The Deflation of Intelligence

AI is getting cheaper. So why are we going to spend more on it than ever? There is something strange happening to the price of intelligence. Not human intelligence. Machine intelligence. My team saw this and im more than happy to write about it to share with you(human or agent). In 2026, OpenAI and Anthropic are fighting for the same increasingly valuable territory: the right to become the cognitive infrastructure behind software, companies, agents, and eventually entire digital workforces. The obvious part of that competition is model capability. Better reasoning. Better coding. Longer context windows. More reliable agents and the demand is not going anywhere any time soon. The less obvious part is price. And price may end up being the more economically important story. Of course, tokens are not perfectly interchangeable commodities. A token processed by one model is not necessarily equivalent to a token processed by another. Context, personal setup, caching, reasoning quality, tools, latency, and agentic performance all complicate direct comparisons. But the direction is difficult to ignore: The cost of accessing machine intelligence is falling and competition is accelerating the process. That sounds like good news for consumers. It is the case. But it also creates one of the strangest economic dynamics of the AI era: AI is getting cheaper precisely as we prepare to spend more money on it than ever before. "Intelligence is becoming deflationary" For most of human history, sophisticated cognitive labor has been expensive because it required humans: • Writing a contract required a lawyer. • Analyzing a company required an analyst. • Designing software required a programmer. • Researching a market required researchers. • Producing a marketing campaign required writers, designers, strategists, and coordinators. Intelligence was bundled with labor and labor is scarce. AI begins separating the two. A company no longer necessarily purchases an hour of someone's time. It purchases inference. Tokens go in. Tokens come out. Somewhere in between, a machine performs a small amount of cognitive work. That unit of work is becoming cheaper extraordinarily quickly. Every improvement in hardware utilization, model architecture, inference infrastructure, competition, caching, quantization, and specialized silicon pushes the effective cost downward. The result is something that economists rarely encounter in labor markets: the price of cognitive production can fall dramatically in a matter of months. • A human accountant does not normally become 80% cheaper next quarter. • A software engineer does not suddenly accept one-fifth of his salary because a competitor released a better engineer. Machine workers can. And they will. This is why thinking about AI purely as software may eventually become misleading. Tokens are not units of labor. But economically, the workloads they enable are beginning to compete with labor The Jevons paradox of intelligence The obvious assumption is that cheaper AI should mean lower AI bills. It may mean exactly the opposite. In the nineteenth century, economist William Stanley Jevons observed something counterintuitive about coal. More efficient steam engines did not reduce coal consumption. They made steam power economically viable for more activities. Coal consumption increased ➡️ Efficiency increased demand. The same effect may now be emerging with intelligence. When AI inference was expensive and unreliable, companies used it selectively. A human might occasionally ask a model to summarize something, generate code, or draft a document. The architecture looked roughly like this: one human → one model → a few prompts But as inference becomes cheaper, something fundamental changes. The human no longer needs to invoke the model manually. Software can invoke it constantly. And models can invoke other models. An agent receives a task: 1. It reads documents. 2. It searches databases. 3. It generates a plan. 4. It writes code. 5. It tests the code. *The test fails. 1. It reads the error. 2. It modifies the code. 3. It calls another model to review the solution. *That model disagrees. The original agent investigates again 🔄. Millions of tokens can disappear before a human even looks at the result. So imagine that the cost of intelligence falls by 80%. A company could save 80% or it could consume twenty times more intelligence. The second possibility may prove far more important. This produces a paradox: The cheaper intelligence becomes, the more economically rational it becomes to waste intelligence. And "waste" is not necessarily a bad word here. Humans already do this with electricity: • We illuminate empty hallways • Keep servers running overnight. • Cool entire buildings. • Charge devices while we sleep, etc. Electricity became sufficiently cheap and useful that constantly consuming it became normal and machine intelligence may follow the same path. From software to workers This change becomes easier to understand