Der Preis von KI ist um das 900-fache gefallen. Selbst eine Flasche Mineralwasser ist teurer als sie.

[Introduction] In the past year, the price of AI models has plummeted a hundredfold! The same sentence that cost 10 yuan last year now only costs a few cents. Meanwhile, the prices of housekeeping, childcare, psychological counseling, repairs... these "manual jobs" are getting more and more expensive. Technology is experiencing extreme deflation, while the cost of living is rising. This is not an economic joke, but a reality created by Jevons and Baumol: when machines become smarter, human labor becomes more expensive.

"The token cost of large language models (LLMs) has been dropping a hundredfold every year. The most advanced models are even dropping at a rate of 900 times."

Last week, this sentence appeared in a post on X.

There was no accompanying picture or long - form argument, just a cold string of numbers.

But it was like a nail, hitting the nerve of the entire AI industry hard. The comment section was flooded with one sentence: "Moore's Law is back."

The poster was Rohan Paul, an analyst who has been tracking the AI cost curve for a long time. He summarized it more straightforwardly:

The price of AI is collapsing at a speed never seen in any technology in human history.

A few days later, Alex Danco, a researcher at a16z, responded:

Whenever a technology becomes cheaper, human labor becomes more expensive.

Two seemingly independent posts unexpectedly formed a cycle.

The price of AI is falling, while the value of human beings is rising. Technological deflation is creating new inflation.

I. Plummeting AI, Inflating Humans

In the past year, the price of AI models has been dropping almost in a "free - fall" manner.

Researchers have calculated that since the end of 2022, the cost of using a GPT - 3.5 - level model has dropped from about $20 per million tokens to the current $0.07 - a full 280 times cheaper!

There are significant differences in the price collapse speed of different - level models.

Another analysis from a16z also points out that the inference cost of LLMs is decreasing at a rate of 10 times per year, comparable to the rebirth of Moore's Law.

The time - based collapse curve of AI model prices. Taking the lowest - cost model that reaches MMLU ≥ 42 points as an example, the cost has dropped from $60 per million tokens in early 2022 to less than $0.1 in 2024.

Previously, generating a novel might cost dozens of dollars, but now it can be done for just a few cents.

AI is changing from a luxury to "tap water." You don't feel bad about using it, and you might even can't help but use more.

But things aren't as simple as they seem. As Rohan Paul said:

A price drop doesn't mean savings; it's a new consumption explosion point.

When model calls become cheap, we start to embed AI everywhere: writing copy, doing translations, writing code, editing videos...

The demand for computing power has not decreased but increased, which has pushed up the prices of energy and hardware.

Meanwhile, those jobs that cannot be done by AI, such as housekeeping, nursing, psychological counseling, and repairs... are becoming more and more expensive.

Repairing an air - conditioner or hiring a nanny costs far more than training a medium - sized model once.

The AI world is experiencing deflation, while the human world is experiencing inflation.

So, a seemingly absurd reality is happening: AI is getting cheaper, but life is getting more expensive.

Behind the collapse of algorithm costs is the reorganization of labor value: the parts that can be replaced by machines are rapidly depreciating, while the parts that cannot be replaced have become new "luxuries."

II. The Cheaper, the More Addictive: The Jevons Paradox in AI

In 1865, British economist William Jevons wrote that famous warning:

Don't think that greater efficiency means saving fuel - it will only make us burn more.

The plummeting cost curve of AI. For the cheapest model that meets MMLU ≥ 42, the cost has dropped from $60 per million tokens for GPT - 3 to $0.5 for GPT - 3.5 - turbo, and then to less than $0.1 for Llama 3.2 3b.

At that time, he was referring to the steam engine. But 160 years later, this sentence is being repeated with AI.

When the model price keeps dropping, we thought it was the "era of saving money."

But the fact is, a decrease in cost only makes people more willing to use it.

Previously, a company had to hold a meeting to approve the purchase of computing power. Now, even small and medium - sized enterprises can call GPT - 4 Turbo with a single click; previously, a person carefully asked ten questions in ChatGPT, but now they open ten tab pages and run scripts crazily.

AI is no longer a tool but a resource with "unlimited refills."

Microsoft CEO Satya Nadella also mentioned in an interview:

The Jevons Paradox is back: the more efficient and cheaper AI is, the more people can't do without it.

AI is changing from a "high - end intelligent service" to a "new public facility" - like electricity, water, and Wi - Fi, everyone is using it.

This is the so - called "Jevons Paradox in AI": when the use becomes too cheap, humans stop being restrained.

