For most of technological history, transformative technologies became cheaper gradually.
Electricity took decades to become dramatically more affordable. Computing followed a remarkably powerful but relatively steady improvement curve. Lithium-ion batteries required decades of manufacturing improvements before their costs fell enough to transform transportation and energy storage. DNA sequencing went from an extraordinarily expensive scientific procedure to something available for hundreds of dollars.
Artificial intelligence may be doing something different.
According to a new analysis from Epoch AI, the cost of obtaining a given level of AI performance has fallen by roughly 47% per quarter since 2023. That translates to approximately a 13-fold reduction every year. Epoch argues that this rate is faster than the historical cost decline of other transformative technologies it examined—including DNA sequencing, computing, lithium batteries and electricity.
That finding is bigger than simply saying that ChatGPT or other AI models are becoming cheaper.
It suggests something much more consequential:
The cost of machine intelligence itself is collapsing.
And if that trend continues, the economic consequences could extend far beyond the AI industry.
The crucial distinction: AI isn't merely getting cheaper
When people hear that AI costs are falling, they often think about cheaper API tokens.
That is only part of the story.
Epoch AI attempted to measure something more meaningful: how much it costs to achieve a particular level of AI capability.
In other words, rather than asking:
“How much does one million tokens cost?”
the more useful question is:
“How much does it cost to make an AI system perform at this level?”
That distinction changes the picture dramatically.
A model can become more expensive per token while simultaneously becoming more efficient at solving a particular problem. Conversely, a model can become cheaper per token without delivering meaningful improvements in capability.
Epoch's analysis attempts to hold performance approximately constant and then track how the cheapest way of achieving that performance changes over time.
Across five major benchmarks covering mathematics, science and games of skill, Epoch estimates an average cost decline of about 47% per quarter since 2023.
That means the same amount of machine intelligence that previously required a dollar might eventually require only a fraction of that amount.
And then that process repeats.
47% per quarter is an extraordinary number
A 47% quarterly decline is difficult to intuitively grasp.
If a cost falls by roughly 47% every three months, it is reduced to approximately 53% of its previous level each quarter.
After four quarters, that compounds to roughly:
0.53⁴ ≈ 0.079
In other words, the cost becomes roughly 8% of its previous level after a year.
That is approximately a 13-fold reduction annually.
And then the next year begins from that already-lower base.
This is why the phrase “AI is getting cheaper” understates what is happening.
The more accurate description is:
The amount of intelligence you can purchase with a fixed amount of money is increasing extremely rapidly.
That is a fundamentally different economic phenomenon.
AI versus the technological revolutions before it
Epoch compares this rate of decline with several historically important technologies.
Its analysis estimates that the rate of AI cost reduction is approximately:
| Technology |
Relative rate compared with AI |
| DNA sequencing |
AI declining ~4× faster |
| Computing |
AI declining ~6× faster |
| Lithium batteries |
AI declining ~18× faster |
| Electricity, through 1973 |
AI declining up to ~54× faster |
These comparisons should be interpreted carefully.
The technologies are measured differently, the datasets cover different periods, and “performance” is not directly comparable across industries.
Epoch itself warns that its dataset is relatively short, incomplete and noisy. It also notes that benchmark performance does not necessarily equal real-world usefulness, and that companies may optimise models specifically for particular benchmarks.
So this isn't proof that AI has mathematically defeated every technological learning curve ever recorded.
But it is still a remarkable signal.
DNA sequencing shows what exponential cost decline can do
Consider DNA sequencing.
The cost of sequencing an entire human genome has fallen from roughly $100 million in 2001 to just over $500 in 2023, according to WIPO.
That collapse helped transform genomics from an elite scientific undertaking into a technology increasingly used in medicine, research and biotechnology.
But AI's reported trajectory is substantially steeper.
And there is an important difference.
Sequencing produces information about biological molecules.
AI can potentially process information, reason over it, generate new information and perform cognitive tasks.
That makes the economic consequences of falling AI costs potentially much broader.
If DNA sequencing becomes cheap, we can sequence more genomes.
If intelligence becomes cheap, we can potentially perform more thinking.
The arrival of “cheap cognition”
This may ultimately be the most important way to understand the trend.
For most of human history, cognitive labour was constrained by the number of human beings capable of performing it.
A company needing 10,000 hours of research could hire researchers.
A government needing thousands of documents reviewed could employ analysts.
A pharmaceutical company needing millions of scientific papers examined could hire teams of scientists.
