Two independent estimates โ a published stylized model from Epoch AI and an operator range from Lambda’s CTO โ converge on facility costs and diverge sharply on servers, and the divergence is the more informative result.
On 14 May 2026, Epoch AI published a cost breakdown by Amelia Michael and Ben Cottier modelling the up-front capital expenditure for a one-gigawatt AI data centre. The paper is explicit on its own status: “This is a stylized model, not an estimate for any specific facility.” Within those stated limits, the model arrives at $37.883B in total CapEx โ $21.188B for servers, $11.433B for the facility, and $4.925B for networking. Epoch’s gigawatt is defined as nameplate capacity for the IT equipment, the stylized system is NVIDIA GB200 NVL72, IT hardware carries a five-year life, and the facility a fourteen-year life. On an annualised basis the same model puts total cost of ownership at $8.5B a year, of which servers account for $5.021B โ the 60% that gives the analysis its headline, Servers account for 60% of the total cost of ownership of a one-gigawatt AI data center.
Separately, on 27 September 2026, Stephen Balaban โ chief technology officer of Lambda โ gave per-gigawatt figures on the MAD Podcast. His figures are an operator estimate spoken in ranges on a podcast, not a filing and not audited. He put a power plant at $2โ3B per gigawatt, the data-centre facility at $10โ15B per gigawatt, and servers at $35โ45B per gigawatt โ the largest slice by a wide margin in his account as well.
These two accounts come from different methodologies, different vantage points, and different moments. They are not fused or averaged here, and neither is presented as correcting the other. What makes them worth placing side by side is what they agree on and what they do not.
The key insight: When two independent estimates converge on one cost line and diverge sharply on another, the divergence is informative about the line itself โ not about the quality of either source. Buildings price in known ways. Servers, apparently, do not.
On the facility the stylized model lands inside the operator’s range. On servers it sits below even the low end.
The Structural Read
Start with the agreement, because that is what makes the disagreement legible. On facility costs, the two accounts land almost on top of each other: Balaban’s operator range is $10โ15B per gigawatt; Epoch’s stylized model produces $11.433B, sitting comfortably inside that band. Buildings are a mature cost category โ steel, concrete, switchgear, and cooling all price in quotable, observable ways. One estimate landing inside the other’s stated range, from entirely different methods, is roughly what you would expect.
Servers are a different story. Epoch’s stylized model puts them at $21.188B. Balaban’s operator range begins at $35B and extends to $45B. Epoch’s figure falls below even the low end of the operator range by a factor of roughly 1.7 to 2.1. The gap is real, and the structural property worth naming is this: a total is only as defined as its largest term. Servers are the largest term in both accounts. If the largest term can be $21B or $45B, every per-gigawatt total in circulation is a range wearing the clothes of a figure.
The most likely explanation for the gap is definitional โ but candidate explanations are not findings, and no one has reconciled these two figures publicly. Epoch is explicit that its gigawatt means IT equipment nameplate capacity. An operator quoting dollars per gigawatt of facility draw is counting a different quantity of computer. Epoch also fixes one specific stylized system, the GB200 NVL72; a spoken range from an operator can span configurations and hardware vintages. These are plausible reasons the two baskets are not the same basket. They are not confirmed reasons, and a definitional gap does not shrink as data improves โ it closes only when people agree what they are counting.
Map of AI โ Infrastructure Layer
The Arithmetic Problem at the Base of the Stack
In the Map of AI framework, the infrastructure layer sits beneath every capability built on top of it. When the cost of that layer cannot be stated as a single number โ when the largest line item can vary by a factor of two depending on what is being counted โ every business model built on infrastructure assumptions inherits that uncertainty. The agreement on facility costs is reassuring. The disagreement on server costs is the number that matters most, which makes it the number least resolved.
