For thirty years, data centre electrical design has rested on one assumption. The grid is big, the data centre is small, and the job is to protect the servers from whatever the grid throws at them.
That assumption no longer holds. AI has changed the shape of data centre demand and campuses now connect at a scale where their behaviour moves the grid. Grid operators across Europe and North America are responding with new connection rules, and the protection obligation now runs both ways: data centres must also protect the grid they depend on.
On 28th September, GreenScale Data Centres published Grid & AI Load Compatibility Requirements, an open and detailed engineering guide to what this means for design, testing and operations. It is well worth every operator's time. But it is, by design, an engineering document. The commercial questions it raises (who carries the risk, who pays, what changes in contracts and connection strategy) are where boards and energy teams now need answers.
What has actually changed
Two things and they compound.
AI loads pulse. A traditional cloud hall runs thousands of servers doing unrelated jobs. Their peaks and troughs cancel out, so the grid sees a smooth, predictable line. AI training is different. Thousands of GPUs compute in lockstep, so a whole site's demand can surge and drop together, many megawatts at a time, on a repeating rhythm.
Think of soldiers marching in step across a bridge. The weight is no greater than a crowd walking normally; the rhythm is the problem. Repeating power swings can set up oscillations in the grid and in on-site generators, and in the worst cases damage turbines and engines.
Data centres flinch together. When the grid suffers a brief fault, most data centres do exactly what they were designed to do: step away from the grid and switch to backup power. One site doing this is invisible. Dozens doing it at once means the grid suddenly loses a huge block of demand, then has to absorb it again when they reconnect.
On 22nd July this year, a routine fault on a 230 kV line in Northern Virginia saw around 3.8 GW of data centre load switch to backup generation without warning. PJM called it the largest such event in its history. Similar events in July 2024 and February 2025 were about 1.5 GW each. The grid held, with operators moving quickly to rebalance it.
The rules are arriving, but not evenly
New connection requirements for large loads are published or in development in Ireland, Northern Ireland, Finland, Belgium, Denmark, France, Germany and Norway, with ENTSO-E pushing for coordination.
Great Britain is heading the same way. NESO has floated large demand technical requirements at its Grid Code Development Forum, modelled on the rules for generators, including restoring 90% of demand within half a second of voltage recovering.
In North America, ERCOT and Alberta's AESO are moving too, and FERC has ordered NERC to write enforceable standards for data centre loads, with a first phase due by 31st December 2026.
The core asks are broadly similar:
The catch is that there is no common standard. Rules vary by country and are still changing and there is not yet an industry standard for describing AI power draw. Requirements may also be applied retrospectively, which is why GreenScale recommends designing to a common baseline now, even where the rules are not yet in force.
For operators building across several European markets, that means one design philosophy, many rulebooks and a moving target.
Four commercial questions nobody has fully answered yet
GreenScale's paper sets out the engineering. The commercial questions it raises are just as pressing and far less discussed.
1. Who carries the disconnection risk?
This is the underlying issue. The operator is accountable to the grid for a load profile it does not fully control, because tenants run the workloads. A single tenant's training run could put the whole site in breach of its connection conditions, with penalties or disconnection on the table.
It also flips a long-standing principle. The operator may need to curtail a tenant's power to protect the grid or on-site generation. Under a traditional lease, that would be an SLA breach.
This is a shared obligation, not a tenant problem: neither side can meet it alone. Expect leases to grow a new schedule: an agreed allowable power draw profile, with monitoring, notification and remedies on both sides. Energy pass-through clauses will need a second look too.
2. Who pays for the extra storage?
Batteries used to be backup. Now they are also shock absorbers, smoothing AI load so the grid and generators see something they can tolerate. That means bigger batteries, more cycling and potentially shorter life.
It also sharpens a question every battery business case must answer: you cannot sell the same charge twice. Capacity held back for smoothing, ride-through and backup is not available for trading or grid services. The strongest cases will allocate state of charge by role, model degradation honestly and still find revenue in the headroom that is left.
3. What happens to on-site power?
Microgrids and on-site generation are rightly gaining ground as a way to get sites energised sooner. But generators have traditionally been specified for occasional load transfers, not constant pulsing. GreenScale warns that cyclic AI loads can affect emissions performance, fuel efficiency, warranty and equipment life.
None of this rules out on-site power. It does mean the generator specification, warranty terms and buffering strategy need settling before contracts are signed, not discovering at commissioning.
4. How does this change the route to connection?
Grid operators increasingly expect evidence before they connect you: modelled load profiles, simulation studies and proof of ride-through performance. That is more work, earlier, across more disciplines.
Handled well, it is also an advantage. A site that can show the grid operator it will behave well is a site that is easier to connect. In a market where power availability now decides where data centres get built, that matters.
What to do now
The bigger picture
In our white paper, Powering the Next Generation of Data Centres, we argued that energy strategy is becoming infrastructure strategy. GreenScale's work shows how literally true that now is. The IT workload, the batteries, the generators and the grid connection have become one tightly coupled system and they have to be planned as one.
The operators who win the next phase will not just be the ones who secure power. They will be the ones who can prove they are good neighbours on the grid and who have the contracts and commercial structures to back that up.
At True Group, we help data centre operators across Europe join those dots: grid connection strategy, battery business cases, procurement and risk, and the regulatory detail in each market. If you are working through what these rules mean for your portfolio, we would be glad to compare notes.
FAQ
What is fault ride-through for data centres?
Fault ride-through is a requirement to stay connected and keep drawing power through a brief grid fault, rather than switching to batteries and dropping off. It already applies to power stations. Grid operators in Ireland, the Nordics, France and elsewhere are now applying it to large data centre loads, alongside rules on how quickly demand must recover afterwards. NESO has floated similar requirements for Great Britain.
Why are AI workloads a problem for the grid?
AI training synchronises thousands of processors, so a site's power demand rises and falls together in a repeating pattern. Those swings can be large and rhythmic enough to excite oscillations in the grid and in on-site generators. Traditional cloud workloads average out and present a smooth, predictable load.
Do these rules apply to existing data centres?
It depends on the market and the connection agreement. Rules are still evolving and some may be applied retrospectively, so operators should check current codes and published drafts for each site, and plan for the requirements likely to be in force by their energisation date.