Start from the server, not the chip: each layer between the GPU and the meter multiplies the load.
Multiply the GPUs by the server power per GPU, add the network and storage, then apply the peak PUE. On illustrative figures, 100,000 GPUs in 8-GPU servers drawing 10.2 kW each need 142.8 MW of IT load and 178.5 MW at the grid, about 1.8 kW per GPU. Running at 85 per cent of capacity, the cluster uses 1.22 TWh a year, about $85.6 million of electricity at $70 per MWh.
Worked in full in The Digital Infrastructure Investor by Julian R. Sterling, with every figure reproduced in a free workbook.See the book on Amazon →
The figure most often quoted for an AI accelerator is its thermal design power, around 0.7 kW for a high-end training part. Multiply 100,000 of them by 0.7 kW and you get 70.0 MW, which is the number that appears in press coverage and early-stage site screens. It is 39.2 per cent of what the building actually draws. A server also carries CPUs, memory, network cards, fans or pumps and power supply losses; the cluster carries a switching fabric and storage; the building carries cooling and conversion losses. Each layer multiplies the one below it, which is why the investor's unit is kW per GPU at the meter, not the chip's rating.
| Input | Value |
|---|---|
| GPUs | 100,000 |
| GPUs per server | 8 |
| Server power at full load, kW | 10.2 |
| Network, storage and management, share of server load | 12.0% |
| Peak PUE, liquid-cooled halls | 1.25 |
| Annualised PUE | 1.15 |
| Average draw as a share of IT capacity | 85.0% |
| Power price, $ per MWh | 70 |
Servers = 100,000 / 8 = 12,500
Server load = 12,500 × 10.2 kW = 127.5 MW
IT load = 127.5 × (1 + 12.0%) = 127.5 + 15.3 = 142.8 MW, or 1.428 kW per GPU
Grid capacity = 142.8 × 1.25 = 178.5 MW, or 1.785 kW per GPU
Annual energy = 142.8 × 85.0% × 8,760 h × 1.15 = 1,222,782 MWh, or 1.22 TWh
Excel: =GPUs/PerServer*ServerkW/1000*(1+Overhead)*PeakPUE for the grid MW
Turned round, 1 MW of IT capacity houses 700 GPUs on these assumptions, and 1 MW of grid capacity 560. At four servers a rack the cluster fills 3,125 racks at 40.8 kW each before networking, a density that rules out most air-cooled halls built for the previous generation, as discussed in how many racks 1 MW of data centre capacity supports.
Note the two PUEs. Capacity is sized on the hottest hour, 1.25; energy is billed on the year, 1.15. The average draw is 139.6 MW against 178.5 MW of connection, and the gap is not waste: it is the capacity a hot afternoon at full training load requires.
| Line | $ million a year | $ per GPU a year | $ per GPU-hour |
|---|---|---|---|
| Electricity, at $70 per MWh | 85.6 | 856 | 0.098 |
| Rent on 142.8 MW at $138 per kW-month | 236.5 | 2,365 | 0.270 |
The rent line uses the $138 per kW-month at which the book's Halvard DC1 case is leased, applied to IT capacity. On these figures the building and its power together cost the user about 36.8 cents per GPU-hour, likely a modest share of what the servers themselves cost to own once their price is spread over a short useful life. That asymmetry is why a tenant will pay for speed to power over a cheaper site, and why the developer's scarce input is the grid connection rather than the shell.
| Server kW (8 GPUs) | IT MW | PUE 1.15 | PUE 1.25 | PUE 1.40 | kW per GPU at 1.25 |
|---|---|---|---|---|---|
| 6.5 | 91.0 | 104.7 | 113.8 | 127.4 | 1.14 |
| 10.2 | 142.8 | 164.2 | 178.5 | 199.9 | 1.78 |
| 14.3 | 200.2 | 230.2 | 250.3 | 280.3 | 2.50 |
Server power drives the answer far more than cooling efficiency. Moving from a 6.5 kW to a 14.3 kW server more than doubles the connection for the same GPU count; moving from a 1.40 to a 1.15 PUE on the 10.2 kW server saves 35.7 MW. For an investor this is the central risk of an AI-dedicated building: a hall designed around one generation's kW per GPU can be stranded, or need a retrofit, when the next generation arrives. The book's Halvard DC1 case prices exactly that difference between a retrofitted and a stranded hall.
The power, chips and lease economics of an AI-era data centre are worked in the book, and the free workbooks for its cases reproduce them by live formula. For the revenue side per MW, see how much revenue a data centre makes per MW.
Budget about 1.8 kW of grid capacity per GPU on today's illustrative figures: 100,000 GPUs need 178.5 MW, use 1.22 TWh a year and spend $85.6 million on electricity at $70 per MWh. Start from server power, not chip power, and size the connection on the peak PUE.
About 700 per MW of IT capacity on illustrative figures: 8-GPU servers at 10.2 kW plus 12 per cent for networking and storage give 1.428 kW per GPU. Per MW of grid capacity, after a 1.25 peak PUE, the figure is 560. Higher-power servers cut both numbers sharply.
On illustrative assumptions, 142.8 MW of IT load running at 85 per cent on average with an annual PUE of 1.15 uses 1,222,782 MWh, or 1.22 TWh, a year. At $70 per MWh that is $85.6 million, or about $856 per GPU and 9.8 cents per GPU-hour.
Because the chip is one layer of several. A 0.7 kW accelerator times 100,000 is 70.0 MW, but servers, networking, storage and cooling take the illustrative cluster to 178.5 MW at the grid, 2.55 times the chip figure. Sizing a site on chip power alone understates the connection it needs.
This article is one calculation from The Digital Infrastructure Investor. The book takes the same case from first principles to the decision, chapter by chapter, and every figure it prints is a live formula in the free companion workbooks.
Get the book on Amazon →Free companion files
Also on Amazon UK · Amazon Germany · Amazon France · Amazon Canada
Reading guide: infrastructure, data centres and energy → · All 1682 articles →
If this book helped, or didn’t, a few lines on Amazon are worth more than they look: they are what the next reader goes on. Write a review. The workbook stays free either way.