Three stories I have been thinking about this week.
Lithuania wants parts of its electricity grid to keep operating even when the main transmission network is damaged. Britain spent £3.1 billion balancing its electricity system in a year. And Norway has temporarily stopped reserving new grid capacity for data centres while it develops technical requirements for connecting them.
Together they touch three requirements for the grid that are also as relevant to any distributed system: resilience, efficiency, and scalability..
Reading them, I kept coming back to a familiar question from software engineering:
How much of a system’s reliability comes from the way its components are connected and operated?
Main signal: can the grid have availability zones?
On 10 November, Lithuania’s transmission system operator Litgrid presented its proposed energy islands model.
The idea is to divide the country into seven electricity-system regions that could temporarily operate independently if the main 330 kV transmission network became unavailable.
To make that possible, Litgrid proposes ten new flexible power plants with a combined capacity of 800 MW, upgrades to existing generation, and integration of battery storage. The initial phase could support limited isolated operation of four regions in 2028; the full seven-region design is targeted for 2033.
Litgrid explicitly points to Ukraine’s experience of attacks on energy infrastructure as one of the reasons for the proposal.
Reading about it, I couldn’t help thinking about distributed systems.
Watertight compartments, or availability zones?
It reminded me of a ship protection technology — watertight compartments.
The idea was to divide a ship into isolated sections so that if one section was breached, the barriers would prevent water from spreading. I learned about this from the story of the Titanic. We all know that it did not save the ship from disaster. The effectiveness of watertight compartments depends on how the damage spreads and how much of the ship can still function.
In cloud infrastructure, there is another familiar idea: availability zones.
You can deploy a service across independent failure domains, usually in different geographical locations, so that healthy infrastructure can continue serving requests when one or more zones go down. Of course, the application must be designed for failover, and the surviving zones need enough capacity.
The ideas are similar, but there is an interesting difference: in the first case, individual sections may survive even when the whole system fails; in the second, the goal is for the whole system to remain available when individual zones fail.
I can imagine the Titanic having some perfectly dry, intact compartments even after sinking, possibly for quite a long time. Not particularly useful for the ship as a whole, though.
And now I wonder: which of these systems is Lithuania trying to build?
Under normal conditions, the seven regions would work together as one interconnected electricity system. During a major disruption, they should be able to separate and continue supplying electricity independently.
That could significantly improve fault containment and regional resilience. But does it also give us the equivalent of cloud failover? If one region loses its ability to supply consumers, could the other six take over part of its load?
It depends on what survives the failure. Power needs a usable transmission path, and neighbouring regions need enough generation and reserves to help. Unlike a load balancer redirecting API requests, the grid cannot simply move demand wherever spare capacity exists.
Each isolated region must also balance its own generation and demand, maintain frequency and voltage, respect equipment limits, and keep protection working under the new topology. Healthy power lines alone won’t help much without enough generation and control capability.
So I see two different goals: keep each region operating when the wider grid is damaged, and keep the wider system supplying electricity when one region fails.
Lithuania is proposing a way to achieve the first. How far it could support the second is what I would like to understand.
The engineering problem behind the proposal
To understand what it takes for a region to survive on its own, suppose a major transmission connection is suddenly lost and that region becomes electrically separated from the rest of Lithuania.
Several things have to happen, often almost at once.
First, detect and establish the new operating state. Which breakers opened? Which sources and loads are still connected? What has happened to frequency, voltage and power flows?
Second, stabilize the island. Some resources must establish or support voltage and frequency. Batteries and fast-responding generation may help, but their capabilities, operating modes and energy limits matter. If demand exceeds supply, some loads may need to be disconnected.
Third, adapt protection and control. Short-circuit levels and fault-current behaviour can be different in islanded operation. Protection settings and coordination that worked in the interconnected grid may not work in precisely the same way after separation.
Fourth, restore and reconnect. Starting a de-energized region presents different challenges from keeping an existing island stable. Restoration requires sequencing and potentially black-start-capable resources. Reconnecting two live islands also requires acceptable differences in frequency, phase angle and voltage.
