The Cheaper the Tokens, The More Insatiable the World

In November 2021, the GPT-3 API opened to the public. The price: $60 per million tokens.

In July 2024, GPT-4o-mini was released. The price for similar tasks: $0.15 per million tokens.

In less than three years, the price dropped by 400x.

If tokens were merely “tools for writing articles,” a 400x price reduction should have brought only incremental growth—people might use ChatGPT a few more times, polish a few more emails, and that would be it. Demand has an upper limit, because a person only has 24 hours a day; you won’t read ten times more articles just because AI writing becomes cheaper.

But what actually happened?

China’s average daily token consumption surged from 100 billion in early 2024 to 180 trillion in February 2026. That’s an 1800x increase in eighteen months.

OpenAI’s enterprise customers saw a 320x year-over-year increase in inference token consumption per organization.

The total volume of tokens processed by Google’s products grew from 480 trillion per month in May 2025 to 1.3 quadrillion in October—a 2.7x increase in five months.

Prices dropped by 400x, consumption grew by 1800x.

This isn’t the Jevons Paradox “possibly holding true.” This is the Jevons Paradox playing out in an extreme form in reality. The elasticity of demand is not slightly greater than 1—it is far, far beyond 1. The demand unleashed by falling token prices is far more violent than the price drop itself.

The question is: Where are all these extra tokens going? Who is “consuming” them? And why can the elasticity of demand be so astonishing?

The answer lies in the layered structure of token consumption.

From FLOPS to Tokens: What Does the Price-Demand Curve for Computational Resources Look Like?

a16z gave this trend a name: LLMflation. For LLMs of equivalent performance, inference costs drop by 10x each year. This pace is faster than the cost reduction of computing power during the PC era (Moore’s Law) and faster than bandwidth cost reduction in the internet era (Edholm’s Law).

This isn’t the first time this has happened.

CPU computing power has decreased by roughly 10^10 (ten billion) times over the past 60 years. During the same period, global total computational consumption has increased by roughly 10^15 (one quadrillion) times. It dropped a trillion-fold, but usage increased a quadrillion-fold. Demand growth outpaced price reduction by a factor of a hundred thousand.

The same applies to bandwidth. And to storage. Every foundational computational resource follows the same pattern: When costs drop to a new order of magnitude, it’s not that “the same people use more” — rather, “a large batch of use cases that were previously completely outside the market suddenly flood in.”

The transition from mainframes to PCs wasn’t about “the same scientists computing a few more equations” — it was about billions of ordinary people starting to use computers for word processing, gaming, and watching videos. The shift from dial-up to fiber optics wasn’t about “the same people viewing a few more web pages” — it was about entire new categories of applications like Netflix, YouTube, and TikTok emerging that were previously impossible.

Tokens are replaying this history. But with a key difference—tokens are not an “infrastructure resource” (like electricity, bandwidth, storage), but rather a pricing unit for “cognitive services.” What you are buying is not “computing power,” but “a piece of thinking.”

This makes the layered elasticity of token demand more complex and also more interesting.

Three Types of Token Consumption, Three Types of Elasticity

Not all token consumption is the same. Viewing “1800x increase in token usage” as a monolithic number misses the most important structural information.

Layer 1: Human-Machine Conversational (Moderate Elasticity)

This is the token consumption mode most people intuitively think of—you open ChatGPT and ask a question, and it answers. You use Copilot to write code, and it offers suggestions. You ask Claude to polish an email.

Consumption at this layer is hard-constrained by human attention and time. You can only converse with AI so many times a day. You can only read and judge so much AI-generated content a day. Even if tokens were free, you wouldn’t ask AI ten thousand questions a day—because you don’t have time to digest ten thousand answers.

Therefore, the elasticity of this layer is moderate. If token prices drop 10x, individual usage might increase 3-5x (from “asking 10 questions a day” to “asking 30-50 questions a day”). But it won’t increase 100x. This layer is more like “translation”—you wouldn’t translate the same article ten times just because translation became cheaper.

