We live in the age of extreme efficiency.
Language models that cost $60 per million tokens in 2021 now cost less than $0,06—a 1.000-fold drop in just three years. When GPT-3 became publicly available in November 2021, it was the only model capable of achieving an MMLU of 42, charging that small fortune per million tokens processed. Today, Llama 3.2 3B delivers the same performance for pennies. Not to mention new models with much better benchmarks.
With each new generation of Nvidia GPUs — Hopper H100, Blackwell B200, and upcoming architectures — inference frameworks evolve, quantization techniques improve, and the cost of processing artificial intelligence plummets on an even steeper curve than Moore's Law. We're talking about a reduction rate of at least 15x per year, with some benchmarks showing drops of up to 900x per year depending on the task.
But there's a paradox that few people are seeing: the efficiency that drives the adoption of IA It could be the same pin that bursts your financial bubble.
The Efficiency Trap
Companies that currently sell inference—that is, charge for the use of AI models—are trapped in an unavoidable dilemma: their costs fall 15 times per year, but so does the price they can charge. It's a exponential deflationnon-inflationary.
Let's do the math: a data center that costs $1 billion today needs to have its return on investment calculated based on prices that will be 15 times lower next year, and 225 times lower in two years. It's like building an oil refinery when the price of a barrel falls 15 times a year—the math simply doesn't add up.
Those who sell tokens wholesale and resell them as tools (chatbots, copilots, corporate assistants) will soon realize that the perceived value for the end user also decreases, as the barrier to entry disappears. A product that today requires a multi-million dollar infrastructure could be replicated by a startup at 1/1.000th of the original cost. Democratization is not just technological— It's economical.
And here's the crucial point: When your product gets cheaper faster than you can create demand, you don't have a sustainable business. You have a race to the bottom.
The Investment Paradox
While costs are plummeting, global capital expenditure (capex) on AI infrastructure — especially in data centers and GPUs — is growing at a record pace. Trillions of dollars are being poured into servers that, in a few years, may be underutilized, operating at negative margins, like the crypto mining farms abandoned in 2018.
The logic is fascinating in its contradiction: investors are funding the construction of an infrastructure based on the premise of scarcity (“the more computing power, the more value we create”), while the technology they are funding is proving that we will have abundance. And digital abundance is, by definition, deflationary.
It's the equivalent of investing billions in power plants while the efficiency of solar panels doubles every year. At some point, you realize you've built too much capacity for a market that needs less and less.
The problem isn't technological—we're winning on that front. The problem is economic. Projections show that by 2029, even in the conservative scenario, the cost per million tokens will fall to US$0,000019. In the optimistic scenario? US$0,000001. This means that revenue per computing unit will fall by more than... 3.000 times in the next five years.
How do you monetize something that tends toward zero marginal cost?
We've Seen This Movie Before
In the 19th century, the world witnessed a similar euphoria: The railroad bubble.
Investors poured fortunes into expanding the United States' railroad lines. The country was being stitched together from coast to coast, and every meter of track seemed a promise of infinite wealth. Companies were created purely for speculation—many without even a realistic operating plan.
When the bubble burst in 1873, the collapse was brutal. Northern Pacific Railroad It went bankrupt. The bank Jay Cooke & CompanyThe railroad, one of the most powerful in America, collapsed. A five-year economic depression followed. Investors lost fortunes. Thousands of railroad companies disappeared overnight.
But here's the twist: The tracks remained.
The legacy of that bubble was monumental. The rail infrastructure built during the speculative frenzy paved the way for the logistics that sustained the United States for more than a century. Farmers shipped grain to distant ports. Manufactured goods reached the interior. People moved between states with ease unimaginable decades before.
The bubble punished investors, but rewarded users.
The same thing happened with the dot-com bubble in the late 90s. Billions were burned on unprofitable startups, absurd business models, and companies that existed only in PowerPoint. pets.com It became a joke. Webvan is a case study in failure. Hundreds of dot-com companies were wiped out when the bubble burst in 2000.
But what remains? The fiber optic cables that connect the world. The TCP/IP protocols we use every day. The digital culture that defines the modern economy. The data centers, the e-commerce platforms, the cloud that stores our lives.
Bubbles build the future, even if they destroy the present for their investors.
The Next Bubble: Artificial Intelligence
The "AI bubble" is, in essence, an efficiency bubble.
Companies and governments are betting on unlimited growth in use and value, without realizing that technological progress itself undermines the economic fundamentals that underpin this expectation. It's like betting that gold will appreciate while you invent a machine to transmute lead into gold—the better your machine works, the less gold is worth.
The marginal cost of intelligence is approaching zero. And when marginal cost goes to zero, profit follows the same path — unless you're a monopolist, and in the world of open source models and commoditized hardware, monopolies are impossible to maintain.
The signs are all there already:
- Infrastructure companies competing for long-term contracts at unsustainable prices.
- Value aggregators without sustainable pricing power.
- Business models based on margins that melt faster than ice in the desert.
- Billion-dollar valuations of companies whose core product will become a commodity in 18 months.
The AI gold rush is creating a magnificent infrastructure—state-of-the-art GPUs, optimized frameworks, ever-improving models. But it's also creating a financial illusion: that this infrastructure will generate returns proportional to the investment.
It will not generate it. It cannot generate it. The math doesn't allow it.
Conclusion: Bubbles Build the Future
Bubbles are destructive — but also productive.
They punish investors who confuse euphoria with analysis. They sweep away companies that exist only to ride the wave. They expose the weak, the opportunists, those who built sandcastles on imaginary foundations.
But they also do something extraordinary: They build the future faster than any rational planning could.
No government committee would have approved the investment necessary to cover the United States with train tracks at the rate that speculative capitalism has done. No central plan would have installed the amount of fiber optics that the companies of the dot-com bubble installed. The irrationality of markets builds on a scale and at a speed that rationality cannot.
Just as the railroad bubble paved the way for industrial progress, and the internet bubble laid the foundation for the digital age, The AI bubble will pave the way for the cognitive infrastructure that the world will use for decades to come.
The bubble will burst, yes. Many companies will go bankrupt. Many investors will lose money. Many data centers will be underutilized.
But when the dust settles, what will remain is a world where artificial intelligence is cheap, ubiquitous, and integrated into everything. A world where consulting a language model will cost less than a fraction of a cent. A world where small businesses will have access to the same cognitive capabilities that are currently only available to tech giants.
The tracks for AI are already being built through speculation. Our task is not to invest in the tracks—it's to learn how to drive the trains.
Call to Reflection
Efficiency is good — until it becomes too efficient.
When intelligence costs zero, who will have a competitive advantage? It won't be those who own the models, because they will be open source. It won't be those who own the data centers, because capacity will be abundant and commoditized.
The advantage will go to those who know how to ask the right questions. To those who know how to integrate intelligence into real problems. To those who build upon the infrastructure that the bubble is creating.
The question remains: are you investing in building data centers, or in learning how to use what they produce?
The right answer could determine who survives when the bubble bursts — and who thrives after the dust settles.