Now look at how much software is involved.
You need reliable topology information, measurements, event detection, state estimation, local control, coordinated protection, restoration procedures and communications that do not themselves become a single point of failure.
And there is one requirement I would borrow directly from resilient software design:
The tests have to cover the failures the system is supposed to survive.
Seven islands on an architecture diagram are not enough. They have to work when the connections actually fail.
Wouldn’t shared capacity be better than isolated capacity?
That brings me back to the availability-zone analogy.
In a well-designed distributed software system, resilience often comes from combining independent failure domains with enough spare capacity and connectivity for surviving nodes to take over the work of failed ones.
Could the seven electricity regions support each other in a similar way?
In some conditions, yes. Geographically distributed generation, storage, reserves and strong interconnections can make a power system more resilient. Surviving regions may be able to support each other when adequate transmission paths remain.
When those interconnections fail, sharing resources becomes impossible. Even healthy connections have thermal and stability limits, so the network can support only certain transfers.
That means the design has to support both modes: share resources when the network is healthy, and preserve essential local supply when it is not.
Of course, both approaches have a price.
In cloud systems, distributing a service across several regions usually costs more. We accept replication, infrastructure overhead and sometimes additional latency in exchange for availability.
In the grid, the trade-offs are different: additional dispatchable generation, reserves, storage, protection and control capability, potentially less efficient dispatch, and sometimes deliberate load prioritization during an emergency.
Litgrid estimates that its proposed isolated-operation services would add 0.67 euro cents per kWh to the electricity transmission tariff after 2033. The figure is a projection for the proposed programme; the final tariff has yet to be approved. Litgrid also expects the new flexible plants to help moderate wholesale price peaks.
At least this gives us a number to discuss when talking about the cost of resilience.
But what exactly would we be paying for: independent regional operation, the ability to maintain service across the country after a major failure, or some combination of both? I’d want to measure how much load each region can serve, for how long, under which failures, and how the system recovers as a whole.
Keeping individual regions alive is useful. I am equally interested in what happens to consumers in a region that cannot survive on its own, and whether the wider grid can help it recover.
Lithuania’s design is still a proposal, with details open for public consultation until 10 December.
Around the grid
Britain has electricity. Can it move it where it’s needed?
On 9 November, Britain’s National Energy System Operator published the second part of its Balancing Act series.
Its reported balancing costs for 2025/26 reached £3.1 billion, 16% above the previous year. Thermal transmission constraints were the largest component. Just under £0.5 billion was paid to wind farms to reduce generation, and NESO says it now performs more than twice as much balancing activity per unit of electricity consumed as in 2018.
Wind curtailment makes an easy headline. The geography of the system explains much of the problem.
Imagine that wind farms in Scotland can produce plenty of electricity while consumers farther south need it. If the transmission network cannot safely carry the required power, the operator may have to reduce generation north of the constraint and increase it somewhere else.
Even with enough generation overall, delivering it where and when it is needed can be impossible within the network’s limits.
NESO gives a useful comparison: October 2025 had lower potential wind generation and lower wholesale prices than December, yet balancing costs were 34% higher. One important difference was that less transmission capacity was available because of outages.
This is where I think better software can genuinely help.
Better forecasts, network models, outage planning, security-constrained optimization and flexible demand can reduce unnecessary cost. NESO estimates that its actions and those of its partners avoided £891 million in 2025/26.
The optimization still has to respect the thermal and stability limits of the network.
We will still need more dependable generation, storage and flexibility for certain hours, regions and future scenarios.
At the same time, coordinating generation, demand and transmission is becoming a major engineering challenge of its own. When that coordination costs billions, I think the models and software deserve much more attention.
Norway needs to know how data centres behave
On 9 November, Statnett temporarily suspended new grid-capacity reservations for data centres. Norway’s government followed with proposed regulatory changes.
Just look at the numbers.
