As we canall see, AI is spreading very rapidly, but behind AI agents or “starter packs,”there is a frantic race to deploy infrastructure in order to keep up with themultiplication of use cases.
Cinema and audiovisual production are no exception to this trend, withincreasing use of AI in production workflows.
The deployment of AI involves a growing number of data centers, particularly“colocation DCs” (from 500 to 10,000 m²) and “hyperscale DCs” (over 10,000 m²),which must adapt to the physical specifics of GPU servers — the fundamentalhardware building block of AI.
GPU servers require large amounts of electricity ⚡ and, consequently, a lot of water 💧 for cooling. Finally, the construction of new data centers will require large quantities of concrete 🪨.
⚡ Energy
A GPU server consumes an average of 5 kW, which is 2 to 10 times more than aCPU server.
Source: ASUS Servers
This means AI (GPU) data centers require much more energy supply thanconventional (CPU) data centers.
The IEA, in its recent report “Energy and AI,” highlights that between2024 and 2030, data center electricity consumption is expected to grow by 15%per year — a rate four times higher than that of other sectors.
Source: IEA, p.63
Semi Analysis, based on the study of over 3,500 data centers in the U.S. —including both existing and under-construction colocation and hyperscale datacenters — has modeled the “Critical IT Power”.
Source: AI Datacenter Energy Dilemma – Race for AI Datacenter Space
(“Critical IT Power” refers to the usable electrical capacity at the datacenter level, available for computers, servers, and network equipment hosted inserver racks. It excludes energy used for cooling, power supply systems, andother facility-related infrastructure.)
It appears that the Critical IT Power of U.S. data centers will rise from 3.3%of U.S. electricity production in 2020 to 14.6% in 2028.
(See chart.)
While AWSspent $650 million to purchase a “nuclear-powered data center” with 960 MWcapacity (source: Data Center Dynamics), most AI data centers will bepowered primarily by fossil fuels.
Source: TIME
The consequence of this massive energy consumption is an increase in waste heat that must be dissipated.
💧 Water
Non-AI data centers cool their servers using :
- Air Cooling, up to 40 kW per rack
- Door Cooling, for 40 to 70 kW per rack
But AI datacenters, with rack power exceeding 70 kW, require heat dissipation that only directcooling — using direct water contact with GPUs — can handle.
This results in significant evaporation of potable water, which cannot then beused for other domestic, agricultural, or industrial purposes.
For instance, Microsoft's AI data center in the Netherlands was temporarily banned after it consumed 4 to 7 times more potable water than expected.
Source: Next
🪨 Concrete
The exact quantity of concrete needed to build a hyperscale AI data center (nospecific data center design in Betie-one click LCA from SNBPE) is not yetknown, but it is likely to be in the thousands of tons.
Based on FDES data from INIES (Floating screeds made of concrete andcement-based mortar [7–10 cm thick] – DEFAULT ENVIRONMENTAL DATA v.1.2),one m² of screed has an impact of 74.41 kg CO₂-eq.
Source: INIES
For a hyperscale data center, the carbon footprint could reach several thousandtons of CO₂-equivalent.
So now what ? 🤷♀️
From an economic standpoint, in light of the necessary investment costs andphysical limits (water, energy, resources), AI use will increasingly bemonetized (no more free starter packs).
From an environmental standpoint, in its current form, the digital sector’sdecarbonization trajectory is at risk.
Since AI is expected to continue expanding, its use will need to be optimized —especially as its rising costs will strongly incentivize efficiency.
It also seems clear that its applications, particularly in cinema andaudiovisual industries, must align with a strategy focused on reducingenvironmental impact.