Blog · 51 min read
AI economics: who keeps the profit and who’s left with the bill
By 2030, neural networks will be grinding through 120 quadrillion tokens a month. That’s 24 times more than now. And the ones paying for it won’t be the labs that build the models. It’ll be the businesses that use them. In other words, you.
I recently finished building an AI agent that takes part of the daily routine off the shoulders of a small company’s owner. As an economist and AI architect, I’ve dug through a huge pile of research, articles, reviews, and interviews on how AI is changing the economy, both inside a single business and globally. Today almost everything written about AI is either a showcase of what it can do or a conversation about risks. I suggest we look at the money behind AI.
Goldman Sachs forecast
By 2030: 120 quadrillion tokens a month
The labs won’t pay for it. Businesses will.
Story 1The “bubble”
Look at what’s happening right now.
Some say this is the greatest technology of the century. Models get smarter every quarter, neural networks write half the code, and trillions are pouring into data centers and chips.
Others answer: come on, it’s a bubble. Companies are burning hundreds of billions, almost nobody is making a profit, and the letters AI get slapped on everything, the way the word “blockchain” was a couple of years ago. My sense is that, at this stage, both sides have a point.
We were told: “The internet will change everything: retail, media, communications, advertising.” And? Yes, it did. Today a significant share of the world economy exists only because of the internet: e-commerce, digital advertising, cloud services, streaming, fintech, marketplaces, and social platforms. Without the global network, these markets simply wouldn’t have appeared, or they’d have stayed many times smaller.
Most of the power, though, is concentrated in a relatively small group of companies. Amazon sets the standards for e-commerce and cloud computing (AWS). Google (Alphabet) controls search and a huge share of digital advertising, plus YouTube. Meta holds the biggest social networks and ad inventory. Add Apple (its device ecosystem and the App Store), Microsoft (the Azure cloud, enterprise software, and AI), and the Chinese giants Alibaba, Tencent, and ByteDance (Douyin/TikTok), which dominate their home market and are pushing hard beyond it. Netflix and other streaming services reshaped how we consume media, and companies like Shopify and Stripe became the infrastructure for millions of small businesses.
These are the players who decide how people find information, shop, communicate, and pay. And by no means all of them built the internet’s infrastructure in the 1990s — many came later and took most of the value that had been created.
A technological revolution and a financial bubble aren’t opposites. If anything, they look like Siamese twins.
How a bubble is born
A revolution creates ideal conditions for a bubble
Huge uncertainty
nobody knows the market size or the winners
A beautiful story about the future
Fear of missing out
FOMO
Capital arrives
The direction is right. The timing and the winners’ names are not.
Why? Because a truly revolutionary technology creates ideal conditions for a bubble. It opens up a gigantic space of possibilities where nobody yet understands the real size of the future market. Nobody knows who the future winners are, which business models will survive and which won’t, or how fast ordinary people’s habits will change. But it’s already obvious to absolutely everyone that something very big is happening.
And the market starts frantically filling that huge uncertainty with its fantasies and expectations. A beautiful story about the future appears. Next comes plain fear of missing out, the famous FOMO effect. And then the capital arrives.
So the market is often right about the direction history is moving in, and badly wrong about something else. In reality, technology only expands what’s possible, while the ability to hold on to profit and survive depends on fundamentals alone. Why a real customer needs your product, and whether your business model has any margin at all.
Remember when the internet was just starting to enter our lives. We were told: the internet will change everything, retail, media, advertising. And?
YES! It did. Half the world economy today lives only thanks to it. But in the spring of 2000 the dot-com bubble burst, the stock index crashed by almost 80 percent, and thousands of “promising” companies evaporated along with their investors’ money. The ones left standing were those who ruthlessly carved up the main markets among themselves.
It was like that with railroads. It was like that with aviation. There’s a famous joke by one investor: if capitalists at the start of the last century had wanted to make money on aviation, they should have shot down the Wright brothers on their first flight. The airplane turned out to be a magnificent technology. Only the enormous value it created didn’t go to the airline owners. It went to passengers, hotels, tourism, world trade. To everyone except those who put money into the planes themselves.
Why is that? Because a truly revolutionary technology creates ideal conditions for a bubble. It opens a gigantic field of possibilities where nobody yet understands the size of the future market, or who the future winners are, or which business models will survive. But it’s already obvious to everyone that something very big is happening. And the market starts frantically filling that huge uncertainty with fantasies. A beautiful story about the future appears. Next, the fear of missing out kicks in. Then the capital arrives. The market is almost always right about the direction. And almost always wrong about the timing and the names of the winners.
Hence the main takeaway for your business. Technology only expands what’s possible. The ability to hold on to profit and survive depends on boring things: 1) “why a real customer needs your product” and 2) “whether your model has any margin at all.” We’re about to go through everything else.
Story 2There’s no free scaling here
First, let’s work out what an AI model actually is.
How a model works
The model doesn’t read text. It guesses the next token
Cuts text into tokens
a whole word, a piece of a word, or a punctuation mark
Calculates the probability of every next token
millions of probabilities per answer
Picks the most likely one
say, with 84% probability
Not a truth detector but a statistical autocomplete
When you open, say, ChatGPT and type in your question, your request goes to what’s called an LLM, a Large Language Model. And this is where it gets interesting. The model doesn’t read your text the way you and I do. It cuts any text into small fragments called tokens. A token can be a whole word, a piece of a word, or just a punctuation mark.
In essence, a super-smart neural network works like a vast, incredibly advanced statistical autocomplete system. It calculates the probability of every possible next token and picks the most suitable one. Once it gets your request, it models millions of probabilities and simply calculates: with a probability of, say, 84%, the next token should be this word.
That’s exactly why models make downright childish mistakes. Because they work with tokens rather than mathematical meaning, AI can get completely stuck on basic arithmetic with large numbers, or still miscount the letter “r” in the word “transformer” on the third try.
So how did this model actually get so smart? It was fed astronomical amounts of data: petabytes of books, articles, Wikipedia, and basically the whole internet. From that, it learned to pick up highly complex patterns, to catch grammar and the logic of how human thought is put together.
