OpenAI's three observations on the economics of AGI

If you zoom out on the history of technology, most breakthroughs eventually turn into cheap infrastructure. But when the cost of raw intelligence collapses at a speed that makes Moore's law look sluggish, how does the rest of the economy adapt?

OpenAI CEO Sam Altman laid out his perspective in a new blog post titled Three Observations. According to him, society is now standing right at the threshold of artificial general intelligence, and the technical trajectory ahead looks remarkably clear.

Crushing Moore's law on cost

Altman's first observation focuses on model capabilities. An AI model's intelligence roughly equals the logarithm of the resources poured into training and running it, primarily training compute, data, and inference compute. Scaling laws have proven accurate across multiple orders of magnitude, meaning capital can be turned into predictable, continuous capability gains.

His second point is about pricing. The cost of a given level of AI drops roughly tenfold every 12 months, triggering massive surges in usage. Altman pointed to OpenAI's own pricing trajectory: between early 2023 with GPT-4 and mid-2024 with GPT-4o, the cost per token plunged by around 150 times. Moore's law delivered a 2x improvement every 18 months; this is radically faster.

Third, the socioeconomic value of linearly increasing intelligence is super-exponential. Because each step up in capability unlocks disproportionately massive economic utility, Altman sees no reason for the exponential surge in AI infrastructure capital expenditure to slow down anytime soon.

He expects the next practical phase to arrive through AI agents acting like virtual colleagues. A software engineering agent, for instance, could soon handle multi-day tasks comparable to a junior developer with a few years of experience at a top firm. It will lack groundbreaking ideas and still demand close human oversight, but fielding thousands or even millions of them simultaneously will upend knowledge work.

Capital, labor, and a decade to 2035

Altman predicts that by 2035, an individual could marshal cognitive capacity equivalent to all of humanity in 2025. While daily life in 2025 will look mostly familiar, long-term structural shifts will be massive.

In this environment, human agency, resilience, and personal determination will become the ultimate competitive moat. Rather than diminishing individuals, Altman argues that AGI will act as an unprecedented force multiplier for human will.

However, broad distribution of these benefits will not happen automatically. Altman warned that the balance of power between capital and labor could easily destabilize, likely requiring early policy intervention. He floated unconventional ideas, such as granting every person on Earth a baseline compute budget, alongside relentless efforts to drive the marginal cost of intelligence toward zero.

Notably, Altman wrapped up the post with a pre-emptive footnote, emphasizing that his casual use of the term AGI does not alter or trigger the contractual definitions governing OpenAI's partnership with Microsoft.

Why it matters

Altman is effectively writing a theoretical defense of the staggering capital expenditure driving the frontier AI race.

While critics increasingly question whether massive data center investments will ever yield proportionate returns, Altman's argument is that spending arbitrary sums remains the only rational playbook if cost drops 90% annually while economic returns compound super-exponentially. It is an explicit signal that OpenAI will continue to build aggressively, forcing hyperscaler competitors to match pace or fall behind.

For builders and creators, the takeaway is stark: building thin wrappers or reselling basic inference is a dead end because commodity intelligence is plunging toward zero cost. The real leverage lies in orchestration, domain expertise, and sheer agency—knowing exactly what complex goals to pursue when you can command a swarm of virtual engineers.

Yet the structural tension in Altman's vision is hard to ignore. If junior-level knowledge work can be spun up on demand for pennies, traditional talent pipelines and career ladders could break before society figures out what replaces them. Compute credits might sound egalitarian, but they will hardly solve the acute displacement of human labor if capital captures the lion's share of that super-exponential wealth.