AI Tokens Explained: Why Falling Costs Are Fueling the AI Revolution

There's a tiny unit of value sitting at the centre of the biggest spending race in human history. It's not Bitcoin. It's not equity. It's called a token. And once you understand it, AI finally makes sense.

Understanding tokens explains everything happening in AI right now. The $600B infrastructure buildout. The data centres in Memphis. The satellites going into orbit. All of it.

What is a Token in AI?

Think of a token as the basic unit AI uses to think and respond. Not quite a word. Not quite a letter. Somewhere in between.

• 750 words ≈ 1,000 tokens
• A quick chat with AI = a few hundred tokens
• An AI agent running overnight = millions of tokens
• Every single token costs real money to produce
• Behind each token: specialised chips costing $25,000–$40,000 each, running in massive datacentres worth billions

Types of AI Tokens

Not all AI tokens represent the same kind of information. A token is simply a small piece of text or data that an AI model processes, but the way tokens are created and used depends on the input, output, and model architecture.

Understanding the main types of tokens helps explain why AI usage costs can vary so significantly from one task to another.

1. Input Tokens

Input tokens are the pieces of information you send to an AI model. They can come from a question, document, system instruction, conversation history, or data provided through an API.

For example, when you ask an AI assistant to summarize a 10-page report, the report itself becomes part of the model's input. The longer and more detailed the input, the more tokens the model typically needs to process.

Input tokens matter for businesses because they can accumulate quickly in applications that repeatedly send large prompts, documents, or conversation histories.

2. Output Tokens

Output tokens are the tokens generated by the AI model in response to an input.

A short answer may require only a few dozen tokens, while a detailed report, software code, or multi-step analysis can generate thousands of tokens.

For many AI applications, monitoring output tokens is particularly important because longer responses can increase both processing requirements and usage costs.

3. Cached Tokens

Some AI platforms allow frequently reused information to be cached rather than processed from scratch every time.

This can be useful when an application repeatedly sends the same instructions, documentation, context, or other information to a model. Depending on the provider, cached input may be priced differently from standard input tokens.

For AI applications with large system prompts or recurring context, caching can therefore become an important part of token-cost optimization.

4. Reasoning or Thinking Tokens

Some newer AI models use additional internal computation to work through complex problems before producing a final response. These are often described as reasoning tokens or thinking tokens.

They are particularly relevant to tasks involving mathematical reasoning, coding, planning, research, and multi-step problem solving.

The important point is that an AI response that looks short to a user may still involve substantially more computation behind the scenes. This is one reason token monitoring can provide a more useful picture of AI usage than simply counting visible words.

5. Tokens Used by AI Agents

AI agents introduce another layer of token consumption. Unlike a simple chatbot interaction, an agent may perform multiple model calls while completing a single task.

An agent might:

1. Receive a user's request.
2. Plan the task.
3. Call a tool or API.
4. Read the returned information.
5. Evaluate the result.
6. Adjust its approach.
7. Generate a final response.

Each model interaction can consume additional input and output tokens. As a result, a single agent workflow can use substantially more tokens than a straightforward question-and-answer exchange.

Why Token Types Matter for AI Costs

For businesses, token usage is more complicated than simply asking how many words an AI application generates. Costs can depend on input tokens, output tokens, cached context, model-specific processing, and the number of model calls involved in an AI workflow.

This is why organizations deploying AI at scale need visibility into token consumption. Monitoring usage by application, model, workflow, or team can reveal which processes are driving costs and where optimization opportunities exist.

As AI moves from simple chat interfaces toward autonomous agents and complex multi-step workflows, understanding these different forms of token usage becomes increasingly important. The better you understand what is consuming tokens, the easier it becomes to control AI costs without limiting useful AI adoption.

The Biggest Price Drop in Technology History

• In 2022, one million tokens cost $20
• By late 2024, the same thing cost 7 cents
• That's a 280x collapse in under two years
• Solar panels took 40 years to move that far, AI did it in two

So companies must be spending less on AI now, right? Wrong.