Two consequences of training
What many enthusiasts would rather not think about
First: the knowledge cutoff
- the model knows the world only up to a set moment
- about yesterday, it either honestly doesn’t know or makes things up
- unless it has access to search
it doesn’t know recent events
Second: the answer mirrors the question
- vague prompt: fluff and platitudes
- clear context, role, and instructions: a working tool
quality starts with the question
But this training has two critically important consequences that many enthusiasts would rather not think about.
First: the model knows about our world only what made it into its training data before a strictly defined moment, the so-called knowledge cutoff date. If something happened yesterday or this morning, the model will either honestly admit it doesn’t know, or, with a perfectly straight face, start filling in the gaps and making things up. Unless, of course, it has direct access to internet search.
Second: the quality of the answer mirrors your question. Ask a vague, fuzzy, generic prompt, and you get fluff and platitudes. Give it clear context, spell out a role and strict instructions, and you get a precise working tool.
One last thing. The model has no mind, no opinions of its own, no emotions. All its empathy, politeness, or “personality” is simply a good imitation of the communication style in the texts it was trained on.
Two processes
Training happens once. Inference, every time
Training
- OpenAI, Anthropic, Google
- billions of dollars
- the output: frontier models
huge investment, once
Inference
- running a model that’s already trained
- every request means computing
- every little request and every line
you pay again and again
When you work with neural networks, it’s important to tell two processes apart.
The first is the training itself. That’s when giants like OpenAI, Anthropic, or Google spend billions of dollars to build the most advanced foundation models, the so-called frontier models. Frontier models aren’t built on some special physical principle. The term simply marks who is at the edge of intellectual capability right now.
The second process is inference, that is, actually running a model that’s already trained. You send a request, and the model computes on the fly and gives an answer. And it’s this inference, every little request and every generated line, that you or your startup will have to pay for again and again.
That’s why a huge ecosystem of applications is springing up around frontier models today. What do these applications do? They connect AI capabilities with data, tools, and workflows. For example, they help write a program, go through documents, handle a customer request, and so on. These AI applications are seeing fast growth in users, requests, and integrations.
And this is where it gets really interesting, because as usage grows, the cost of inference starts growing just as fast. The main economic feature of inference is that it costs money for every single request.
Unlike ordinary software, where adding new users or requests costs almost nothing, every answer from a neural network is a physical process. In data centers, thousands of GPUs run billions of math operations. Every second the model runs costs real money: it eats huge amounts of expensive electricity, and the servers and chips wear out over time and have to be depreciated.
Story 2
There’s no free scaling
Classic software: the next customer costs pennies
AI product: every answer is real computing
So the difference between the economics of traditional IT products and AI products lies in the cost of serving users. In classic software, most of the team’s budget goes into the initial development. Say you build a messenger, automate a system, or make a mobile banking app. Once the program is ready, serving each additional customer costs the company literally pennies. As a result, a business like that has extremely high margins. The more users, the higher the company’s net profit, because infrastructure costs barely grow. That, by the way, is how digital monopolies are born.
In AI applications, this familiar model falls apart completely. Every time a model answers a user, writes code, recognizes speech, or performs an agent step, expensive computing power is working somewhere. As a result, an AI business doesn’t get free scaling. As your audience grows to a million people, compute costs grow almost proportionally, eating a significant part of the margin.
Then agents come onto the stage. An agent is no longer just a chatbot. An agent is a system that uses the model as a brain to carry out a complex assignment. It plans the next step itself, calls the tools it needs, looks at the result, fixes mistakes, and moves on. The model only generates answers, while the agent system organizes the entire process around it.
Agents
One command, dozens of model calls
“Analyze the market and prepare a presentation”
With AI agents, the cost of inference becomes even more noticeable. For example, a user might give a single command: “Analyze the market and prepare a presentation.” But inside, the system may make dozens of inference requests: separately build a plan, find data, read it, compare, double-check, write the text, edit it. So the user took one action but paid for dozens of the model’s compute operations.
The cost formula
Token price is not the cost of the task
This leads to a very interesting conclusion about the economics of AI. Comparing only the price per million tokens isn’t enough. Costs are the product of the number of tasks, the average resource consumption, and the price of those resources.
Let’s walk through a hypothetical example. The price per token fell by 80%, and consumption per task grew 10 times. So the cost of the task actually doubled. But the task may have become much more useful to the user along the way.
Why does this matter? Because neither a falling token price nor a growing total bill for those tokens says anything, on its own, about how effective the adoption is.
Designing the whole system means balancing three things: the token price has to leave room for a margin; the number of requests has to be reasonable, so the agent isn’t generating 10,000 requests for nothing; and all of it has to produce a result that’s valuable to you. It’s a complicated formula. And it’s what sets AI business models apart from ordinary traditional software.
This has an interesting effect on how apps make money. If your costs depend directly on how much computing work the neural network does, then selling customers unlimited access for a flat subscription fee looks very risky.
A $20 subscription
Same subscription, hundreds of times the cost
Quiet developer
- a couple of times a week
- asks to explain a bug in his code
costs a few cents
Automation fan
- agent runs for hours
- hundreds of tests, thousands of lines
- works while he sleeps
hundreds of times more
Picture two customers with exactly the same subscription, say $20 a month. The first is a quiet, modest developer. A couple of times a week he logs into the service and asks it to explain where he messed up in his code. Serving him costs the company a few cents.
The second is a bold automation fan. Every day he runs the agent at full throttle, and it spends hours plowing through the project’s architecture, running hundreds of tests, and rewriting thousands of lines.
Each of them brings you exactly $20, but the cost of serving them doesn’t differ by a few percent — it differs by a factor of hundreds. Especially if the agent starts working autonomously. While the person is drinking coffee or even sleeping, the neural network burns through expensive hours, and you’re the one paying for them.
That’s where the classic subscription model breaks down. Selling unlimited AI access for a flat, say, $20 a month is financial suicide for the business model. So services are urgently changing how they charge. Cursor, for example, sells a convenient development environment but builds in limits and extra charges for heavy model use. Or take a service like ElevenLabs, which works with speech and digital avatars: it sells credit packs. You don’t buy unlimited use. You buy a specific number of minutes or characters of generated audio.
