultaneously creating a valuable new resource and undermining the scarcity that gives many existing products and skills their value.
This article combines publicly available research with my own observations and opinions. AI assisted with portions of the research and editing, but factual claims were reviewed against the cited sources. The conclusions and predictions are my own.
The Cost of Creating a Model
Developing an AI model requires far more than purchasing a collection of GPUs and feeding them data.
The companies creating the largest models employ highly compensated researchers, engineers, security professionals, and infrastructure specialists. They build or acquire access to enormous computing environments. They collect, purchase, license, prepare, and filter immense amounts of data. They must then train the model, evaluate it, secure it, improve its behavior, and build the infrastructure and safeguards necessary to make it available to millions of people.
A failed experiment can consume substantial resources without ever producing a commercially useful model.
At the frontier, these investments can reach hundreds of millions of dollars. The largest developers continue to use more compute and spend more money as they attempt to push beyond what existing models can do. Epoch AI estimates that the cost of training frontier language models has increased rapidly since 2020, even as the underlying methods and hardware have become more efficient. Epoch AI trends
That initially sounds contradictory, but it reflects two different trends.
It is becoming cheaper and easier to reproduce a previously achieved level of capability. The methods are better understood, more people have the necessary expertise, hardware continues to improve, and developers can build upon years of published research and practical experience. A level of performance that once required one of the largest laboratories in the world may eventually become available through a smaller model running on comparatively modest infrastructure.
At the same time, organizations attempting to create the best model in the world continue to expand the scale of their experiments. Efficiency improvements allow them to do more, but they often reinvest those savings into larger training runs, additional data, and new forms of reasoning.
Recreating yesterday’s frontier is becoming cheaper. Moving the frontier forward may still become more expensive.
Once a model has been trained, it can be reused repeatedly at a much lower incremental cost. A developer can purchase access to millions of tokens for what appears, relative to the original investment, to be a remarkably small amount of money.
In that respect, an AI model is like a factory. Designing and building the factory requires a substantial initial investment. Once it is operational, it can produce an enormous number of products without being rebuilt each time. Every unit still consumes energy, materials, equipment capacity, and labor, but at a fraction of the cost required to create the factory itself.
Training builds the factory. Inference operates it.
The model is not consumed when it generates a response, but each request uses very real resources: GPUs, electricity, cooling, networking, data centers, software, security, and the people required to operate the service. Tokens are simply the units through which much of that activity is measured and billed.
The cost of training determines whether a model can be created. The cost of inference helps determine whether operating it can become a sustainable business.
The Growing Value of Data
Compute is only part of the investment. Models also require data, and the economics of that data are changing.
Early developers of large models had access to an enormous amount of information available on the public internet. Some of it was licensed or freely offered. Some was collected under legal theories that remain contested. In my opinion, some AI pioneers were willing to take liberties with data that larger or more risk-conscious organizations might not have accepted.
Regardless of where those legal arguments ultimately settle, much of the readily accessible public corpus has already been collected repeatedly for AI development. The remaining challenge is not necessarily a shortage of raw information. It is a shortage of new, distinctive, high-quality data that developers can legally and reliably use.
People and organizations have also become much more aware of how their data might be used. Publishers are negotiating licensing agreements. Content owners are restricting automated collection. Governments are considering new rules. Companies that possess valuable proprietary information increasingly recognize that it may be one of their most important assets.
This may create an advantage for organizations that naturally produce or control large collections of legally usable data through research, licensed content, commercial relationships, or their ordinary operations.
The important distinction is not simply who possesses the most data. It is who has the right to use valuable data for a particular purpose.
Future improvements may depend more heavily on licensed material, specialized datasets, private organizational knowledge, expert human feedback, real-world interaction, and synthetically generated examples. Each comes with its own costs and limitations.
Compute may become more efficient and technical expertise more widely available while distinctive, legally usable, high-quality data becomes more valuable.
Open Weights and Proprietary Models
After investing in a model, its developer must decide how to distribute and monetize it.
Some companies keep their models proprietary and sell access through APIs, subscriptions, enterprise agreements, or cloud platforms. Others release model weights that can be downloaded, modified, and operated by someone else.
These approaches are often described as closed and open, although “open weight” is usually more precise than “open source.” A company may publish the trained weights without disclosing its complete training data, development process, or source code. Its license may also impose restrictions on how the model can be used.
The economics of a proprietary model are relatively straightforward. The developer retains control and charges customers or partners for access. A model may appear on several competing cloud platforms, but that does not mean it is open weight. It usually means the developer has established commercial distribution arrangements that allow customers to use the model through platforms they already trust.
The economics of open weights are less obvious. Why would a company spend millions of dollars creating a model and then allow other organizations to use it without paying for every token?
The answer depends on where the company expects to capture value.
An open-weight model can build recognition, attract researchers, encourage integrations, establish a technical standard, and create an ecosystem around the developer. The company may sell its own hosted version, support, consulting, specialized models, or future proprietary services. It may also benefit from improvements contributed by outside researchers.