AI Token consumption by Model

Cheaper Tokens = More Spending. Every Time.

• Enterprise AI spending: $11.5B in 2024 → $37B in 2025
• That's a 320% increase, while prices were collapsing
• Average company now spends $85,000/month on AI
• 45% of companies spend over $100,000/month, that figure doubled in one year

This isn't an anomaly. It's history repeating itself.

Look Back at the Industrial Revolution

When the steam engine made coal cheap and efficient, the assumption was simple: less fuel needed. Smarter engines, lower consumption.

The exact opposite happened.

Cheaper coal unlocked things that were previously too expensive to bother with. New factories opened. New machines were built. Entire new industries appeared from nowhere. Coal consumption exploded, not from waste, but from new possibility.

Every time the token price drops, 10 new use cases become possible that weren't before.

• Contracts that wouldn't get reviewed by AI
• Research that wouldn't get funded
• Products that wouldn't get built
• Automation projects that would never have made it through the business case

We are watching the Industrial Revolution, but for intelligence.

AI Agents Are Pouring Petrol on This Fire

Normal AI chat uses a few hundred tokens. AI agents are a fundamentally different category.

An agent doesn't just respond. It thinks, acts, checks the result, rethinks, and acts again, in loops, for hours, sometimes overnight.

• A coding agent fixing a single bug = 50,000 to 500,000 tokens in one session

• A research and build task = millions of tokens, easily
• Agents consume 10–100x more tokens than a normal conversation
• And they're becoming mainstream fast

The steam engine didn't just power one factory. It powered the era. AI agents are doing the same thing, at a speed nobody anticipated.

The Infrastructure Arms Race This Creates

When demand explodes, infrastructure has to keep up. The numbers here are genuinely staggering.

• Hyperscalers (Amazon, Google, Microsoft, Meta) spent $260B on AI infrastructure in 2024
• That jumped to $450B in 2025
• It will cross $600B in 2026
• Goldman Sachs projects $1.15 trillion in total hyperscaler spend by 2027
• Elon Musk spent ~$18B filling three buildings in Memphis with 500,000 chips
• The Stargate project, OpenAI, SoftBank, Oracle is targeting $500B over four years
• SpaceX filed to put one million computing satellites in orbit, data centres in space, powered by the sun, cooled by the coldness of space

These aren't projections or wishful thinking. These are real filings. Real buildings. Real capital already deployed.

The Shift Nobody Is Talking About

For the first time ever in 2026, running AI costs more than building it.

• Inference (generating responses for users) now accounts for 55% of all AI cloud spending
• That's projected to reach 70–80% by 2030
• The industry has moved from creating intelligence to delivering it at massive scale

• The factory is built. Now we're running it at full capacity.

This is the inflection point. The centre of gravity has moved.

What This Means for Your Business

If you're deploying AI or evaluating it, token economics should already be informing your decisions.

• Not all tasks need expensive models, knowing which is which saves serious money
• AI agent costs compound fast if you're not actively tracking them
• Per-seat software pricing is dying, usage-based models are replacing it
• By 2028, seat-based pricing for AI-enabled workflows is projected to be obsolete

The businesses that understand this now will make better vendor decisions, smarter build-vs-buy calls, and clearer ROI cases for AI investment.

The Bottom Line

Token prices keep falling. Demand keeps exploding, just like coal did in the 1800s. Agents multiply that demand by 10–100x. And infrastructure investment is responding at a scale the world has never seen.

The Industrial Revolution ran for over 150 years. We are two years into the AI version.

The companies and people who understand token economics will make better decisions, about what to build, where to invest, and where this is heading.

Follow the tokens. Everything else follows from there.

About Fruition Services

Fruition Partners with AI market leaders like monday.com, Atlassian Hubspot we specialise in AI strategy and implementation, with 750+ client implementations across Australia, the UK, and the US. We help business leaders understand and deploy AI in ways that drive real operational results.