OpenAI's CEO puts the five-year cost of one gigawatt of AI capacity at $70 billion and says it needs to come down sharply for the industry to scale
Building the machinery behind artificial intelligence is staggeringly expensive. OpenAI CEO Sam Altman thinks the price tag needs to shrink by a lot.
Altman says the five-year cost of building and operating one gigawatt of AI infrastructure should fall from $70 billion to about $20 billion. That is a cut of more than two-thirds, and it is a target, not a done deal.
What Altman actually said
The comments came during a recent podcast appearance. Altman framed the figure as total cost of ownership, or TCO, over five years.
For AI data centers, that means chips, power delivery, facilities and the electricity to run them all. Altman’s current estimate for that full bundle is approximately $70 billion per gigawatt.
He described a drop to about $20 billion as essential for the industry to scale in a healthier way.
Altman pointed to two main levers. One is lower energy costs. The other is thinner margins across the supply chain.
He expressed skepticism that simply eliminating those margins would be enough without parallel gains in energy efficiency and supply chain dynamics. Energy supplies, he noted, remain tightly constrained.
The analysts mostly agree
Semiconductor analyst Dylan Patel estimates that current gigawatt construction costs exceed $60 billion.
That is up from roughly $40 billion in 2024.
Patel believes that without the profit margins collected by chipmakers like Nvidia and Broadcom, costs could realistically fall to approximately $20 billion.
Analysts have also flagged rising capital expenditure pressure from chips, power delivery and facilities as build-outs continue.
Why OpenAI cares so much
OpenAI has a very specific reason to obsess over this number. The company has made infrastructure commitments exceeding $1.4 trillion, tied to forecasts of revenue growth.
One concrete example is Project Camellia in Georgia. It is a $20 billion campus with a contracted capacity of 3.2 gigawatts, set for a phased rollout between 2028 and 2032.
The company is also pursuing vertical integration across its operations. That push reportedly spans semiconductor production, robotics and data centers.
The long-term ambition is striking. If costs come down, OpenAI aims to eventually add 1 gigawatt of capacity per week.
A precedent from the memory chip world
DRAM makers, the companies that produce computer memory, went through their own brutal capital cycle.
They moved from spending $20 billion in capital expenditures against $19 billion in revenue to generating $100 billion in annual free cash flow per maker.
What this means for the industry
Chipmakers like Nvidia and Broadcom currently enjoy margins that both Altman and Patel identify as a cost driver. A serious effort to compress those margins could shift the balance of power between AI labs and their suppliers.
Energy is the harder constraint. Supplier margins can be negotiated, competed away or engineered around through in-house production. Power availability cannot be conjured with a purchase order, and Altman himself pointed to tight energy supplies as a limiting factor.
OpenAI’s $1.4 trillion in commitments are tied to revenue forecasts, and those forecasts assume the scaling story plays out. If costs stay stubbornly near current levels, the gap between spending and revenue gets harder to close.
For now, Altman has put a number on the problem. Getting from $70 billion to about $20 billion will require cooperation from chipmakers, utilities and the broader supply chain.
