Gartner now expects datacenters to pull 1,200 TWh by 2030, surpassing even Schneider Electric's most aggressive scenario. AI-optimized servers are the reason, and grid operators may simply run out of headroom to connect new capacity.
Gartner just moved the goalposts on datacenter energy consumption again, and the new numbers are uncomfortable for anyone who has watched grid interconnection queues stretch out past five years. The firm projects global datacenter electricity consumption hits 565 terawatt-hours (TWh) in 2026, with power demand climbing from 104 GW in 2025 to 132 GW this year. That is a 26 percent jump in a single year, and the curve keeps bending upward through the end of the decade.
What makes this notable is not just the size of the number but how badly it overshoots earlier predictions. Two years ago Gartner pegged AI-optimized servers at roughly 500 TWh per year by 2027. We are already past that pace in 2026. When a forecast from a major research outfit undershoots reality by a full year inside of 24 months, it tells you the underlying growth rate is steeper than the models assumed.
The forecasts keep getting beaten
It is worth laying the competing estimates side by side, because the spread tells the real story. Goldman Sachs projected datacenter energy use would roughly double by the end of the decade. Schneider Electric published four scenarios at the start of 2025 covering a range of AI buildout outcomes. Gartner's latest 2030 figure clears even Schneider's most aggressive case.
| Forecast source | Metric | Estimate |
|---|---|---|
| Gartner (2024) | AI-optimized server consumption by 2027 | ~500 TWh/yr |
| Gartner (2026) | Total datacenter consumption, 2026 | 565 TWh |
| Gartner (2026) | Total datacenter consumption, 2030 | >1,200 TWh |
| Schneider Electric (2025) | Most aggressive 2030 scenario | below Gartner's number |
| Goldman Sachs | End-of-decade growth | ~2x current |
The pattern is consistent: every conservative estimate ages badly, and the aggressive ones turn out to be the realistic ones. For capacity planners that is a brutal position, because you provision substations and transformers years ahead of load, and the load keeps arriving faster than the lead times allow.
AI silicon is eating the power budget
The driver is no mystery. Gartner estimates AI-optimized servers will account for 31 percent of all datacenter power consumption this year. By 2027 their combined draw is expected to surpass every conventional server in operation, the entire installed base running databases, web tiers, virtualization, and analytics combined.
This tracks with what anyone racking GPU nodes already knows. A dense conventional 2U server might pull 400 to 800 watts. An 8-way accelerator node lands in a completely different regime. An NVIDIA HGX H100 system draws on the order of 10 kW. The newer GB200 NVL72 rack-scale designs push past 120 kW in a single rack. You do not need many of those before a hall that was provisioned for 5 to 10 kW per rack is wildly underbuilt. Older facilities were never wired for this density, which is why so much new construction is greenfield rather than retrofit.
The cooling math compounds the problem. At those rack densities air cooling stops working and direct-to-chip liquid or immersion becomes mandatory. Liquid cooling is more efficient per watt of heat moved, but it does not reduce the watts going into the chips. It just changes how you get the heat out. The facility power overhead, the PUE, can improve, yet the IT load itself is what is exploding.
Why the grid is the actual constraint
Gartner director analyst Linglan Wang framed it directly: "AI capacity is now constrained by power availability, making datacenter power security the new battle ground for scaling and protecting margins in the global AI race."
That is the part that should concern homelab builders and hyperscalers alike, because it is the same physics at different scales. You can buy the silicon. You can buy the cooling. What you cannot conjure is a grid interconnect. In the largest US energy market, datacenter demand has already pushed prices up by around 75 percent, and operators are stuck between developers demanding capacity and residents protesting both the construction and the bills. Energy analysts do not see a clean exit.
The generation side cannot turn on a dime. A new gas peaker plant takes a couple of years. Transmission lines take the better part of a decade once you account for permitting and right-of-way fights. Meanwhile the federal response has included pumping funds into keeping coal plants online in the name of energy security, which tells you how tight the supply picture has become when the policy answer is to stop retiring old thermal generation.
What operators are actually supposed to do
Wang's prescription, that infrastructure and operations leaders should "prioritize efficiency upgrades and secure grid access" and invest in high-efficiency cooling and edge computing, is sensible but reads as advice for people who already know they are in trouble. Securing grid access is the whole game, and there is only so much access to secure.
For anyone planning capacity over the next few years, a few things follow from these numbers. Power, not floor space or capital for hardware, is the binding constraint, so site selection now means following stranded generation and favorable interconnect queues rather than network latency to population centers. Efficiency work that used to be a nice-to-have, tuning PUE, raising inlet temperatures, moving to liquid, becomes load-bearing because every watt saved on overhead is a watt you can redirect to compute you already cannot get enough power for. And the smaller operator running a serious homelab or a colo cage faces the same squeeze in miniature: the breaker on your rack is a hard ceiling, and measuring actual draw at the PDU matters more than the spec-sheet TDP that never matches reality under sustained load.
The broader pattern here is that the AI buildout has converted an industry that competed on chips and software into one that competes on electrons. Whether the 1,200 TWh figure proves high or low by 2030 almost does not matter for planning purposes. The direction is set, the grid is the bottleneck, and the forecasts that looked aggressive a year ago are the ones coming true. Gartner's full research is available through its datacenter and infrastructure coverage, and The Register's ongoing reporting on grid constraints is worth tracking for anyone whose roadmap depends on a substation that may not exist yet.
Comments
Please log in or register to join the discussion