What a gigawatt of AI data center actually costs
Four respected research houses priced the same one-gigawatt build between $35 billion and $60 billion. An average is tempting. It would remove exactly the information that the spread contains.
Four surveyors, one bridge
Imagine that you ask four licensed surveyors for the length of one bridge. Their answers go from 300 to 600 meters. You would not average them and continue. You would suspect, correctly, that they measured different things: one the main span, one the full deck, one the deck with the approach roads. The spread is not error. It shows the different definitions, and it is the most informative thing that you have.
The cost of an AI gigawatt has the same problem. The published estimates are:
- Bernstein: roughly $35 billion.
- Morgan Stanley: $35, $39 or $49 billion, depending on which NVIDIA generation fills the racks.
- Barclays: $50 to $60 billion per gigawatt, from converted industry plans. NVIDIA itself likes to quote this range.
- JLL: $45 to $55 billion, for whole campuses.
- Epoch AI: about $38 billion, from a bottom-up model.
That is twice the money for the same nominal object, from people who are all good at arithmetic.
So the useful task is not to decide who is right. It is to find the questions that each number answers without saying so. There are three. When you can see them, the whole spread makes sense. You can also apply the three questions to any per-gigawatt figure that anyone quotes to you. That is worth more than any single answer.
The question of the meter
Start with the assumption that nobody states: what, physically, is the gigawatt?
There are three defensible readings:
- A gigawatt of IT load. Every watt arrives at a rack.
- A gigawatt at the utility meter. Cooling and conversion take their share first. An AI-optimized facility runs a power usage effectiveness (PUE, total facility power divided by IT power) around 1.14. So only about 877 MW reaches compute.
- A gigawatt of grid interconnection, sized to carry redundant feeds. Here both the PUE and the redundancy factor reduce the power before a single chip receives any.
Calculate the cost of each reading with a GB200 NVL72 rack, which reportedly draws 132 kW in deployment:
| “One gigawatt” means | IT load | Racks | Silicon at $3.4M per rack |
|---|---|---|---|
| IT load | 1,000 MW | 7,575 | $25.8B |
| Meter reading, PUE 1.14 | 877 MW | 6,645 | $22.6B |
| 2N interconnect | 439 MW | 3,322 | $11.3B |
The same phrase buys 2.3 times more compute at one end of the table than at the other. That is a difference of about $14.5 billion in silicon alone. For comparison, Morgan Stanley’s full increase from GB200 to Vera Rubin is $14 billion, and analysts write notes about that increase. The definition is worth as much as the technology generation. Unlike the technology, it never appears in the headline.
The question of the generation
The second hidden question is the visible one, so it needs less work: which chips? Morgan Stanley’s own figures are $35 billion per gigawatt for GB200, $39 for GB300 and $49 for Vera Rubin. The cause is the silicon content per rack. Bernstein’s teardown of a Vera Rubin rack finds about $3.96 million of GPUs (72 packages near $55,000 each) plus roughly $3.2 million of memory and storage. HBM4 (the fourth generation of high-bandwidth memory) alone costs more than a million dollars per rack.
Notice where the increase is: in memory, more than in logic. This detail matters to anyone who reads these numbers for signals about supply chains, because the memory market has three suppliers and is sold out.
The question of the spare
The third question is inside the electrical plant: how much redundancy did the estimate buy?
The topologies are standard. Tier III means N+1, one spare unit per group. Tier IV means 2N, two fully independent power paths. Either path can carry the site alone.
A common misreading is that 2N doubles the cost of the data center. It does not. It doubles only the electrical plant: generators, transformers, UPS (uninterruptible power supply) and switchgear. From Bernstein’s category shares, these items are a bit under a third of the build. In the model below, full duplication adds about a fifth to the total on top of an N+1 plant. On top of a plant with no spare, it adds about a quarter. Neither is a factor of two, because the design does not duplicate the building or the chillers.
The subtler cost of 2N is not money at all. It doubles the grid connection that you must secure. Grid connections, not dollars, are the queue everyone is standing in.
When all three questions are explicit, you can calculate any of the published numbers yourself. The model below starts with middle values: GB200, a meter-reading gigawatt, N+1 and PUE 1.14.
Build your own gigawatt
Every control below is one of the assumptions the published estimates make silently. Change them and watch a single gigawatt move across the whole range the analysts report.
Basis: The gigawatt is measured at the meter, so cooling and losses come out of it first. IT load is the gigawatt divided by PUE.
N+1: Uptime Tier III. One spare unit per group. The 1.2x factor assumes roughly five blocks plus one spare: an assumption, not a published figure, since the real factor depends on how the plant is divided.
GB200 NVL72: Bernstein splits a ~$5.9M NVL72 rack into ~$3.4M of compute hardware and ~$2.5M of physical infrastructure. Only the compute half is used here; infrastructure is modeled separately so it is not counted twice. NVIDIA specifies 120 kW; deployed racks have been reported drawing 130-132 kW.
| Category | Capex | Share |
|---|---|---|
| Compute racks | $22.6B | 52% |
| Scale-out network & optics | $5.6B | 13% |
| Power & electrical | $11.2B | 26% |
| Cooling | $1.3B | 3% |
| Shell, land & site works | $2.3B | 5% |
- IT load
- 877 MW
- Facility draw
- 1,000 MW
- Interconnect
- 1,200 MW
- Racks
- 6,645
- GPUs
- 478,440
- All-in per rack
- $6.5M
Against what the analysts published
The last column runs the model backwards: holding your facility and power assumptions fixed, it asks what each published total implies a rack costs. When that number drifts far from the rack prices actually reported, the disagreement is about scope, not about pricing.
