Four published figures put the cost of one gigawatt of AI capacity anywhere from about $6.25 billion to $70 billion, by our arithmetic. The highest is 11.2 times the lowest. They are not estimates of the same thing: each counts a different scope, from a data-centre campus to five years of building and running the whole system.
Sam Altman’s $70 billion covers building and running a gigawatt for five years. AM Intelligence’s $6 billion for about 100 MW, which includes Nvidia Rubin GPUs, works out to about $60 billion per gigawatt. Cushman & Wakefield’s facility cost, cited by Runware, is $17.6 billion per gigawatt without chips. OpenAI’s Project Camellia works out to about $6.25 billion per gigawatt.
Business Engineer Course · The FDE Playbook
An evergreen Business Engineer course on forward-deployed engineering, how AI gets put to work inside organisations. It teaches the general idea only and does not discuss data-centre costs.
The key insight: Four published figures imply between about $6.25 billion and $70 billion per gigawatt of AI capacity by our arithmetic, 11.2 times apart, and each counts a different scope: a campus price over contracted power, a facility without chips, a first phase with GPUs, and five years of building and running.
Chip financing, a fourth scope (30 Oct). The FT, as summarised on X by Wall St Engine, reports that Anthropic’s $63B Project North Star financing supports the leasing of 1.5 GW of Google’s AI chips, with more than $42B of Broadcom-guaranteed senior loans and $18B of junior debt backed by Anthropic’s lease payments. By our arithmetic, $63B divided by 1.5 GW is about $42B per GW. That is a chip-financing figure: it is not a building cost like the $17.6B-per-GW facility figure, and not an all-in build-and-run figure like Altman’s ~$70B, so the scopes are not added together. Details: Blackstone expects a BB rating for Anthropic after its IPO.
Altman gave the highest figure on the Cheeky Pint podcast. He said: “People give different numbers for the cost per gigawatt measured in the five-year total cost of ownership right now. But let’s say, let’s pick $70 billion as a number.” Asked whether that was the cost to install, he said: “And run it for five years.” We covered his remarks in our Altman piece.
AM Intelligence said on 5 October that it plans more capacity “in addition to the $6 billion already committed for the initial 100 MW”. Its release puts that first phase at “approximately 29,000 Rubin GPUs and close to 100 MW of capacity”. By our arithmetic, $6 billion over 0.1 gigawatts is about $60 billion per gigawatt, a figure that includes the GPUs. See our AM Intelligence piece.
Runware’s 30 October release cites a cost guide that “puts average upfront greenfield development costs for modern facilities in the United States and Canada at $17.6 million per megawatt, excluding chips and GPUs”. That is $17.6 billion per gigawatt, by our arithmetic, for the building alone. We covered it in our Runware piece.
The lowest figure comes from OpenAI’s Georgia campus. The Effingham Herald reported that “OpenAI plans to build a $20 billion, 1,400-acre data center campus in Effingham County”, and OpenAI says it is “contracting with Georgia Power Company for 3.2 gigawatts of power”. By our arithmetic, that is about $6.25 billion per gigawatt; the dollar figure is the newspaper’s and the gigawatts are OpenAI’s.
Dollars per gigawatt implied by four published figures, each with a different scope: Altman’s illustrative five-year build-and-run cost ($70B), AM Intelligence’s $6B for about 100 MW including GPUs ($60B), Cushman & Wakefield’s facility cost excluding chips cited by Runware ($17.6B), and OpenAI’s Camellia campus ($6.25B). Per-GW values are our arithmetic; do not add or average them.
Why the Gap Is Scope
The figures differ in what they include. Cushman & Wakefield’s facility cost excludes chips and GPUs. The Camellia figure divides a campus price reported by a newspaper by the power OpenAI has contracted; neither source says whether chips are included. AM Intelligence’s commitment includes the GPUs for its first phase. Altman’s figure adds five years of running costs to the build.
Altman also named a target. He said: “I think if that number were 20 instead of 70, we should be much happier.” By our arithmetic, $20 billion would be about 0.29 times his $70 billion figure.
Two other large projects carry no comparable cost. Chevron’s Project Kilby, a plant of about 2.67 gigawatts for a Microsoft data centre, discloses no construction cost; see our Kilby piece. For the Nvidia-backed Ohio site, an X post’s figure of about $50 billion per gigawatt is the poster’s estimate, not a disclosed cost; see our Ohio piece.
The Structural Read
As we read the sources, the figures cannot be added or averaged, because each one stops at a different point in the system.
As we read Runware’s release, the $17.6 million per megawatt it cites from Cushman & Wakefield excludes chips and GPUs, so it measures the building, not the compute inside it.
As we read Altman’s remarks, his $70 billion is a figure he chose to illustrate a five-year total cost of ownership, and he said that at $20 billion “we should be much happier”.
“People give different numbers for the cost per gigawatt measured in the five-year total cost of ownership right now.”
Three Implications
FOR READERS OF CAPEX NUMBERS A per-gigawatt figure needs its scope: Cushman & Wakefield’s North American facility average excludes chips, while AM Intelligence’s first-phase commitment includes GPUs.
FOR ALTMAN’S TARGET Altman said that if the number were 20 instead of 70, “we should be much happier”; by our arithmetic, 20 is 0.29 times 70.
FOR CAPACITY BUYERS Runware says its pods cost $1 million per usable megawatt per year, a rate that excludes compute hardware and electricity.
The framework puts it this way: “The structural bottleneck isnโt computing, itโs electrons.”
The framework is the owner’s general argument about AI data centres, not evidence about these four figures; it does not compare costs per gigawatt.
What Is Not Established
None of the four figures is an audited cost. Altman called his figure a number he would “pick”, and AM Intelligence’s figure is a commitment for a first phase. The $17.6 billion is an industry average for North America cited by Runware, and the Camellia dollar figure is a newspaper’s. The per-gigawatt values are our arithmetic, and because the scopes differ they should not be added or averaged.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Four published figures imply between about $6.25 billion and $70 billion per gigawatt of AI capacity, 11.2 times apart, by our arithmetic. The spread reflects scope: a facility cost without chips is $17.6 billion, a first-phase commitment that includes GPUs is about $60 billion, and Altman’s $70 billion counts five years of building and running.
A note on sourcing. Figures come from Sam Altman on the Cheeky Pint podcast, AM Intelligence’s 5 October 2026 release, Runware’s 30 October 2026 release citing Cushman & Wakefield, OpenAI’s Project Camellia post and the Effingham Herald, as covered in our earlier pieces. Per-gigawatt values and ratios are our own arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