There is also a scope difference worth naming explicitly, because it is easy to misread as a contradiction. Balaban prices a power plant at $2โ3B per gigawatt. Epoch’s capital lines are servers, facility, and network; energy appears in Epoch’s model not as capital but as an operating cost of approximately $594M per year, inside a total annual OpEx of $907M against $37.883B of up-front capital. One account builds generation; the other buys electricity. That is a difference in how the business is structured, not a discrepancy in the physics of the site. It is worth keeping because the choice of whether power sits in CapEx or OpEx moves billions between columns without changing a single physical fact โ and it explains why the capital question dominates these conversations: in Epoch’s model, the entire annual operating bill is under a billion against nearly $38B of up-front spend.
Epoch AI โ Amelia Michael & Ben Cottier, May 14 2026
“This is a stylized model, not an estimate for any specific facility.”
Three Implications
IMPLICATION 1 โ FACILITY COSTS ARE THE SETTLED TERM
Epoch’s stylized $11.433B for the physical facility falls inside the $10โ15B range the operator gave. That containment is signal. For any analysis that depends on construction cost โ permitting timelines, financing structures, real-estate strategy โ the facility number appears to be the most stable input available. It is the one term in this cost equation that does not appear to be in dispute.
IMPLICATION 2 โ EVERY PER-GIGAWATT TOTAL IS A RANGE IN DISGUISE
Servers are the largest cost category in both accounts and the most contested. A factor-of-two spread on the biggest line item means any single per-gigawatt total โ whether from a stylized model or an operator โ carries an implicit range that dwarfs rounding errors. Treating published figures as point estimates understates the uncertainty baked into the underlying arithmetic.
IMPLICATION 3 โ THE UNIT DEFINITION IS THE PRIOR QUESTION
Before any two gigawatt-scale cost figures can be compared, the definition of a gigawatt must be shared. IT equipment nameplate capacity and facility power draw are different quantities. Until that definitional question is settled publicly โ and no such settlement has been published โ the gap between estimates will persist regardless of how much primary data accumulates. The measurement problem precedes the data problem.
Business Engineer Framework
The Map of AI โ Infrastructure Layer
The Map of AI traces where value is created and captured across nine layers of the AI stack. The infrastructure layer โ where data centres, compute, and power sit โ is the foundation every upstream capability depends on. When the cost of that layer cannot be stated as a resolved number, every business model built above it inherits the uncertainty. Understanding which layer you occupy, and what the coststructure of the layer below you actually is, is the prior strategic question.
A stylized model from Epoch AI and an operator estimate from Lambda’s CTO land compatibly on the physical facility โ Epoch’s $11.433B falls inside the operator’s $10โ15B range โ and that compatibility is worth something. But they disagree by a factor of 1.7 to 2.1 on servers, which is the largest cost in both accounts, and no one has publicly reconciled why. Until the unit definition is shared and the server figure is resolved, every per-gigawatt total in circulation is a range with a false precision problem, and the infrastructure layer of the AI stack remains the most consequential number that the industry cannot yet pin down.
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
Two different kinds of number appear above and they are never fused or averaged.The Epoch AI figures come from what its authors describe as a stylized model, not an estimate for any specific facility, published 14 May 2026 by Amelia Michael and Ben Cottier; its gigawatt means nameplate capacity for the IT equipment, and it assumes NVIDIA GB200 NVL72 with a five-year IT hardware life and a fourteen-year facility life. Those assumptions are doing real work. The Lambda figures are an operator estimate, spoken in ranges by Stephen Balaban on a podcast on 27 September 2026 — not a filing, not audited and not a published model.Nothing above claims either estimate is wrong, more accurate, or corrected by the other, and nothing above reconciles them. The definitional explanations offered for the gap are candidates only; no public reconciliation of the two figures has been published. The difference between costing a power plant and treating energy as an operating expense is a difference of scope, not a contradiction. No reconciliation, no costs for any named site, no Lambda build costs, no utilisation or financing terms and no industry-wide buildout total appear above — a limit of this reporting rather than evidence that no such figures exist. Nothing above predicts costs, capacity, prices, or the conduct of Nvidia, Lambda, Epoch or any other party.
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.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