Layer 2: Agent/Automation-Based (Elasticity Approaching Infinite)

This is the fastest-growing part of token consumption and the biggest contributor to the 1800x increase.

An AI Agent is not constrained by human time. It can run 24 hours a day non-stop. It can launch a hundred parallel branches. It can repeatedly attempt, self-correct, and iterate in loops until it reaches its goal. Each loop consumes tokens.

Let’s look from a programmer’s perspective.

Traditional usage: A programmer asks AI, “How do I write a sorting algorithm?” The AI responds with 500 tokens. Asking 50 questions a day = 25,000 tokens.

Agentic usage: A programmer gives Claude Code a task: “Refactor this authentication module.” The Agent starts working autonomously—reading code, analyzing dependencies, making a plan, writing code, running tests, finding bugs, fixing them, running tests again… A single task might consume hundreds of thousands or even millions of tokens. Developer reports show that monthly costs for agentic coding with Claude Code range between $500-2000 (pure API cost). The Agent team mode consumes about 7 times more tokens than regular conversations.

Andrew Ng did the math: If an AI agent runs continuously at 100 tokens/sec, at $4/M tokens, the cost is $1.44/hour—lower than the US minimum wage.

What does this mean? It means when tokens become cheap enough, programmers’ work methods will undergo a fundamental shift—from “think first, then write” to “let the AI run all possible solutions and pick the best one.” From “conserving tokens” to “splurging on tokens.”

In the past, when you debugged a bug, you: first thought of possible causes, tried two or three fixes, and if that didn’t work, consulted the documentation. Now you can let an agent run 20 fixes simultaneously, automatically test each one, and tell you which one passed. The token cost is a few cents. The human effort saved is several hours.

Why is the elasticity at this layer close to infinite? Because its constraint is not human time—but rather “the economic value of solving this problem > the token cost consumed.” As long as the value of solving a problem exceeds a few cents in token cost, it’s “worth it to run the agent.” And in the real world, there are almost infinitely many problems whose value exceeds a few cents.

The CEO of Alibaba said at the 2025 Apsara Conference: “Token consumption speed doubles every two to three months.” This pace is primarily driven by the agent layer.

Layer 3: Embedded/Platform-Based (Elasticity Depends on End-User Scale)

This is the primary source of Google’s 1.3 quadrillion tokens per month. It’s not humans conversing with AI—but AI being embedded into every request of products like search, email, documents, and maps. Users might not even be aware they are “consuming tokens”; they are just using Google Search, but behind every search, Gemini is running.

The elasticity of this layer depends on the user growth of the end products. When tokens become cheap enough to embed into every interaction of every product, token consumption essentially equals “the total requests of all digital products worldwide × the number of tokens per request.” This number has virtually no ceiling.

There is also an emerging sub-layer: Video and Multimodal. ByteDance’s Seedance 2.0 requires about 350,000 tokens to generate a 10-second 1080p video. A complete animation project might require hundreds of millions of tokens. When the token cost of video generation drops to an affordable level, the demand explosion in this sub-layer will be dozens or even hundreds of times greater than the current text layer.

From a Single Line of Code to a Billion People: Three Scales of Token Elasticity

By looking at the three layers of elasticity together, we can understand what “The cheaper the tokens, the more insatiable the world” truly means.

The Programmer Scale: From Tool to Environment

For programmers, tokens are evolving from a “tool” into an “environment.”

Tool Phase (2023-2024): You ask AI questions, and AI answers. You are the driver, and AI is the assistant. Token consumption is limited by how often you ask questions.

Environment Phase (2025-): You set goals, and AI explores autonomously. Your token consumption is no longer limited by “how many questions you asked,” but by “how large a search space you defined for the agent.”

