Try this thought experiment.
Every time you ask AI to write, read, summarise, rewrite, check, explain, translate, or think through something, a meter starts running.
Not metaphorically.
Literally.
AI companies don’t charge per question.
They charge per token.
A token is just a small chunk of text. Roughly:
- 1 token ≈ ¾ of a word
- 1,000 words ≈ 1,300 tokens
So every time you paste something in or get something back, you’re paying for the amount of language being processed.
Pricing varies by model, but as a rough guide:
- input tokens cost a few dollars per million
- output tokens cost several times more
That sounds abstract, so make it real.
If you paste in a 1,000 word document and ask AI to rewrite it:
- ~1,300 tokens go in
- ~1,300 tokens come out
Total: ~2,600 tokens
On typical models, that’s on the order of a couple of cents.
Cheaper models can be less.
More powerful models can be significantly more.
Which is why none of this feels expensive.
Because it isn’t.
At least, not for you.
The reason it feels free is that it’s being subsidised.
Massively.
Companies like OpenAI are valued in the tens or hundreds of billions, not because of today’s profits, but because of a bet:
That in the future, most human communication will be mediated through AI systems.
Emails. Documents. Customer support. Code. Search. Learning. Thinking.
If that happens, they don’t just sell software.
They sit in the middle of language itself.
And every time language flows through them, the meter runs.
Now bring it back to you.
One AI interaction might cost a couple of cents.
So you don’t think about it.
But imagine AI becomes your default layer for communication.
You use it to:
- rewrite an email
- summarise a meeting
- check a proposal
- draft a reply
- explain a document
- polish a Slack message
- turn notes into a blog post
Say that’s 20 interactions a day.
At a few cents each, that’s:
- around 50 to 60 cents a day
- roughly $15 to $20 a month
- around $200 a year
Still feels manageable.
Now zoom out slightly.
A company with 1,000 people working like this:
- $500 to $600 a day
- $15,000 to $20,000 a month
- around $200,000 a year
And those are conservative numbers.
If the average interaction is longer, includes documents, multiple prompts, or uses more advanced models, the cost climbs quickly.
And here’s the important part.
Those numbers are not even necessarily what you are paying today.
They are what the language costs somewhere in the system.
Right now, much of that cost is being absorbed.
By AI companies.
By SaaS products bundling AI into existing plans.
By employers covering usage.
By investors funding aggressive growth.
That’s why it still feels free.
The bill exists.
It just hasn’t reached you yet.
And subsidies change behaviour.
They make it easy to:
- ask instead of think
- generate instead of write
- summarise instead of read
- rewrite instead of edit
Each step adds another layer of processing.
Another layer of tokens.
Another layer of cost.
There’s a deeper issue.
These systems are not perfectly reliable.
They predict what sounds right.
Which means:
- you pay to generate words
- you pay to check those words
- you pay to fix those words
Even a small error rate pushes the real cost higher.
You’re not just paying for language.
You’re paying for uncertain language.
And here’s the uncomfortable part.
We are not just using AI.
We are being trained to depend on it for communication.
To mediate thinking through it.
To route everyday language through paid systems.
Once those habits set, they are hard to undo.
Not because we can’t.
Because we won’t want to.
The canary in the coal mine
The thing to watch for is pricing.
Not the demos.
Not the launch videos.
Not the benchmark scores.
Pricing.
Right now, most AI products still feel like subscriptions.
$20 a month. $30 a month. Predictable.
But underneath, many of them are already running on usage-based economics. Tokens in. Tokens out. Meter running.
Companies like Microsoft are starting to expose this at the edges, especially in enterprise tools and developer platforms.
That’s the transition phase.
The real shift comes when that usage-based pricing moves to the surface.
When:
- subscriptions start including credits
- limits become visible
- overage charges appear
- better models cost more per use
That’s when the subsidy ends.
That’s when the language tax stops being hidden.
And starts showing up on the invoice.
And when that happens, usage changes.
Because metering changes behaviour.
When every prompt has a visible cost, people ask fewer throwaway questions. Companies start asking whether every workflow really needs AI. Teams begin optimising prompts, restricting models, and cutting unnecessary usage.
The magic starts looking more like a budget line.
And that’s when valuations begin to normalise too.
Because if usage falls once pricing becomes visible, the market has to reprice the dream.
Not every AI interaction will be worth paying for.
Not every “AI-powered” product will justify its cost.
And not every company sitting in the flow of language will deserve a premium multiple.
That might be the real signal.
Not that token pricing appears.
But that once it does, people start using less.
Words used to cost time.
Now they cost tokens.
And without really noticing, we’ve started building a world where every sentence carries a small, persistent charge.
A language tax.
Not just for us.
For everyone who comes after.
Leave a Reply