So the good old subscription isn’t disappearing entirely. But now it has to be designed far more carefully, with a constant eye on the real cost.
What exactly is included in the plan, where the hard limit is, how much overage costs, and what real benefit the customer gets.
What revenue growth means
Where every new sales dollar goes
What’s more, the meaning of revenue growth changes completely. In ordinary software, of every dollar received, about 70 cents stay with the company for marketing, salaries, and growth. In AI applications it often works differently: of every new dollar in sales, 80–90 cents go to the model provider, for calls and compute. A company can grow its revenue quickly and still earn almost nothing.
With agents, the picture gets even more complicated. An agent doesn’t need you sitting in front of the screen. You give it one task, close your laptop in the evening, and go to bed, and it works on its own all night: digging through databases, researching the market, writing code, running tests, fixing bugs, sorting email, and getting documents signed off. In the morning, you get a finished result.
For those who rent out computing power, this is convenient: their servers are loaded around the clock. For your business model, it’s a risk — the cost of all those overnight calls affects what you earn from each customer and can eat your profit fast.
Story 3Who’s who in the new AI economy: chips, infrastructure, models, and applications
You know what changes everything in this story? The customer couldn’t care less about tokens. The customer pays for a solved problem. Tokens are like a fuel-consumption meter. It shows how much you use. It doesn’t show whether you got there. So in the economics of AI, you need to tell apart four different roles. At each level, money and risk are distributed differently.
Story 3
Four levels, four different business models
Applications
Cursor, ElevenLabs, Harvey
must keep customers every month
Models
OpenAI, Anthropic
sell access through an API
Infrastructure
Azure, AWS, Google Cloud, Oracle, Together AI
earn back the investment over years of rentals
Chips
Nvidia, Broadcom, memory
paid at the moment of sale
On one side are the makers of the hardware, the chips. That’s Nvidia, Broadcom, and the makers of specialized memory. They sell the “shovels” in the gold rush: the very accelerators everything is trained and run on.
On the other side is infrastructure. These are the cloud providers: Microsoft Azure, Amazon Web Services, Google Cloud, Oracle, and others. There are also niche services for running models, like Together AI. They buy the hardware, build giant data centers, and rent out the capacity.
The third player is the developers of the most advanced models: OpenAI, Anthropic, and others.
The fourth level is the app makers. For example, Cursor for programming, ElevenLabs for voice work, apps for searching corporate knowledge, Harvey for lawyers, and so on.
The key point is that money and risk work differently at each level. Nvidia gets its billions right away, the moment it sells a chip. The cloud provider that bought that chip expects to earn back the investment only after several years of rentals. And the app maker has to keep its customers every single month so they go on paying for the subscription.
So Nvidia’s record profits don’t at all mean that the buyers of its chips will ever recoup their investment. They tell you only one thing: hardware is in demand right now.
Nvidia is often accused of inflating the bubble. Demand for its chips exists only as long as the whole AI story keeps developing. So it finances its partners and gives them loans, so they can use that money to buy the very same Nvidia chips. In economics this is called circular financing, and many economists see it as a warning sign. But for now, nobody can do anything about it.
Like the railroads
Tracks first, trains later
Huge capital
Tracks, stations, bridges, tunnels
Trains with passengers who bought a ticket
Will paying demand catch up with the investment?
The main economic question is this: will demand from paying customers catch up with the enormous scale of investment now pouring into infrastructure?
In this sense, today’s AI economy looks a lot like building railroads. First you have to lay the tracks, invest huge sums, and build stations, bridges, tunnels, and trains. Only then does enough economic activity start moving across the network — those very trains carrying passengers who bought a ticket.
There’s a famous joke by Warren Buffett: if capitalists at the start of the 20th century had really wanted to make money on aviation, they should have shot down the Wright brothers on their first flight. Of course, the point isn’t that the airplane turned out to be a bad technology. Quite the opposite: it turned out to be a magnificent one. But the enormous value aviation created went not so much to airline owners as to passengers, aircraft makers, hotels, tourism, and global trade. Put simply, to everyone who gained the ability to travel around the world.
The same can be said about AI. Even if AI becomes infrastructure as important as electricity, railroads, or airplanes, it doesn’t at all follow that AI companies will automatically become the most profitable in history.
This raises another important question: what if the models themselves gradually become an ordinary commodity? That doesn’t mean all models are the same or that they’ll stop being needed. It means only one thing: for most tasks, a buyer will be able to choose among several good-enough options and will start comparing them more carefully on price, speed, and ease of access.
That creates an uncomfortable prospect for model developers. The technology gets more useful, people use it more often, but charging a high price becomes very hard. It’s like the situation with airlines: huge investments, complex production, and a constant fight to keep capacity full. This is a possible path, not proof that every AI lab is bound to become low-margin.
The main sign that models are turning into a commodity isn’t cheapness in itself. It’s the moment a customer says: “For my task there’s almost no difference. I’ll take the cheaper one.”
That’s exactly where the economics change. Technological capability keeps growing, but the ability to make money on uniqueness shrinks. It’s the “good enough already” principle. If you feel the current model already handles the job, you have no incentive to buy a new one and overpay.
Three levers
All of AI economics comes down to three speeds
Price ↓
how fast the price falls
Cost ↓
how fast the cost falls
Usage ↑
how fast paid usage grows
Price −80% → sell 5 times more to keep the same revenue
Can usage growth make up for the falling price? If the price drops by, say, 80%, the company needs to sell five times more to get the same revenue. But every extra request also has to be served. So a fivefold rise in volume doesn’t yet mean keeping your profit.
To simplify, the economics of AI come down to three things: how fast the price falls, how fast the cost falls, and how fast paid usage grows. These are the three main levers.
If cost falls faster than price, a company can earn more even at lower rates. If price falls faster than cost, growing demand has to make up for the shrinking earnings on each operation.