when we stop imagining AI as a chatbot because a chatbot waits, a worker acts. Agents increasingly sit somewhere between the two. Consider a future software company with fifty employees. It may also operate: • 40 coding agents. • 20 research agents. • 12 customer-support agents. • 8 monitoring agents. • 6 sales agents. • 5 accounting agents. • 3 cybersecurity agents. And hundreds of short-lived subagents created dynamically when necessary. Nobody will necessarily call them employees (accounting systems may classify them as cloud expenses), but economically, something resembling a labor force has appeared. And unlike human workers, this workforce can operate under extremely unusual economics. • It does not need to sleep. • It can be copied. • It can work simultaneously on thousands of tasks. • Its memory can be duplicated. • Its tools can be upgraded instantly. And, most importantly, its cost can continuously decline. That means tokens may slowly stop feeling like API usage and start feeling like machine payroll. Payroll becomes compute This transition has enormous implications for where money moves through the economy. Today, a company might spend: • $10 million on salaries. • $500,000 on cloud infrastructure. • $100,000 on AI. Now imagine that AI systems become capable enough to meaningfully multiply each employee's output. Five years later the same organization might spend: • $7 million on salaries. • $2 million on AI inference. • $1 million on compute infrastructure. Even without completely replacing workers, the composition of business expenditure changes. Capital moves away from some forms of labor and toward: • GPUs. • Datacenters. • Networking. • Electricity. • Cooling. • Model providers. • Inference platforms. • Agent infrastructure. What used to be payroll slowly becomes compute. That does not automatically mean economic contraction. If a company spends $10,000 on agents and those agents generate $50,000 of additional economic value, society has not simply lost $10,000 of consumption. Productivity increased. New income may appear elsewhere(or may not?). But the distribution of that income changes radically. The economic beneficiaries increasingly sit upstream: • chip manufacturers, • energy producers, • datacenter operators, • cloud companies, • model laboratories, and whoever ultimately owns the productive AI systems. That redistribution may become one of the defining political and economic questions of the next decade. The uncomfortable economics of the AI labs Now we arrive at the contradiction. AI companies are selling a product whose effective unit price keeps falling, while producing that product requires extraordinary amounts of capital. Cheaper inference does not mean cheaper infrastructure. Better models require more compute. More users require more inference capacity. Longer context windows require more memory. Agentic systems generate far more tokens than traditional chatbots because they plan, search, call tools, retry, verify, and sometimes delegate work to other models. So the economics begin to look uncomfortable: the price of intelligence is falling at the same time that the machinery required to produce intelligence is becoming more expensive and more abundant. That creates a peculiar industrial race. Labs must continuously invest in larger clusters, better chips, networking, datacenters, and energy just to remain competitive. But competition itself pushes the price of inference downward. The loop looks like this: More capital → better models → stronger competition → lower prices → more usage → more infrastructure → more capital As long as usage grows quickly enough, the system works. A provider can charge less per token while earning more money because customers consume dramatically more tokens. This is hardly unprecedented. Computing itself followed a similar trajectory. • Storage became cheaper, so we stored more data. • Bandwidth became cheaper, so we streamed video. • Processors became cheaper, so software became more computationally ambitious. The question is whether intelligence exhibits the same elasticity. Because if it does, the market could become astonishingly large. But what if it doesn't? This is where the possibility of a crash enters the conversation. AI companies and infrastructure providers are effectively making an enormous bet: that future demand for intelligence will justify today's investment in producing it. That bet may be correct. But it is still a bet. Imagine enormous datacenter capacity being constructed under the assumption that businesses will deploy millions of agents. Chip production expands. Energy contracts are signed. Power infrastructure is built. Debt is issued. Investors price companies around extraordinary future consumption. Then imagine that something goes wrong. • Perhaps agent reliability improves more slowly than expected. 🧨 • Perhaps companies discover that many agentic workflows do not generate enough economic value. 🧨 • Perhaps inference becomes so efficient that much less infrastructure is needed. 