For companies with other major businesses, the benefit can be even more indirect. Widely available AI can increase demand for cloud infrastructure, chips, advertising, consumer products, or other services. An organization does not necessarily need to earn money each time its model is used if the model strengthens another part of its business.
Open weights can also place strategic pressure on competitors. If a freely available model provides adequate performance, it becomes harder for another company to charge a premium for similar capabilities.
Releasing a model is therefore not necessarily an act of charity. It may simply move the profitable portion of the market somewhere else.
The Hidden Bet Behind an AI Subscription
The economics become even more interesting when AI is sold as a flat monthly subscription.
Traditional subscription software often has relatively low marginal costs for each additional interaction. Generative AI has a more direct and potentially significant computing cost attached to every request, even if caching, model routing, and infrastructure optimization reduce that cost.
A person who submits a few short prompts each week costs much less to serve than someone who continuously analyzes documents, generates software, performs research, creates images, and operates agents.
A flat subscription is therefore a prediction about the combined behavior of an entire population of users.
Some people will pay for the service and use it only occasionally. Others will consume as much capacity as the provider allows. The company is betting that lighter users will subsidize heavier users and that average consumption will remain below the point at which the subscription becomes unprofitable.
In that respect, AI subscriptions resemble gym memberships and unlimited data plans. Their economics depend partly on many customers consuming considerably less than the maximum amount available to them.
Providers can protect themselves with usage limits, slower processing, less expensive models, restricted access to premium features, and higher-priced service tiers. They can route simpler requests to smaller models, cache repeated information, or charge separately for high-performance processing and agent capabilities.
But every restriction introduces another tension. AI companies need people to use their products. Customers will not renew services they rarely find useful. They will not recommend a model that is constantly unavailable, slow, or constrained. Businesses will not redesign important processes around capacity they cannot reliably access.
The provider must encourage consumption without allowing that consumption to overwhelm the economics of the service.
Cheaper Tokens Can Still Produce Larger Bills
AI inference is becoming considerably more efficient.
Stanford’s 2025 AI Index found that the cost of querying a model performing at approximately the level of GPT-3.5 declined more than 280-fold between November 2022 and October 2024. Hardware costs and energy efficiency also continued to improve. Stanford AI Index
That does not mean every new frontier model will be cheaper than the one before it. The newest and most capable systems frequently command higher prices than older or smaller alternatives. They may perform more reasoning, accept more context, call more tools, and attempt much more complicated work.
The cost of obtaining a particular level of intelligence is declining. Yesterday’s frontier capability eventually becomes available through smaller, faster, and less expensive models. At the same time, the frontier continues moving outward, and customers ask the newest models to perform work they would never have attempted with the previous generation.
Someone who once used AI to rewrite an email may eventually use it to analyze thousands of pages of documents. A developer who once requested a code sample may ask an agent to build, test, troubleshoot, and document an application. A researcher may run several models simultaneously and compare their results.
The cost of each unit of intelligence may decline while the total amount of intelligence consumed increases much faster.
This resembles the Jevons paradox, the economic observation that greater efficiency can increase total resource consumption. When something becomes cheaper and more useful, people discover enough new uses for it that aggregate consumption grows rather than declines.
AI appears particularly susceptible to this effect because the potential demand for inexpensive intelligence may be nearly unlimited.
Agents Change the Unit of Consumption
Agents magnify this issue.
In a simple chat interaction, a person submits a prompt and the model produces a response. An agent may interpret the request, develop a plan, search several systems, call tools, evaluate the results, revise its approach, correct an error, and continue until it completes the task.
The person sees one request. The provider sees a chain of model interactions.
In describing its multi-agent research system, Anthropic reported that agents used approximately four times as many tokens as ordinary chat interactions, while its multi-agent system used approximately fifteen times as many. The ratios will vary by system and task, but the direction is clear. Delegating work can consume substantially more resources than asking a question.
That does not make agents uneconomical. An agent that consumes several dollars of computing resources but saves an employee several hours may produce an excellent return.
The relevant question is not simply how many tokens the agent consumes. It is whether the completed work is worth more than the combined cost of the model, infrastructure, development, supervision, security, and occasional failure.
This suggests that the future will not be built around using the most capable model for every task. Organizations will route work among different models, reserving expensive frontier systems for problems that require them and assigning routine work to smaller, less expensive alternatives.
The smartest model is not automatically the most economical model.
When Intelligence Stops Being Scarce
The companies developing AI are trying to recover enormous investments by selling access to increasingly capable models. Yet those same models allow their customers to avoid spending money in other parts of the economy.
A small-business owner can use an AI service to create a logo instead of hiring a graphic designer. They can build their own website instead of paying a web developer. A marketing professional can generate an image rather than licensing one from a stock photography service. An employee can vibe code a small application to solve an internal problem instead of recommending that the company purchase another enterprise software product.
The list goes on. AI can help people write reports, create videos, translate documents, develop training, analyze contracts, produce presentations, compose music, and perform other tasks that once required specialized software or professional assistance.