So an impressive rise in the number of processed requests doesn’t, on its own, answer the question of how much money the model provider has left once all the work is done. Follow that thought, and a conclusion emerges: competition heats up first around common use cases. To make up for falling prices, you have to compete for mainstream users and common business tasks, where what matters is combining good-enough quality with an acceptable cost of repeated use.
Story 4Is buying API access worth it, and where is real value created?
There are companies that develop the most advanced AI models: OpenAI, Anthropic, Google, and others. They invest huge sums in research and training, assemble strong teams, and secure access to vast computing resources. Then they give other companies access to their models’ capabilities through an API.
This is where the growth is: new startups are springing up like mushrooms after rain... roughly 90% of them are AI applications for all kinds of tasks. These apps send a request, the model on, say, OpenAI’s servers processes it and sends back the result. For a startup, this setup sharply lowers the barrier to entry. The frontier model itself becomes just a basic building block.
- You don’t build your own data center; you use the cloud.
- You don’t train a model from scratch; you pay for the API as you go.
Within a week or two, it becomes clear whether customers need the product. Yes, these days, instead of building some huge infrastructure for their idea or service, startups test the hypothesis on a real customer.
Cloud computing once let companies launch services without buying their own servers. Today, access to ready-made models lets you build products without training your own model. You’re essentially renting its capabilities: you send a request and pay for usage. Early on, this is often the rational choice. Instead of big investments in your own model, you pay as you go and test the hypothesis quickly. A startup’s main job is to test an idea as frugally as possible and in a short time.
Frontier models create something important: they lower the cost of testing and entering the market. They let thousands of companies experiment. This is one of the mechanisms through which a foundational technology builds a market around itself. But then an interesting effect appears. When you start growing and calculating the actual costs and returns, the question “which model is the smartest” stops being the main one. Another question, critical for a business, takes priority: how do you make the product’s economics work?
Gross margin
A great product, but a bad business
The customer pays $10
A middleman who buys intelligence wholesale and sells it retail
For a business, the metric that matters is gross margin: what’s left after the direct costs of delivering the service. If you sell a service to a customer for $10 and immediately hand $8 of that ten to OpenAI or Google for the API, you may have a great product but a bad business. You become a middleman who buys intelligence wholesale, wraps it in nice packaging, and sells it retail. The faster you grow, the more of other people’s money you have to burn to fund that growth.
This applies to every business model with low gross margin. Where variable costs are high (delivery, couriers, and so on), it’s very hard to build a sustainable model. Markets like that often end up with one or two big players, and the rest don’t survive.
How do you get around this obstacle? You have to make the customer a different promise. The user doesn’t care how many tokens they spent. They pay for a solved problem.
Yes, an autonomous agent can spend 10 times more tokens than a simple chatbot. But if, instead of a canned brush-off, it actually wrote a working app, found a critical bug in the code, or did market research, the perceived value of that result jumps.
The main metric is the cost of getting a successful result, and tokens are just a meter of compute spent. They show how much “gas” was burned, but they don’t tell you whether you reached your destination.
And look around you. Companies already using LLMs run into a problem early on: a cheap open model can turn out insanely expensive if a person then has to spend three hours rewriting the text. And an expensive model obviously pays off if it reliably gets the task done the first time.
Customers’ willingness to pay may grow, because AI is no longer just an advanced search bar or a picture generator. It’s starting to solve complex tasks. When ChatGPT was just a convenient way to get a summary, its economic value was limited. But once tools appear that replace five hours of a qualified analyst’s work, that’s a whole different kind of money and a different kind of value.
If a system spent $10 of compute on a task, is that expensive or cheap? If it wrote three banal paragraphs, then yes, expensive. But if it replaces the work of a financial analyst earning $3,000–4,000 a month, that’s a very good deal.
When the cost of getting a valuable result falls thanks to optimization, while quality and value grow, the economics of an AI product finally start to work. A business that only yesterday looked doubtful because of expensive compute starts showing good margins.
Story 5How tech bubbles are born: the gap between expectations and real economics
So where do financial bubbles come from? A bubble forms not only because investors overestimate how fast the prices of hardware or tokens will fall. It arises from a gigantic gap between the expected value and the real value of the result.
What does this trap of inflated expectations look like?
First, hype and fear of missing out create an illusion. The market values today’s companies as if an agent will become an autonomous employee tomorrow. Give it the task “increase sales by 20%,” and it’ll do everything perfectly, and the company will save millions on staff. In practice, for complex, high-stakes tasks in law, medicine, finance, or software development, the model still delivers the work of an intern with a vivid imagination.
Second, there’s a hidden cost: the cost of oversight. To turn a raw answer from a neural network into a result you’re not embarrassed to hand a client or put into operation, a business has to assign a qualified specialist who double-checks, polishes, and fixes the errors. If the model’s work cost $2, and then an expert spent 2 hours checking and correcting it, the real cost turned out to be huge, and the economics don’t work.
We run into the fundamental difference between “done fast” and “made what was needed.” A neural network can generate a 10-page report or 500 lines of code in 5 seconds. That looks impressive. But if nobody reads the report, and the code contains a hidden vulnerability that will bring the system down in a month, the economic value of that result is zero, or even negative.
Story 5
A bubble is the gap between expectation and result
The essence of the bubble is overvaluing finished, valuable results at the current stage. The market is assuming that AI already delivers the work of a strong specialist with 95% reliability. Reality says it delivers junior-level work with 70% reliability, and errors in the remaining 30% can cost a company its reputation or its profit.
Were we wrong that AI will someday learn to deliver flawless, valuable results? Most likely it will. But the market has already paid trillions of dollars for that future today, counting on the ideal employee showing up tomorrow morning. When it turns out that a reliable result still needs several more years of process tuning, testing, and losses, the market value of many overvalued products will inevitably go down.
The main question analysts are worried about right now: how quickly will this future arrive? The market isn’t valuing today’s technology; it’s valuing the expectation of what that technology will become tomorrow. If cost really falls by a factor of 10 in a couple of years and the result is consistently valuable, great. But what if companies’ market caps already assume prices falling by a factor of 100 or 1,000? Or if customers start paying generously only in 10–20 years, while the loans for data centers have to be serviced this month? That’s exactly how bubble dynamics are born.