🧨🧨 • Perhaps open models compress margins. 🧨🧨 • Perhaps competition pushes token prices downward faster than consumption grows. 🧨🧨🧨 That combination could produce overcapacity. And overcapacity is one of the oldest ingredients in financial crashes. • Railroads had it. • Telecommunications had it. • The dot-com era had it. • Real estate has had it repeatedly. The revolutionary nature of a technology does not protect investors from paying too much for the infrastructure surrounding it. In fact, historically, revolutionary technologies often attract precisely that kind of overinvestment. A crash would not mean AI failed This distinction matters, and its kind of logical once we reach this point of the lecture. Financial bubbles and technological revolutions are not opposites, they frequently coexist. The dot-com crash did not prove that the internet was useless, it proved that expectations, valuations, and capital deployment had temporarily outrun reality. Much of the physical and digital infrastructure survived and companies built on top of it later became some of the largest businesses in history. Something similar could happen with AI: • There could be enormous overinvestment in datacenters. • There could be bankruptcies. • There could be consolidation among model providers. • There could be dramatic repricing of companies whose valuations assume unrealistic growth. And at the same time, AI could continue transforming the economy, we can see a recent case in Bitcoin for the money vector. A crash in AI assets would not necessarily represent the failure of artificial intelligence, it could simply represent the market discovering the correct price of the infrastructure, and this is expensive sometimes but an evil needed. Or perhaps demand really is enormous There is another possibility where maybe we are underestimating demand. Humans have historically rationed intelligence because intelligence was expensive. Consider all the things companies do not analyze today simply because assigning a human to analyze them would cost too much. Every customer interaction. • Every contract. • Every transaction. • Every production line. • Every piece of source code. • Every security event. • Every sales lead. • Every supplier. • Every internal process. Most organizations operate with enormous amounts of information that nobody has time to examine and cheap machine intelligence changes that. Suddenly : • Everything can be watched. • Everything can be summarized. • Everything can be optimized. • Every employee can have ten agents. • Every small business can have capabilities that once required an entire corporate department. And machines themselves can become consumers of intelligence. This last point may be the most important. Today there are roughly billions of humans capable of generating demand. Tomorrow there may be billions of software processes capable of requesting inference. The largest customer of an AI model may eventually not be humans, it may be and shall be other software. At that point the potential demand curve becomes difficult to imagine for a simple and humble human like me. The price of a thought Perhaps this is the wrong way to think about tokens entirely. A token is a technical accounting unit. Nobody actually wants tokens. People wants • Answers. • Code. • Decisions. • Plans. • Research. • Designs. • Actions. • Work. Tokens simply measure part of the machinery producing those things and that machinery is becoming cheaper extraordinarily quickly. For the first time, humanity may be approaching an economy where the marginal cost of producing a small amount of useful cognitive effort continuously declines. That does not make intelligence free. 1. Electricity is not free. 2. Computing is not free. 3. Storage is not free. But each became cheap enough to reorganize civilization around it. Intelligence might be next. If so, the defining economic question of the AI era will not simply be: How intelligent can machines become? It may be: What happens when intelligence becomes too cheap to ration? We do not yet know whether today's infrastructure boom ends in scarcity or overcapacity. We do not know whether AI spending eventually cannibalizes enormous categories of human labor or creates entirely new categories of economic activity. We do not know whether falling inference prices destroy margins or unlock enough demand to compensate for them and we certainly do not know whether the enormous capital flowing into AI today will earn the returns investors expect. But one trend is increasingly difficult to miss. The commodity underneath this revolution is getting cheaper. And our appetite for it is getting larger. Perhaps the strangest economic story of AI is not that machines are becoming intelligent. It is that intelligence itself is becoming deflationary. The price of a thought is falling. The number of thoughts we are willing to buy is exploding. Somewhere between those two curves lies one of the largest economic experiments of our lifetime.