This does not mean that AI produces work equal to that of an experienced professional. A generated logo is not the same as a carefully developed brand identity, just as a vibe-coded application is not automatically equivalent to secure and supported enterprise software. Quality, reliability, maintenance, legal protection, and professional judgment still have value.
But the AI-generated alternative does not have to be equally good. It only has to be good enough for the customer’s immediate need.
AI also challenges the way professional work is priced. Consider a software company that traditionally estimates a project according to the number of developer hours required to complete it. If AI allows the company to create equally capable, high-quality software in one-quarter of the time, the development cost may fall dramatically. The value of the finished software, however, does not necessarily fall with it.
This creates a difficult adjustment for companies accustomed to equating time with value. The developer may feel uncomfortable charging the same amount for something that took far less time to produce. The customer may also expect a much lower price once they understand how quickly it was created. Yet neither perspective fully accounts for the value of the outcome, the expertise required to direct and validate the AI, or the years of experience that made rapid development possible.
A useful application is not worth less simply because an experienced developer found a faster way to build it. AI may therefore push software developers, consultants, designers, and other professionals away from hourly billing and toward pricing based on outcomes, intellectual property, risk, and value delivered.
That is where the economic disruption begins.
AI does not need to eliminate graphic designers, web developers, photographers, or software companies to affect their markets. It only needs to reduce the number of occasions when someone decides to pay for their products or services, while changing how the remaining work is produced and priced.
A single AI subscription may replace several smaller purchases. The value that once flowed to a collection of software vendors, creative professionals, consultants, and specialized services can instead move toward the companies providing models, computing infrastructure, and AI applications.
This is more than productivity improvement. It is a reallocation of economic value.
What Happens to the Zero-Day Market?
The conversations at DEF CON also made me think about vulnerability discovery.
Unknown vulnerabilities have value because they are scarce. A previously undisclosed vulnerability can provide access to a device or system whose owner does not yet know how to prevent it. Governments and law enforcement organizations may use vulnerabilities in support of investigations or intelligence operations. Security vendors and researchers use them to improve defenses. Criminal organizations and hostile governments use them to gain unauthorized access.
This has created legitimate vulnerability-research and bug-bounty programs alongside a much less transparent market for buying and selling exploits.
AI could disrupt both by reducing the cost of discovery.
As models become better at reading code, reasoning about software behavior, automating reverse engineering, and testing potential weaknesses, more vulnerabilities may be discovered by more people. Defensive teams can use the same capabilities to find and repair flaws before software is released. Multiple researchers may independently discover the same weakness, reducing its exclusivity.
The result may not be that the zero-day well completely dries up. Software will continue to contain vulnerabilities, and increasingly complex systems may introduce new opportunities for exploitation. AI will also help attackers identify and operationalize weaknesses more quickly.
What may change is the scarcity, exclusivity, and useful lifespan that give an individual vulnerability its value. A flaw that might once have remained undiscovered for years could be found independently by several AI-enabled researchers. Once it is reported and patched, its economic value falls dramatically.
AI could therefore increase the total number of vulnerabilities discovered while reducing the value of possessing any one of them.
The same technology might place pressure on bug-bounty programs. If AI dramatically increases the number of researchers capable of finding relatively simple vulnerabilities, programs may receive more duplicate or low-value submissions. Rewards may become concentrated around the smaller number of flaws that still require exceptional skill, access, or creativity to discover.
The well may not be drying up, but more people may be drawing from it with increasingly powerful pumps.
Who Ultimately Captures the Value?
The economics of AI are difficult to predict because cost, consumption, capability, and competition are all changing at the same time.
It is becoming cheaper to reproduce an established level of AI capability, even as the largest developers spend more to advance the frontier. Hardware and inference become more efficient, but people consume more intelligence as its cost falls. Agents require more model interactions, but they can complete work that previously required hours or days of human labor.
Open-weight models place additional pressure on the market by giving organizations access to powerful capabilities without requiring payment to the original developer for every use. Proprietary model companies respond by offering greater capability, easier access, stronger enterprise protections, and complete applications rather than model access alone.
Meanwhile, the intelligence produced by these systems reduces scarcity throughout the rest of the economy. It makes software easier to create, specialized knowledge easier to access, creative products easier to generate, and discoveries easier to reproduce.
That leaves an unresolved question: who ultimately captures the economic value created by AI?
It could be the companies training the most capable models. It could be the cloud and chip providers operating the infrastructure. It could be the application developers who turn models into useful products. It could be the businesses that use AI to reduce costs and increase production. It could ultimately be consumers who receive increasingly capable services at steadily declining prices.
Most likely, the answer will continue shifting among all of them.
On the way home from DEF CON, I kept returning to the same contradiction. The AI industry is investing extraordinary amounts of money to make intelligence abundant. Its success depends on convincing us to consume more of that intelligence every day. Yet the more abundant intelligence becomes, the more it threatens the scarcity on which many existing business models depend.
That may be the strangest part of the economics of artificial intelligence: the industry is racing to create something enormously valuable by making it progressively less rare.