The most interesting technology bubbles come from getting the timing wrong. The dot-com market wasn’t wrong that e-commerce would become a gigantic business. It was just wrong about how fast that would happen. Investors weren’t wrong that the internet would change all of global business. They were wrong to think that absolutely any company with “.com” in its name deserved a billion-dollar valuation right now.
The main rule of such revolutions: the winner is almost never the company that was the bright symbol of the first bubble.
All the real economic value, meanwhile, was taken by completely different players, who came later or adapted in time: Google, Amazon, Meta, Netflix, the cloud providers.
A financial bubble can completely misread the future and completely fail to guess who will end up owning it.
Story 6Compact models, specialization, and AI agents: how to cut costs and build a competitive advantage
Here are three key insights on how to survive in the AI economy without going broke.
First insight: open models often cut 70–90% of costs, but they need work.
Many analysts note that open models now trail the very top tier (ChatGPT or Claude, for example) by only a few months. And running them costs 70–90% less. The logic is simple: why pay for the elite “brain” of a flagship model if all you need is to sort support tickets?
You take an open model, do what’s called post-training on your real data, and turn someone else’s general-purpose intelligence into your own specialized one.
Don’t wait for “the next model.” It won’t solve your problem
Everyone talks about new models: “Just wait till GPT-6 comes out, and everything will work.” In reality, the giants have already spent billions to create a general-purpose brain. It knows language, code, and how the world generally works. That’s the foundation. It’s already there. But what makes you stronger than your competitors isn’t the foundation. It’s the ceiling. The difference is simple:
Don’t wait for the next model
The foundation exists. Your ceiling is post-training
Pre-training
- companies burn billions
- a general-purpose model
- “knows a little about everything”
the foundation
Post-training
- your tasks
- your risks
- your quality criteria
what makes you stronger
Pre-training is when companies burn billions and create a general-purpose model that “knows a little about everything.”
Post-training (fine-tuning) is when you take that ready model and tune it to your tasks, your risks, and your quality criteria.
Imagine you need a great taxi for your city. You can wait for someone to release “the ideal car of the future.” Or you can take a good base car and set it up: routes through your city, the way you talk to customers, the rules for handling complaints, the quirks of local roads and laws. The second option usually works out faster and cheaper. Same with models.
When fine-tuning makes sense, and when it doesn’t
The advice is simple: don’t sit around waiting for the next “breakthrough” model. Take a decent open base model and bring it to perfection on your own tasks. But there’s a catch. Fine-tuning and maintaining it cost money too. So:
At the start, it’s almost always cheaper to just buy access to a ready-made service.
Fine-tuning your own model only makes sense when the savings across millions of repetitive tasks outweigh the cost of development and maintenance.
If you handle 50 requests a day, a ready-made solution will most likely be enough. If you have 50,000 similar documents or requests a month, and a 2% error rate costs you real money, then it’s time to run the numbers.
The main question a team should know how to ask
Not “which model should we use,” but:
Where does extra reasoning quality actually change the result, and how much are we willing to pay for it? Here are three typical situations:
Extracting fields from a document (an invoice, a contract, an application). A fast, cheap model is often enough. A mistake here is rarely critical, and it can be double-checked.
Drafting a reply or a letter. Here style and precise wording matter more. You can use a better model.
Sorting out an ambiguous customer case (a complaint, a disputed situation, a complex request). Here a mistake is costly: reputation, money, time. This is exactly where it makes sense to pay for stronger reasoning or a fine-tuned model.
Choosing a model stops being a technical setting. It becomes a management and product decision:
The mindset “we always use the very best” doesn’t answer that question. It just burns the budget.
Second insight: the winner isn’t the smartest intelligence but the best “harness.”
Choosing a model stopped being a technical setting long ago; it’s a management decision. You have several important criteria: quality, price, and latency.
Second insight
The best harness wins
Product A
- model reinvents the order of analysis
- reads every document
- calculates the metrics itself
lots of paid model work
Product B
- sequence of steps is set
- ordinary code calculates the ratios
- model handles only ambiguous information
- checks catch the gaps
same analysis, less model work
Imagine an agent that evaluates companies. In one product, the model reinvents the order of analysis on every run, reads all the documents, and calculates the metrics itself. In another product, the sequence of steps is already set: ordinary code pulls the data and calculates the ratios, the model handles only the ambiguous information, and checks catch the gaps. The customer gets the analysis they need, but the amount of paid model work can differ substantially.
Third insight: the architecture of a modern AI product is an ensemble.
Expensive resources go only where they change the result
The most expensive mistake in AI systems isn’t the one where the model gets something wrong.
It’s when you’ve sent a simple task to an expensive model. Or a complex one to a cheap model. Picture contract processing.
Third insight
Each task gets its own model
Contract
A fast, cheap model can identify the document type.
Extracting standard terms: a smaller model, fine-tuned specifically on your templates.
Untangling an unusual contradiction or ambiguity: a stronger one.
And the decision on whether to accept a legal risk stays with a human.
That’s what a normal workflow looks like. Not “one smart model for everything,” but a precise allocation of tasks. Expensive resources kick in only where they critically affect the quality of the result. Everything else gets done faster and cheaper. If the system sends simple work to an expensive model, you’re simply overpaying.
If it sends complex work to a model that isn’t reliable enough, the savings dissolve into rework, fixes, and the consequences of the mistake. Economically, the best option is always the same one: the one that does the specific job with the reliability it needs at an acceptable total cost. Advantage comes not from the number of agents but from how precisely you divide the work among them.
But there’s a question more important than routing
Even a perfectly tuned workflow is useless if you don’t understand what “good” means for you.
Utility function
What does “good” mean for you?