Those activities were limited by:
- human time,
- salaries,
- training,
- geography,
- availability of expertise,
- fatigue,
- coordination.
AI changes the equation.
If the cost of obtaining a certain level of machine intelligence keeps falling, organisations can potentially purchase dramatically more cognition.
One researcher could eventually command hundreds or thousands of AI systems.
One small company could conduct analysis that previously required a large department.
A student could access a tutor capable of generating personalised explanations indefinitely.
A software developer could deploy AI agents to inspect code continuously.
A scientist could have machines analyse enormous quantities of literature before beginning an experiment.
The bottleneck begins moving from:
“Can we afford enough intelligence?”
toward:
“What should we use all this intelligence for?”
The paradox: cheaper AI does not necessarily mean less AI spending
This is one of the most counterintuitive implications.
Normally, when something becomes dramatically cheaper, people assume total spending will collapse.
That isn't necessarily what happens.
If the price of intelligence falls 90%, users may simply consume vastly more intelligence.
Imagine that analysing one scientific paper costs $1.
A research organisation might analyse only 100 papers.
Cost: $100.
If the same analysis eventually costs one cent, the organisation could analyse 100,000 papers.
Cost:
$1,000.
The unit price collapsed.
Total spending increased.
This is an example of what economists call an elastic demand response.
Recent analysis from Citadel Securities similarly points to rapidly increasing AI consumption among firms as AI costs fall, suggesting that lower prices can encourage substantially greater usage rather than simply reducing expenditure.
Epoch makes a similar point: even extremely low per-task costs could produce enormous aggregate spending if the amount of cognition demanded is sufficiently large.
Intelligence could become an abundant resource
Human intelligence has historically been relatively scarce.
Highly trained mathematicians are scarce.
Experienced engineers are scarce.
Doctors are scarce.
Researchers are scarce.
Software developers are scarce.
Management consultants are scarce.
Translators are scarce.
Teachers are scarce.
Even ordinary human attention is scarce.
AI potentially changes the economics of that scarcity.
If sufficiently capable AI can be replicated at near-zero marginal cost, then an organisation doesn't need to ask:
“Can we find another expert?”
It can increasingly ask:
“How many instances of this intelligence do we need?”
That is an extraordinary transition.
We have already seen something similar with software.
Once software has been created, reproducing another copy is extremely cheap.
AI potentially adds something new:
software that performs cognitive work.
That could make intelligence itself partially reproducible.
And the capabilities are rising while the price falls
This is what makes the current AI trajectory particularly unusual.
Technological revolutions normally involve some combination of:
higher capability + lower cost.
AI appears to be experiencing both simultaneously.
The cost of reaching a fixed capability level is falling.
But the frontier of capability is also moving upward.
Epoch's research illustrates this with individual benchmarks.
For example, it reports that the cost of achieving a particular performance threshold on GPQA Diamond fell dramatically between OpenAI's o3 and later models.
It also reports that the cost of achieving a particular FrontierMath performance threshold fell hundreds of times in roughly 18 months.
This creates a powerful combination:
AI is becoming cheaper at yesterday's intelligence while simultaneously becoming capable of tomorrow's intelligence.
That is potentially much more disruptive than simple price competition.
The frontier may actually fall fastest at first
Another fascinating finding from Epoch's analysis is that cost declines aren't uniform.
When a particular performance level first becomes possible at the frontier, the cost of achieving it can initially fall extremely quickly.
Across its five primary benchmarks, Epoch estimates that the cost of newly achieved state-of-the-art performance fell around 66% per quarter, equivalent to approximately a 75-fold annual decline.
Two years later, the decline slows to approximately 32% per quarter, or about 4.7-fold annually.
There is an intuitive explanation.
When a new capability first appears, it is expensive.
Only a few organisations can provide it.
Then competition begins.
New models appear.
Inference becomes more efficient.
Hardware improves.
Algorithms improve.
Developers discover better ways to use the models.
Prices fall.
Eventually, the technology becomes mature and further cost reductions become harder.
This means the cheapest AI available today may represent only the beginning of the economic impact of a capability that was expensive when first introduced.
This could change the economics of software
Software has already been transformed by cloud computing and open-source development.
AI adds another layer.
Imagine a startup with five employees.
Instead of hiring:
- a junior developer,
- a researcher,
- a data analyst,
- a customer-support team,
- a copywriter,
- a QA team,
the company could potentially deploy AI agents to handle portions of those jobs.
