Sam Altman wants to build a gigawatt of AI infrastructure a week

Sam Altman doesn't want humanity to choose between curing cancer and giving every kid on Earth a personal AI tutor.

In a new blog post titled Abundant Intelligence, the OpenAI CEO laid out an eye-popping goal: building an industrial machine capable of delivering one gigawatt of new AI infrastructure every single week.

To put that in perspective, one gigawatt is roughly the output of a commercial nuclear reactor. Building that much capacity weekly is less like a tech upgrade and more like a wartime mobilization.

The gigawatt factory and the 2030 clock

Altman argues that widespread access to intelligence will eventually become a fundamental human right. Delivering it, however, requires a complete overhaul of the physical stack, spanning custom chips, grid power, construction, and robotics.

He specifically noted that OpenAI wants to build much of this capacity in the United States, warning that other nations are currently moving much faster on semiconductor fabrication and energy deployment. Formal plans, partners, and novel financing models will be announced later this year.

Meanwhile, speaking in Berlin after receiving the Axel Springer Award, Altman added an aggressive timeline to his hardware ambitions. He predicts artificial general intelligence and genuine superintelligence will arrive before 2030.

Altman also claimed that 30% to 40% of all tasks in today's economy will be taken over by AI in the not-too-distant future. Yet he dismissed existential dread, pushing back against the idea that superintelligence will view humans like ants. Instead, he echoed calls for building maternal instincts into systems, betting AI will treat humanity more like a loving parent.

On the political front, Altman welcomed the pro-business climate under President Donald Trump and predicted more world leaders will rely on AI for governance. He did draw one clear line, however: he doesn't ask ChatGPT for relationship advice.

Why it matters

We think Altman is deliberately shifting the AI narrative from algorithmic breakthroughs to raw industrial might.

For everyday developers, this confirms that the frontier layer is out of reach. If capital expenditure shifts to gigawatts per week, only nation-backed consortiums can play. Everyone else will simply buy inference tokens downstream.

Yet the vision faces a hard collision with physics. Power grids, transformer manufacturing, and nuclear licensing operate on decade-long lead times, not weekly software cycles. Framing compute as a human right looks like clever positioning ahead of OpenAI's upcoming financing reveal, keeping investors anchored to an inevitable future where compute equals revenue.