The same doctor–patient conversation
Clinic
- chart time −70%
- no critical medical errors
its definition of “good”
Insurer
- extract diagnoses accurately
- procedures and reimbursement codes
its definition of “good”
Take transcripts of doctor–patient conversations. A standard speech-to-text feature optimizes one thing: the fewest word errors. For a medical product, that’s not enough. A typo in a preposition is trivial. But a mistake in a dosage, or “no allergy” written down instead of “has an allergy,” is a potential lawsuit. So the company starts doing more than collecting data: it builds a utility function. It answers a simple question: what exactly counts as a good result for us in this scenario? For one company, it might be: cut the time spent filling in charts by 70% while allowing no critical medical errors.
For another (an insurer, for example), it’s extracting diagnoses, procedures, and reimbursement codes as accurately as possible. The doctor’s original conversation is the same. The base model can be the same. The data is similar. But the idea of a “good result” is completely different.
Evals are now a strategic asset
For a utility function to work, it has to be turned into concrete checks. Not an abstract “do a good job,” but a set of clear criteria that tell correct model behavior from incorrect. That’s what evals are: an evaluation system.
Two companies in the same market can use the same base model and have similar data, yet define a good decision in completely different ways. One bank has one risk appetite and set of standards; another bank has different ones. Specific evals become the organization’s formalized knowledge of exactly how work should be done here.
This is where real competitive advantage takes shape. If your product spent several years collecting signals about which errors matter and which don’t, and what users consider a good result, and systematically turned that feedback into improvements, a gap opens up that’s hard to close.
Search had a similar story. In the beginning, almost anyone could build a search engine. But Google gained its advantage not from one secret line of code, but because using the system constantly generated signals: what people search for, what they click on, which result turns out to be relevant. Loops like that are very hard to reproduce. A competitor can’t just copy the product — it has to reproduce the product’s whole history of interaction with the market.
The same thing is happening now with AI products. The winner isn’t whoever has “the smartest model.” The winner is whoever understands more precisely what “good” means for their task, and knows how to measure and improve it systematically.
Story 7The biggest fear of AI startups: being absorbed by the platform
While you’re small, you enthusiastically build a nice service on top of someone else’s API. You spend investors’ money, look for an audience, feel your way to a winning use case, say, automated review of legal contracts. In other words, you prove to everyone that the market exists.
Then comes a classic risk called platform envelopment. First the platform (OpenAI, for example) helps you build a business, and then it discovers that what you’re doing on top of it is valuable enough to fold into its own product. The base model from OpenAI or Google gets a bit stronger, and right inside its own system the platform adds one small button: “Analyze contract for risks.” And that’s it. Your business, which was a nice wrapper around someone else’s API, vanishes into thin air.
It’s exactly the same story as when Apple or Microsoft killed dozens of third-party utilities with a built-in operating system feature. This risk is especially interesting because the foundational technology expands its own capabilities. Today frontier models can write text, tomorrow they’re already pretty good at analyzing documents, and after that they can use a browser, work with files, and carry out a multistep process on their own. Each new leap in capability potentially eats part of the application layer built on top of these models.
The boundaries of the AI market keep shifting. That shift is quickly squeezing out those who just put a nice wrapper on someone else’s model
If all your value came from taking a base model, adding a nice interface, and selling access, the space between the platform and the user keeps getting thinner. Platforms themselves are moving toward the user. What was a product yesterday becomes a feature today. To stay in the market, you need something a general-purpose platform will never have. Not “a slightly better prompt” and not “one more agent.” Something that can’t be copied quickly. What actually protects a business:
Story 7
The boundary shifts, squeezing out the wrappers
What actually protects a business
- proprietary data
- a process in the customer’s routine
- deep integration
- distribution
- experience from thousands of operations
- your own specialized model
Proprietary data that public networks can’t reach.
A specific process already built into the customer’s daily routine, one they’d find it awkward to live without.
Deep integration into corporate software, internal databases, and operational systems.
Distribution: direct channels where you already have a foothold and that are expensive for a competitor to enter.
Accumulated experience from hundreds of thousands of real operations and scenarios, already turned into an understanding of what works and what doesn’t.
Your own specialized model, built not “for everything” but for a narrow task with clear economics.
Everything else is temporary. Because of this, the competitive picture changes fundamentally. Let’s set aside the “whose model is smarter” or “whose benchmark is higher” arguments. That’s not where the game is anymore. Real advantage no longer sits at a single point. It’s spread across the whole system: from proprietary data and strict quality checks to how you route tasks between models, how you connect tools, and how you systematically turn feedback into improvement. The winner isn’t whoever has “the smartest model.” The winner is whoever has a system that understands the customer’s work better and does it more cheaply.
Story 8The cheaper intelligence gets, the more energy it devours
Up to now, we’ve looked at the economics of a single company. Now let’s add up thousands of these companies and see what happens in the physical world. First there was the chatbot: one question, one answer.
Story 8
The cheaper intelligence gets, the more of it we consume
Chatbot
one question, one answer
Agent
one assignment, dozens of requests
Multi-agent system
one click, dozens of parallel processes
Jevons paradox: ×24 by 2030
Then came agents: one assignment spawns dozens of internal requests, searches, and checks.
Now the era of multi-agent systems is arriving. One user click sets off a coordinated operation: one agent researches companies, a second checks facts, a third pulls the conclusions together. From the outside, it looks like a single action. Inside, dozens of parallel processes start up.
That’s why the idea that “inference is a temporary problem that will fix itself once models get cheaper” is most likely wrong. Demand for inference will grow at a huge pace precisely because applications are becoming more and more agentic, at the same time as the models themselves and their usage keep growing. Goldman Sachs forecasts consumption rising to 120 quadrillion tokens a month by 2030 — 24 times the current level.
This is where the classic Jevons paradox kicks in: “The cheaper and more accessible a resource becomes, the more of it people consume.” When car engines became more efficient, we didn’t burn less gasoline — we just started driving hundreds of times more. When a single inference call gets cheaper, companies don’t save. They start piling on more models, automated tests, routing between agents, and extra checks.
Each individual company still benefits from buying access through an API and lowering the cost of its task. But across the whole industry, these improvements widen the range of profitable uses so much that the total need for capacity only grows.