The more you use it, the deeper your dependence.

So we see that companies call models on a large scale for A/B testing; creators use AI for batch generation; research institutions fine - tune models repeatedly; developers run parallel inferences on a dozen channels.

Every price drop brings a new round of "abuse." Every efficiency improvement leads to new waste.

Computing power, energy, and chips are being re - consumed in this "addictive growth" process.

III. As Machines Get Smarter, Humans Become "Luxuries"

While the price of AI is dropping, a strange contrast is happening: algorithms are getting cheaper, but human labor is getting more expensive.

In the United States, the hourly fee for housekeeping services has risen to $45; in the UK, the hourly wage of a plumber even exceeds that of a lawyer.

Human inflation in the era of AI deflation: the wage curve of repair workers. The change in the annual salaries of various skilled workers in the United States from 2020 - 2025. The wage growth rate of air - conditioner repair (HVAC) is significantly higher than the national median.

A machine - learning engineer might only need a few cents to complete an inference, but a repair worker might charge hundreds of dollars to come to your door.

This is not an accident but an inevitability predicted by economists.

In the 1960s, Jevons proposed the Baumol's cost disease theory.

He found that in industries with higher productivity (such as manufacturing and technology), prices are more likely to drop; while in industries where efficiency cannot be improved, such as performing, teaching, nursing, and repairing, prices are "dragged" up by the overall economic wage level because they need to retain workers.

In short: The dividends of efficiency make low - efficiency jobs more expensive.

Applying this model to today, it's not hard to find that AI is a typical high - productivity industry, with efficiency increasing a thousand times and costs plummeting a hundredfold;

But in those fields that cannot be replaced by AI: education, psychological counseling, manual repairs, and elderly care are being caught up in the price - increase wave.

As a16z wrote in its analysis:

Technological deflation often creates human inflation.

When computing power becomes like tap water, the truly scarce things are human time, emotions, and presence.

So we see that algorithms are depreciating, while human personalities are appreciating.

Jobs that can be automated are dropping in price, while jobs that require a "human touch" have become the "luxuries" of the new era.

This might be the most ironic paradox in the AI era - as machines become smarter, being a "human" becomes the most expensive thing.

IV. Technology is Getting Cheaper, Power is Concentrating

The repeated price drops of models sound like a "universal benefit for all."

Everyone can use it, and everyone can access it. AI finally seems to have become a public resource.

But the reality is the opposite - the cheaper AI gets, the more concentrated the power becomes.

Over the years, the main players in the price - drop have never been the open - source community, but a few giants: OpenAI, Anthropic, Google, and Amazon.

While they are "generously" reducing API prices, they are also redefining the entry points.

The cheaper the models are, the more developers rely on them; the more popular the services are, the more they monopolize computing power, data, and algorithm standards.

This is the new pattern in the AI era: the price is dropping, but the control is increasing.

An article in MIT Technology Review wrote:

When a technology becomes "free," monopoly often becomes invisible.

The price drop of AI, seemingly democratic, is actually an acceleration of platformization.

Every time we call a model, we are unknowingly training the next - generation models for large companies.

The price collapse doesn't make the world more equal; it just allows us to hand over more power at a lower cost.

Meanwhile, a group of "sandwiched humans" has emerged.

They don't write models and won't be replaced by models for now. Instead, they are sandwiched between the giant systems and algorithms: prompt engineers, data annotators, AI reviewers, fine - tuning operators...

Their jobs are short - term and repetitive, but they maintain the operation of the entire AI system; they are the labor amplified by AI and the people digested by algorithms.

AI has indeed become cheaper, but what's truly expensive is human time and attention that can control AI.

So we see a new paradox: technology is in deflation, while power is in inflation.

The more affordable the models are, the more closed the ecosystem is; the more popular AI is, the more centralized the center is.

Perhaps this is the deepest hidden line in the "era of AI deflation" - we thought we were moving towards universal benefit, but in fact, we are entering a world where costs approach zero and power is unified.

What we are experiencing is not just a technology price drop but a value reorganization.

Models are getting cheaper, algorithms are getting faster, and efficiency is flooding like a tide. But what's truly diluted is the "definition of a human."

While AI is swallowing up repetition and reducing costs, it is also increasing scarcity in the opposite direction: creativity, emotions, judgment, and companionship, these parts that cannot be calculated by algorithms, have become the new generation of "high - value assets."

In the future, perhaps it's not AI replacing humans, but "AI pricing humans."

Some people will be replaced; some people will have their prices increased.

It's the models that are dropping in price, and it's the human heart that is appreciating.

References