A concrete example. Previously: You spend 30 minutes writing a function and 10 minutes debugging. Total token consumption: a few thousand (asked a few questions). Now: You spend 2 minutes describing what you want, and the agent spends 10 minutes autonomously trying 50 implementation approaches, running 200 tests, and selecting the optimal solution. Total token consumption: hundreds of thousands. But you saved 38 minutes.

When token costs fall below your time cost, the rational choice is “let the agent run more.” This is why token elasticity is extremely high from a programmer’s perspective—because a programmer’s time is valuable ($50-200/hr), while tokens are cheap ($1.44/hr for continuous operation). With each price drop, a new batch of problems becomes “worth it to let the agent run.”

The Company Scale: From Point to Surface

OpenAI’s Enterprise AI Report shows: Enterprise API call volume increased 8x year-over-year, and inference tokens per organization grew 320x.

A 320x increase is not “doing the same thing more often.” It means “processes that previously weren’t using AI have started using it.”

A company’s AI adoption path typically looks like: 1. A single team pilots (the product team uses AI to write documentation) → Token consumption: Low 2. Core processes go live (customer service uses AI for initial responses) → Token consumption: Medium 3. Comprehensive proliferation (Finance, Legal, HR, Operations all connect to AI) → Token consumption: High 4. Agentic transformation (AI no longer just assists humans; it starts autonomously executing tasks) → Token consumption: Explosive

Most companies in 2024 were in stages 1-2. Starting in 2025, they entered stage 3. The 320x growth reflects the proliferation from “point” to “surface.”

When tokens drop another 10x, stage 4 becomes economically viable. At this point, token consumption is no longer calculated “per headcount”—but “per task count.” A company might have 1,000 employees but 100,000 daily tasks that can be automated by agents. Each task consumes 100,000 tokens = 10 billion tokens per day = the cost might only be a few hundred dollars.

This is why token elasticity at the enterprise level can reach 320x—because the number of tasks “worth automating but previously too costly” within a company far exceeds the number of employees.

The Societal Scale: From None to Have

The largest elasticity is hidden in the third scale.

Globally, 1.2 billion people have never had access to personalized education services. It’s not that they don’t want it—a human tutor previously cost $30-100 per hour, unaffordable for most families. The marginal cost of an AI tutor is a few cents per hour in tokens. When tokens become cheap enough, these 1.2 billion people all become “newly entering consumers.”

A similar logic applies to: - Basic legal consultation (billions of people worldwide have never consulted a lawyer) - Mental health support (stigma + cost are dual barriers; AI solves both simultaneously) - Multilingual information access (vast amounts of quality content is English-only; AI translation lets users of smaller languages enter the market) - Data analysis for small and medium-sized businesses (previously required hiring an analyst; now costs a few hundred dollars/month)

The elasticity in these scenarios is not “existing users use more”—it’s “billions of people who were never in this market suddenly enter it.” This is the purest form of the Jevons Paradox: cost reduction unlocks a massive market that previously didn’t exist at all.

World Bank data corroborates this: 86% of EdTech companies in emerging markets have integrated AI, 70% in AgTech, 54% in Fintech. These are not Silicon Valley toys—these are foundational services about to cover billions of people.

What the Data Says

Putting all the numbers together:

Metric	Time	Value	Growth Rate
China’s Avg. Daily Tokens	2024.01→2026.02	100B→180T	1800x / 18 months
Volcengine Avg. Daily	Late 2024→2026.01	2T→63T	31.5x / 13 months
Google Monthly Volume	2025.05→2025.10	480T→1300T	2.7x / 5 months
Microsoft Azure Quarterly	2024 Q3→2025 Q3	~20T→100T	5x YoY
OpenAI API Avg. Daily	2025.10	8.6T	—
OpenAI Enterprise Inference/org	YoY	—	320x
Token Price (Equivalent Perf.)	YoY	—	-10x

Prices drop 10x each year. Consumption grows 100-1000x each year.