Hence a direct risk. If inference becomes unavailable, slow, or too expensive at peak load, the core function of a large number of products simply stops. Voice chat turns into a conversation full of pauses. An agent that was supposed to solve a task in seconds starts taking minutes. Users wait in a queue or have their requests turned down.
So the quality of an AI product no longer depends only on how smart the model is. What matters is whether it can do the needed work on time, under load, and at an acceptable price.
That’s why the debate is gradually shifting. There’s less and less talk about which model is “smarter.” More and more about whether there will be enough data centers, energy, and standardized infrastructure when thousands of apps all start spawning dozens of agents and round-the-clock inference at once.
Story 9Why AI giants suddenly started buying up power plants
Story 9
The firm’s boundary follows the bottleneck
Ford
a chain from raw materials to factories
Amazon
logistics with planes and trucks
Tesla
its own batteries
Netflix
its own films
Now AI giants are buying power plants
Throughout this whole AI story, we keep running into one of the oldest plots in economics. Companies start doing something themselves at exactly the moment external suppliers can no longer keep up. Henry Ford built his own chain from raw materials to factories. Amazon built logistics with planes and trucks. Tesla went into making batteries. Netflix, having started with other people’s films, began making its own. The reason is the same in every case. Buying from external suppliers works as long as they deliver the resource you need at the right quality, at a reasonable price, and without disruptions. As soon as it becomes the main bottleneck to growth, the incentive for vertical integration appears instantly.
When scarcity stops being a question of price
When demand is huge, there comes a moment when the computing capacity you need simply isn’t on the market. And then the problem stops being a question of “how much does it cost.” It becomes a question of access. If you have gigantic demand from customers but no capacity, you physically can’t turn that demand into revenue. Without the resources coming in, the business can’t deliver anything.
A company always integrates toward wherever its critical constraint is. When Amazon realized that logistics was hurting the customer experience, it started building logistics. When an AI product realizes that computing capacity has become the constraint, it starts thinking about its own capacity. The boundary of the firm keeps moving to follow the bottleneck.
The digital revolution has run into transformers and kilowatts
At the start of the boom, it seemed that the main advantage was software, algorithms, and smart request routing. But when capacity becomes severely scarce, owning physical resources in itself becomes the main protective barrier. And so, at the culmination of the digital revolution, it turns out that its main constraints are things as concrete as electricity, land, transformers, data centers, cooling, chips, and capital.
Look at what’s happening right now. Google is buying an infrastructure giant so it can build data centers together with power generation. Microsoft is signing a twenty-year energy contract and, to make it happen, bringing a shut-down nuclear reactor back to life. Amazon is investing in developers of small modular nuclear reactors. OpenAI and its partners are laying the groundwork for Stargate and looking for land, gigawatts, builders, and equipment.
For the last 20 years, the IT industry got us used to the “asset-light” model: an attractive business without heavy physical assets. Facebook didn’t own content, Uber didn’t own cars, Airbnb didn’t own hotels. Software scaled almost for free. And suddenly AI, the pinnacle of software development, brings us back into the harsh material world.
The most physically grounded business
Behind a one-second answer, a power plant
For a smart answer to appear on a screen for a second, somewhere out in reality there has to be a chip made at a $20 billion factory. That chip sits in a server, the server sits in a data center, and the data center draws the output of an entire power plant.
Digital intelligence has turned out to be the most physically grounded business we’ve seen in the last twenty years.
So the more abstract the product is for the user, the more gigantic the physical industry standing behind it. Building this kind of physical infrastructure takes years, while demand for it is shooting up exponentially. Physical supply moves with a huge lag. And when furious demand slams into slow supply, a colossal shortage appears. The speed of the digital revolution starts to depend on how long it takes to build things in the physical world.
So it may well turn out that AI, having started as a software revolution, will end as a new industrial revolution.
Electricity is becoming the raw material of intelligence
Conversations about AI are turning more and more often into conversations about energy. Electricity is gradually becoming one of the inputs for producing intelligence.
A country or region with cheap energy, infrastructure, the ability to build fast, and capital to invest gets an advantage not only in producing kilowatts. It gets an advantage in producing the capacity without which the new digital economy doesn’t work.
Historically, the supply of labor changed very slowly. For one more trained worker to appear in the economy, it took years of schooling and growing up. AI, for the first time, lets us partly increase productive capacity through investment in capital and electricity. You build a data center, install computing power, launch agents, and get an additional volume of digital work.
In this sense, data centers can be seen not only as token factories but also as factories of digital workers.
Profit always runs to where the bottleneck is
The bottleneck in an economy keeps traveling along the value chain, and profit keeps traveling right behind it. This is one of the most useful insights about economics in general. Profit concentrates not where the most beautiful or most hyped technology is created. It concentrates where the scarce complementary resource is.
That’s why AI is right now heating up completely unexpected, deeply conservative markets: energy, transformer manufacturing, industrial cooling systems, construction, and specialized project finance. Digital intelligence is launching a gigantic physical investment supercycle.
Right now we see a market with a wide middle. Many tasks need good-enough intelligence at an affordable price, while maximum capabilities are needed by a smaller share of companies. Frontier models widen the range of tasks that can be automated at all. At the same time, cheaper “good enough” quality lets already-available capabilities be adopted at mass scale. Technical progress turns into the economics of future business.
Two futures. And in both, someone ends up with no margin
In five to ten years, the AI industry will most likely settle into one of two paths. It’s not yet clear which one will win. But you can already see who will fight, and over what.
In five to ten years
Two futures. In both, someone is left without margin
Centralized path
- a few frontier labs
- AI as a utility, by subscription
- little in-house expertise needed
the platform takes the profit
Fragmented path
- base models remain
- specialized systems around them
- like the auto industry: platforms and thousands of companies
margin goes to those who build their own layer
The first path is centralized. A few frontier labs build general-purpose models so strong that it stops making sense for most companies to maintain their own layer of intelligence. In this world, AI becomes like electricity or the cloud: you buy it like a utility, through a subscription. Simple, cheap, good enough. Your own expertise is barely needed.
