The inference model R1 launched by the Chinese startup DeepSeek is comparable in performance to OpenAI's o1 model but at a significantly lower cost. The cost is only about $0.14 to $0.55 per million input tokens, far lower than the $7 to $15 for the o1 model. This gap has drawn market attention: will the cost of using AI continue to decrease as technology matures?

The cost of using AI decreases by about 10 times each year

Just today, OpenAI CEO Sam Altman published an article sharing his three observations on AI economics, one of which focuses on the abrupt decline in AI costs.

Altman pointed out that the cost of using any specific level of AI decreases by about 10 times each year, and lower prices will lead to more usage. He cited GPT-4 and GPT-4o as examples, where the price per token dropped by about 150 times from early 2023 to mid-2024, enough to prove the authenticity of this trend.

He further emphasized that the speed at which AI costs are declining is astonishing compared to Moore's Law, which doubles performance every 18 months.

As the costs of AI continue to decline, Altman anticipates:

  • The prices of many goods will ultimately drop significantly because many current costs come from limitations in intelligence and energy.

  • Conversely, the prices of luxury goods and inherently scarce resources (such as land) may experience more severe increases.

Altman's Law of the AI Industry?

Looking back at the history of the semiconductor industry, Moore's Law successfully predicted the trajectory of chip performance doubling every 18 months, becoming the golden guideline for driving industry progress. Many tech giants have also followed suit, proposing growth forecasts specific to their fields.

Now, does Sam Altman's argument that 'AI costs decrease by 10 times each year' have the potential to become the 'Altman's Law' of the AI industry?

If this observation continues to hold true, it will become the golden rule for the cost evolution of the AI industry model and set a new benchmark for future technological development.

The other two observations

Additionally, the other two observations made by Sam Altman are:

1) The level of intelligence of AI models is approximately equal to the logarithm of the resources used to train and run them.

These resources mainly include training computing power, data, and inference computing power. Altman observed that as long as enough funds are invested, continuous and predictable performance improvements can be achieved, and this phenomenon aligns with scaling laws across multiple orders of magnitude.

2) Linearly increasing AI intelligence has super-exponential social and economic value

This means that Altman observed that even if AI intelligence only exhibits linear growth, the resulting socio-economic value will show super-exponential explosions. He further stated that a conclusion can be drawn that, in the short term, there is no reason to believe this level of investment boom will stop.

Altman stated that if these three observations continue to hold, the impact of AI on society will be tremendous. Economically, it may become a significant scientific breakthrough that can be applied on a large scale and deeply integrated into various industries, much like transistors.

Just as we don't particularly focus on transistors or related companies, their value permeates every corner of life. Similarly, in the future, we may no longer specifically discuss AI itself, but we will naturally expect products such as computers, televisions, cars, and toys to possess stunning intelligent features.

Extremely reducing costs helps democratize AI

On the other hand, as AI technology continues to advance, Altman believes the risk of power imbalance between labor and management may come faster than expected, thus requiring 'proactive intervention' rather than waiting for problems to worsen before attempting to solve them.

Altman proposed a somewhat 'strange' idea, such as providing a certain 'compute resource budget' for everyone on Earth, allowing everyone to fairly access a large amount of AI resources.

At the same time, he noted that continually lowering AI costs could achieve similar effects. If costs approach zero, then AI resources could become as ubiquitous as air or tap water, benefiting everyone without additional policy intervention.

Finally, Altman elaborated on his goals:

By 2035, anyone should be able to utilize the equivalent of the collective intelligence of humanity in 2025, and everyone should be able to use unlimited genius and guide it according to their imagination.

There are now many talents who cannot fully showcase their potential due to lack of resources. If we can change this situation, the creative output generated globally will bring great benefits to all of us.