This means the total global token market spend (price × volume) is still growing rapidly—even as the unit price plummets. a16z points out an interesting phenomenon: The output token price for OpenAI’s top model, o1, is still $60/M—the same as GPT-3’s price three years ago. This shows that the market’s high end is using “more expensive tokens” to trade for higher-quality inference, while the low end is using “cheaper tokens” to cover more scenarios.

This is precisely a characteristic of a market with extremely high demand elasticity: Volume at the low end, price preservation at the high end. Similar to the smartphone market—an iPhone sells for $1000, but global smartphone shipments are 1.2 billion units/year because low-cost Android covers everyone.

What Token Elasticity Tells You to Bet On

Back to economics.

The third condition of the Jevons Paradox is “the elasticity of demand must be sufficiently high.” The actual data from the token market tells us: this elasticity is not slightly greater than 1, but high to a shocking degree. Prices drop 10x, consumption grows 100-1000x.

But the elasticity is not uniformly distributed.

Human-Machine Conversational Layer (you ask ChatGPT questions): Moderate elasticity. Constrained by attention. This layer is close to “translation”-type demand—it won’t be used ten times more just because it’s cheaper.
Agent/Automation Layer (AI autonomously loops and executes tasks): Extremely high elasticity. The only constraint is “task value > token cost.” This layer is close to “computational resources”—each order-of-magnitude drop unlocks a batch of new scenarios.
Globalized Services Layer (people who previously couldn’t afford it enter the market): Extremely high elasticity. A leap from 0 to 1. This layer is most similar to “cotton cloth”—billions of suppressed demands waiting to be released.

The implication for individuals is clear:

If you treat tokens as “a tool for writing articles,” you are consuming in the layer with the lowest elasticity. Your value is limited by the attention ceiling—you can only write/read so many articles a day.

If you treat tokens as “fuel for agents,” you are consuming in the layer with the highest elasticity. Your value is not constrained by your personal time, but by how many “problems worth solving” you can define.

If you treat tokens as “infrastructure for globalized services,” you are consuming in the market with the largest elasticity. You are facing not “existing users using more,” but “billions of new users starting from zero.”

The biggest mistake is: understanding tokens merely as “cheaper writing/translation/coding tools”—this is a low-elasticity perspective. The essence of tokens is the unit of measurement for universal cognitive infrastructure. Its demand elasticity curve is more like that of electricity, bandwidth, or computing power—each order-of-magnitude price drop spawns an entirely new application layer, and the total volume consumed by that new layer will far exceed the previous generation.

The 1800x growth we see today is likely just the beginning.

Appendix: Sources Cited in This Article

LLMflation Trend (a16z, Guido Appenzeller): https://a16z.com/llmflation-llm-inference-cost/
Global Token Consumption Tracker Data: https://robonomics.substack.com/p/token-tracker-and-implications
OpenAI State of Enterprise AI 2025: https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-enterprise-ai_2025-report.pdf
LLM Inference Cost Comparative Analysis (Introl): https://introl.com/blog/inference-unit-economics-true-cost-per-million-tokens-guide
Implications of Token Price Drops for AI Companies (Andrew Ng / DeepLearning.AI): https://www.deeplearning.ai/the-batch/falling-llm-token-prices-and-what-they-mean-for-ai-companies/
Agentic Coding Cost Analysis (Vantage): https://www.vantage.sh/blog/agentic-coding-costs
Claude Code Token Management Documentation: https://code.claude.com/docs/en/costs
AI Diffusion in Emerging Markets (World Bank): https://blogs.worldbank.org/en/psd/how-ai-travels–diffusion-among-firms-in-emerging-markets
Google Token Processing Volume (Official Blog): https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/gemini-enterprise-sundar-pichai/
ChatGPT User Data (TechCrunch): https://techcrunch.com/2025/07/21/chatgpt-users-send-2-5-billion-prompts-a-day/