The second path is fragmented. Base models remain. But around them grows a dense ecosystem of specialized systems: medical, legal, financial, industrial, robotic. The economy starts to resemble the auto industry. There are a few big platforms, and around them, thousands of companies making components, assemblies, and solutions for specific tasks. The general-purpose model can do a lot, but without a specialized layer, the result stays raw and unusable for real work.
It’s hard to say right now which scenario will prevail. But the structure of incentives is clearly visible. The labs building frontier models want their model to do everything and capture as much margin as possible. Companies building products for specific scenarios want the opposite: to prove that without their process, their data, and their quality criteria, the general-purpose model produces a result nobody can use.
The more capability stays inside the base model, the more profit the platform takes. The more value is created through specific understanding of the task, routing, checks, and an embedded process, the more stays with those who build products and applications.
Let me take a short detour through a few examples: some for companies that make money specifically on their own AI products, and one for a company with a real, physical business that has spent years making workwear for people in heavy industry.
What we mean by “margin” here
What margin means
Why does the customer pay you?
The customer pays 100
you have what the general-purpose model doesn’tthe margin stays with you
you neatly wrapped what the platform already doesthe margin goes to the platform
Margin in this context is the part of the money that stays with a company after it has paid for the base intelligence (model access), and that a competitor would struggle to take away quickly. Put more simply: a customer pays you 100 units.
You pay the platform 20 for tokens and compute.
That leaves 80. The question is why the customer is willing to pay those 100 to you specifically, rather than straight to the platform or to another player. If the answer is “because you have something the general-purpose model doesn’t,” the margin stays with you. If the answer is “because you just neatly wrapped what the platform can already do,” the margin will drift to the platform over time.
Example 1. Contract review
A company sells automated contract analysis.
If the system mostly just reads text well and picks out standard clauses, that’s something the base model itself will learn to do over time. It’ll become easier and cheaper for the customer to buy it straight from the platform. Your margin shrinks.
But if the system knows this particular company’s internal rules, remembers which wording led to disputes before, can check compliance with their specific procurement policy, and passes the lawyer only what really needs attention, that’s a layer the platform won’t reproduce on its own. The customer pays for that layer. The margin stays with you.
Example 2. Insurance claims processing
A general-purpose model can extract the diagnosis and procedures from a doctor’s conversation. Over time, that will become cheap and widely available.
But if your system understands which wording is critical specifically for calculating reimbursement at your insurer, which errors are expensive, how to route complex cases properly, and how to cut the number of manual corrections specialists have to make, the customer (or internal client) pays for that accuracy and the embedded process. That part of the value stays as your margin.
In short: you keep the margin when you give the customer a reason to keep paying you rather than the platform or a competitor who just plugged in the same model.
In the “token costs” example, I was talking about companies whose product itself is an AI service. A workwear factory has different economics. Here AI isn’t what you sell the customer. It’s a tool inside your cost of goods. Token spending really might be less than 1–2% of total costs. But that doesn’t make the question “where does the margin stay” go away. It just shifts.
What “margin” means here
A factory’s margin is the difference between the price you sell workwear at and the full cost of making it and servicing the orders. AI can affect that margin in two ways:
Lowering your internal costs (processing orders faster, fewer errors in specifications, less manual work).
Letting you take on more orders, or more complex orders, without growing headcount proportionally.
The question is whether these savings will remain a lasting advantage of your factory in particular, or whether competitors will easily reproduce them.
What this looks like in practice
Say you hand routine processes over to agents: sorting incoming customer requests, checking specifications, preparing quotes, tracking deadlines, answering standard questions about sizes and materials.
A workwear factory
Who keeps the savings?
Agent “out of the box”
- savings right now
- a competitor has the same in a year or two
savings become common to the industry
Agents in your process
- your templates and fabric suppliers
- error history by customer
- learns from your orders
the savings stay with you
If you simply plug in a ready-made, general-purpose agent “out of the box,” you get some savings right away. But in a year or two, a competitor will be able to install the same agent. The savings will become common across the industry. Your margin won’t grow relative to others — everyone will just become a bit more efficient, and the pressure on price will increase.
If instead you build agents into your specific process (your order templates, your fabric suppliers, your kitting rules, your history of errors with specific customers, the particulars of workwear certification), and the system learns from your real orders over time, then the savings become part of your operating model. To get the same effect, a competitor won’t be able to just buy access to a model. They’ll have to reproduce your process and your accumulated data. In that case, the lower costs or higher throughput stay with you longer and have a stronger effect on margin.
To put it briefly, even when the spending on AI itself is pocket change, the strategic question is the same. Either you’re just speeding up current processes with something everyone will soon have, or you’re building intelligence into what’s specific to your factory and hard to copy. In the first case, the margin erodes over time. In the second, it stays with you.
The finale
Progress moves scarcity further along
It just moves it further down the chain. First, frontier models make experiments cheap. Thousands of apps and agents appear. The successful ones start devouring tokens. A margin problem arises. Companies look for cheaper specialized models. Specialization requires your own data and post-training. Data becomes an advantage. The company digs deeper into its niche. And the deeper it specializes, the more inference it needs.
Then it turns out the bottleneck isn’t the model anymore. The bottleneck is capacity.
And capacity requires infrastructure.
And at the end of the chain, suddenly, there’s a power plant. That’s how it always goes. The biggest opportunities in technological revolutions appear not in the technology itself but a couple of steps away from it. The internet generated wealth for website builders, but also for data centers and digital advertising businesses. The car enriched not only the factories but also gas stations, suburbs, motels, and road builders. If programming becomes ten times faster, there will be a shortage of strong product decisions.
If research speeds up tenfold, the bottleneck will be labs and testing grounds.
If agents start performing billions of tasks, first inference will run short, then chips, then megawatts. Right now, thousands of startups are simultaneously stress-testing hundreds of niches. Most will fail spectacularly. Like those .com companies whose names nobody remembers anymore. But it’s exactly out of this noise that it will become clear which scenarios were obvious, which were too early, and which ideas died as standalone companies — and, ten years later, came back to life inside someone else’s product. When the froth of cheap money settles and the fear of missing out dies down, the only thing left on the market will be what can create proven economic value. Everything else is just noise.