The result isn't necessarily mass unemployment.
It could also mean massive increases in the productive capacity of small teams.
A two-person company might be able to operate like a much larger organisation.
The economic question becomes whether AI primarily:
substitutes for workers
or
augments workers so dramatically that new economic activity expands faster than existing jobs disappear.
History offers evidence for both possibilities.
But the unprecedented speed of AI cost declines makes the adjustment period particularly important.
The developing world could experience this differently
This is where the story becomes particularly important for countries such as Nigeria and other African economies.
For decades, developing countries have faced a structural problem:
expertise is expensive and infrastructure is scarce.
A rural entrepreneur might not have access to:
- an experienced lawyer,
- an accountant,
- a software engineer,
- a marketing strategist,
- a research analyst,
- a language tutor,
- an international business consultant.
AI potentially changes the economics.
If high-quality cognitive services become extremely cheap, geographical distance becomes less important.
A Nigerian startup could potentially access sophisticated software development assistance from an AI system.
A farmer could use AI for crop planning.
A small exporter could receive assistance with documentation.
A student could have an always-available tutor.
A journalist could process thousands of documents.
A small manufacturer could use AI for design and quality control.
This does not eliminate the need for physical infrastructure.
You still need:
- electricity,
- internet connectivity,
- computers and phones,
- data centres,
- payment systems,
- logistics,
- education,
- institutions.
But the cost collapse could remove one historically important constraint:
access to cognitive expertise.
Yet there is a major danger: cheap intelligence does not mean equal access
There is a temptation to assume that because AI becomes cheap, everyone automatically benefits.
That is not guaranteed.
The infrastructure behind AI remains highly concentrated.
The most advanced models require enormous quantities of:
- computing power,
- specialised chips,
- electricity,
- data,
- capital,
- engineering talent.
And the companies controlling these systems can determine:
- pricing,
- access,
- geographic availability,
- data policies,
- model restrictions,
- API terms.
This creates a paradox.
Intelligence may become cheap for the consumer while the infrastructure supplying that intelligence remains concentrated in a small number of companies and countries.
That distinction could become strategically important.
The new geopolitical resource may be compute
Oil shaped geopolitics in the twentieth century.
Semiconductors are shaping geopolitics today.
AI could make compute and energy even more strategically important.
If intelligence becomes cheap because enormous computing systems are becoming increasingly efficient, countries without access to compute may become dependent on foreign infrastructure.
That creates questions of technological sovereignty.
Who owns the data centres?
Who manufactures the chips?
Who supplies the electricity?
Who controls the models?
Where is the data stored?
Who sets the rules?
Who captures the economic value?
These aren't abstract questions.
They determine who benefits from the falling price of intelligence.
There is another paradox: cheaper AI could increase the demand for energy
AI may be becoming cheaper per unit of intelligence while requiring enormous amounts of physical infrastructure.
More efficient AI encourages more usage.
More usage requires more inference.
More inference requires more computing.
More computing requires electricity.
More electricity requires generation and transmission infrastructure.
This produces an unusual technological feedback loop:
AI gets cheaper → people use more AI → demand for compute rises → demand for electricity rises → investment in infrastructure rises.
That helps explain why the AI boom can simultaneously produce falling inference costs and intense competition for chips, data centres and electricity.
The falling price of intelligence does not mean the physical world disappears.
Quite the opposite.
Cheap digital cognition could increase demand for expensive physical infrastructure.
AI may become like electricity—but faster
The comparison with electricity deserves particular attention.
Electricity was one of the foundational general-purpose technologies of industrialisation.
Once electricity became widely available, it didn't simply replace steam engines.
It reorganised factories.
It transformed communications.
It enabled appliances.
It changed cities.
It altered transportation.
It created entirely new industries.
AI has the potential to behave similarly because cognition is embedded in almost every economic sector.
But there is a critical difference.
Electricity powers machines.
AI can potentially help operate the machines.
Electricity provides energy.
AI provides decision-making and information processing.
That could make AI a particularly powerful general-purpose technology.
The biggest economic question may not be “How smart will AI become?”
It may be:
How much intelligence will society consume once intelligence becomes extremely cheap?
This is a different question.
Suppose AI eventually becomes 100 times cheaper.
Companies don't necessarily use the same amount of AI and pocket the savings.
They may use 100 times more.
And if AI becomes 1,000 times cheaper?
Perhaps they use 10,000 times more.
The demand for cognition may be enormous because the world contains an almost unlimited number of tasks that humans currently leave undone because they are too expensive.