- Your model$43.1B$3.4M/rack
- BernsteinGB200 NVL72$35.0B$2.3M/rack
- Morgan StanleyGB200$35.0B$2.3M/rack
- Morgan StanleyGB300$39.0B$2.9M/rack
- Morgan StanleyVera Rubin$49.0B$4.2M/rack
- Epoch AIGB200 NVL72$38.0B$2.7M/rack
- Barclaysnot stated$50–60B$5.0M/rack
- NVIDIAforward generations$50–60B$5.0M/rack
- JLLnot stated$45–55B$4.3M/rack
Sources: Bernstein, via Investing.com, "How much does a GW of data center capacity actually cost"; Morgan Stanley per-generation GW deployment costs, via Orennia; Epoch AI, total cost of ownership of a one-gigawatt AI data center; Barclays AI capacity tracker, via Yahoo Finance; JLL gigawatt-scale campus build costs, via Core Insights Review.
Where the numbers refuse to meet
Use the platform selector, and you will find the place where the published record genuinely breaks. That place is more interesting than any place where the record agrees.
Set the model to Vera Rubin at Bernstein’s reported $9.1 million rack and a meter-reading gigawatt. The total is near $88 billion. Secondary reports from October 2026 say that Bernstein lowered its rack estimate to about $7.5 million. At that price, the total is still near $75 billion. Morgan Stanley’s published figure for the same generation is $49 billion. Now run the model backward and hold the facility assumptions fixed. Then $49 billion implies a rack near $4.1 million, about half of either Bernstein figure. Careful people published both numbers, but they cannot both describe the same gigawatt.
At least one of three things must be true, and the sources disclose none of them:
- The per-gigawatt figure may assume far fewer racks, through a different power basis or partial deployment.
- The teardown price may be a bill-of-materials view that no hyperscaler actually pays. Other estimates for the same rack are $7.0 to $8.8 million, and NVIDIA never confirmed list pricing.
- The scopes may differ completely. For example, JLL prices a campus, while Epoch prices a building.
Bank of America appears in every roundup of these numbers, but it is absent from ours. The reason is worth stating: we could not find a sourceable per-gigawatt figure from them. We found only aggregate capital expenditure (capex) forecasts, and we prefer to show a gap than to fill one.
We do not reconcile the contradiction. An honest reconciliation would need assumptions that none of these houses published. The practical conclusion is clearer anyway. A per-gigawatt number without its power basis, its GPU generation and its scope is not a number yet. It is a headline that still needs its footnotes.
Where they quietly agree
The disagreements get the attention, but the agreements contain the information, because different methods produce them.
The headline number itself agrees better than the noise suggests. For a GB200-class gigawatt, there are four paths:
- Bernstein’s bottom-up rack teardown gives $35 billion.
- Morgan Stanley’s deployment accounting gives $35 billion.
- Epoch’s academic cost model, built from entirely different inputs, gives about $38 billion.
- You can calculate a fourth path yourself with only the numbers on this page. 877 MW of IT load at 132 kW per rack is 6,645 racks. At $3.4 million of silicon each, that is $22.6 billion. Add a facility at $13 million per megawatt plus its N+1 electrical spare, then add networking. The result is near $43 billion.
These are four methods from four owners, and their results are close together. When methods this independent agree within about 25 percent, you can depend on the number. When you need to update it in two years, any one of the four recipes will do.
Networking is a smaller line inside that total. Bernstein works bottom-up from component supply chains and puts scale-out networking at roughly 13 percent of total capex. For a gigawatt, that is on the order of $5 billion in switching, cabling and optics. This line item grows as rack power increases, because shorter electrical reach moves more of the network into photonics.
The other quiet agreement is about time, not money. Epoch’s model puts annual cash operating cost (opex) near $0.9 billion against about $38 billion of capital. Think about that ratio: the machine costs forty times more to build than to run for a year. In the annualized picture, the assumption that changes the answer most is the equipment’s assumed life, not the electricity price. Three years instead of five changes the annual cost from $8.5 to $12 billion. Depreciation policy has a much larger effect than power price, and it deserves its own essay.
The constraint no estimate prices
Every number above prices a building that you have permission to energize. In 2026, that permission is the scarce input.
Grid interconnection takes four to seven years in most US markets. ERCOT (the grid operator for most of Texas) alone had a large-load queue near 410 gigawatts in April 2026. Close to 90 percent of that demand is data centers.
The alternative is to generate your own power, but that has its own queue. Reports say that turbine makers allocated their heavy-duty turbine slots through 2029. Lead times for the transformers that must go with them went above 140 weeks. The response was to order both anyway. US developers announced roughly 101 gigawatts of behind-the-meter generation.
This changes the question that this article started with. You can buy a gigawatt of GPUs and get delivery within a year. A gigawatt of permission takes half a decade. The interesting price is moving from what a gigawatt costs toward when a gigawatt is possible. None of the surveys show that number.
What to take away
- Never average per-gigawatt estimates. Ask the three questions first: what does the gigawatt measure, which generation fills it, and how much redundancy does the estimate include? The first question alone is worth 2.3x in compute.
- The published record contains a genuine contradiction. $49 billion per Vera Rubin gigawatt and Vera Rubin racks at $7.5 to $9.1 million cannot both describe one site. Be suspicious of any reconciliation that you get. Its inputs are not public.
- Trust the agreements. Four independent methods put the total within about a quarter of each other. Networking is about 13 percent of the build in Bernstein’s count, and that share grows as networks become optical.
- Capex is roughly forty times annual cash opex. The dominant operating assumption is equipment life, not electricity.
- None of the surveys put a price on the wait for power. Increasingly, that wait is the real cost of a gigawatt.