There are billions of documents nobody reads.
Millions of businesses without sophisticated analytics.
Countless scientific papers nobody has time to examine.
Millions of students without personalised tutors.
Huge quantities of software that could be improved but isn't.
Millions of small businesses that cannot afford professional expertise.
If intelligence becomes cheap enough, many of these previously uneconomic activities become economically viable.
But benchmarks aren't the real economy
This is where the Epoch research needs to be interpreted carefully.
The researchers themselves acknowledge several limitations.
Their study examines a limited set of benchmarks.
Benchmark performance isn't identical to real-world economic productivity.
AI companies can optimise for benchmarks.
Users don't always choose the cheapest model capable of a task.
And the dataset spans only a few years.
Epoch therefore says its headline estimate should not be interpreted as an exact universal law.
That caveat is extremely important.
Passing a difficult mathematics benchmark for fractions of a cent does not mean that an AI can replace a scientist for fractions of a cent.
Real-world work includes:
- context,
- coordination,
- verification,
- liability,
- physical action,
- communication,
- institutional knowledge,
- trust.
The cost of generating an answer is not necessarily the cost of producing a useful outcome.
Nevertheless, the direction is difficult to ignore
Even with those caveats, the broader trend is significant.
AI models are becoming:
more capable,
more efficient,
more autonomous,
and increasingly cheaper to deploy.
That combination has few historical precedents.
The most important consequence may not be that today's AI applications become cheaper.
It may be that entirely new applications become possible because their economics previously didn't work.
When something costs $1,000,000, only a few people use it.
At $10,000, more people can use it.
At $100, it becomes mainstream.
At $1, it becomes routine.
At $0.01, people start using it everywhere.
That is the fundamental economic significance of technological deflation.
The “AI revolution” may actually be a revolution in the price of cognition
The public conversation often focuses on whether AI is intelligent.
But an equally important question is:
How much does intelligence cost?
Human intelligence is expensive because humans need:
- years of education,
- food,
- housing,
- healthcare,
- salaries,
- rest,
- social infrastructure.
AI systems don't have those same requirements.
They require computation, energy, hardware and software.
If engineers continue reducing the cost of computation and improving model efficiency, the cost of machine cognition could continue falling.
That could turn intelligence from a scarce service into something approaching an abundant commodity.
And abundant commodities reshape economies.
The historical significance may only become obvious later
People rarely recognise the full significance of a technological cost curve while they are living through it.
When computing became cheaper, people didn't initially know that smartphones, cloud computing, social media and generative AI would eventually emerge.
When sequencing became cheaper, the full implications for biotechnology weren't immediately obvious.
When electricity became cheaper, nobody could predict every industry that electrification would eventually create.
AI may be similar.
The biggest applications of extremely cheap machine intelligence may not exist yet.
That is perhaps the most important implication of the Epoch research.
The question isn't simply what today's AI can do.
The question is what becomes economically possible when tomorrow's AI capabilities cost a tiny fraction of today's.
The coming decade could be defined by abundance of intelligence
If the current trend persists—even at a substantially slower rate—the economic consequences could be enormous.
We could move from:
scarce human cognition
to
abundant machine cognition.
From:
“Can we afford an expert?”
to:
“How many AI experts should we deploy?”
From:
“Which tasks deserve human attention?”
to:
“Which tasks should humans still perform themselves?”
And eventually perhaps from:
“How can we make AI affordable?”
to:
“How do we manage a world in which intelligence is cheap?”
That last question may be the most difficult one.
Because when intelligence becomes abundant, intelligence itself stops being the scarce resource.
Judgment becomes scarce.
Trust becomes scarce.
Energy becomes scarce.
Compute becomes strategic.
Good data becomes valuable.
Human attention becomes more valuable.
And perhaps most importantly:
Knowing what to ask an intelligent machine to do becomes an economic skill in itself.
The 47% quarterly figure should therefore not be read simply as another impressive AI statistic.
It points toward something much larger.
For the first time, humanity may be experiencing a technological revolution in which the cost of performing cognitive work is falling at extraordinary speed while the capability of the machines performing that work is simultaneously rising.
If that continues, the biggest AI disruption may not be a single model, chatbot or robot.
It may be the moment when intelligence itself becomes cheap enough to be used everywhere.
And history suggests that when the cost of a fundamental capability collapses, the world doesn't merely buy more of the old product.
It invents things that were previously impossible.
Comments
Post a Comment