Norway has around 800 MW of connected data-centre demand, with 4,200 MW reserved, 9,100 MW in the connection queue and another 18,000 MW in additional enquiries. The country’s peak demand is just over 25,000 MW.
Much of that demand is still at the reservation or enquiry stage. Even so, the scale is remarkable.
The queue is an obvious infrastructure bottleneck, with more proposed projects than substations and transmission lines can accommodate. Statnett is also concerned about how these loads operate: rapid changes in demand, software-controlled loads, sensitivity to disturbances and the possibility of contributing to network oscillations.
It is considering requirements for how quickly load can change, when facilities must remain connected during disturbances, and how compliance will be demonstrated.
This is the part that caught my attention. We usually describe a data centre by its power demand in megawatts. The system operator also has to consider its behaviour as a huge, software-controlled load.
A software workload scheduler can affect a physical power-demand profile.
The connection study therefore needs to answer several questions: How many megawatts? How quickly can demand change? How will the facility behave during a disturbance?
Europe’s ability to deploy computing infrastructure, and the speed of inovation in general, increasingly depends on how quickly it can build and safely integrate the electrical systems, alongside securing chips and developing models.
What if hosting capacity could be recalculated instead of assumed?
At Enlit Europe in Vienna, Plexigrid presented its approach to continuously updated hosting capacity.
Hosting capacity is the additional generation or consumption a network can accommodate without violating its operating limits.
Traditionally, an operator may use conservative studies to decide how much additional solar generation, EV charging or other load can connect to a feeder.
The problem is that the feeder does not behave the same way throughout the day.
A transformer may have substantial spare capacity overnight and very little during the evening peak. Solar generation may cause voltage rise on one feeder at noon, while EV charging causes voltage drop later. Network switching and equipment outages change the picture again.
And it made me wonder:
What if the DSO could regularly recalculate available capacity using topology, measurements and forecasts?
In principle, a system could combine network models, telemetry, state estimation and power-flow calculations, then generate a time-dependent operating envelope for flexible loads and generation.
There is a catch, though. Someone also has to make sure the equipment respects the calculated limits.
The model may be wrong. Telemetry may be missing. A charger may ignore its requested limit. A communications link may fail.
To make dynamic capacity usable safely, the operator needs uncertainty margins, controllable devices, enforcement mechanisms and conservative fallback behaviour.
Plexigrid’s presentation is a good starting point for exploring this problem. I’d still want to see the actual deployments and their results before judging how well the approach works.
When that calculation starts influencing physical equipment, the planning software becomes part of a distributed control system.
Generation, storage and transmission should be designed together
On 12 November, Turkey announced a preliminary cooperation agreement with China’s National Energy Administration.
The stated ambition is approximately $15 billion in investment, including 7 GW of wind and solar generation, new storage, and transmission infrastructure intended to handle 4 GW.
The agreement is preliminary, with the projects still to be developed. I like that generation, storage and network capacity appear in the same plan.
We often treat a solar farm, a battery and a transmission connection as separate projects. In operation, they are parts of the same power system. Their locations, capacities, controls and constraints determine what the system can actually deliver.
I would like to know how that combination behaves when generation varies, a line becomes congested, or part of the system fails. The number of announced gigawatts tells us little about that.
A hidden dependency can take down a very visible system
A suspected drone attack on Iraq’s Khor Mor gas field on 8 November halted gas production. Regional authorities reported that about 2.8 GW of electricity generation was affected.
The electricity generators lost a common upstream dependency: the gas supply.
In software terms, this is a common-mode dependency.
You can deploy multiple services, or multiple generators, and still have one point of failure if they all depend on the same upstream resource.
You can count as many redundant instances as you like. If they share a critical dependency, you still have a single point of failure.
AI infrastructure is becoming energy infrastructure
On 12 November, Reuters reported SpaceX’s plans for more than $74 billion in AI computing campuses in Texas and Arizona.
The proposed Texas campus would generate its own electricity using a gas-fired plant, while the Arizona project involves coordination with local electric cooperatives. Both are proposed investments, and the Arizona campus still needs local approval.
I was reminded of Elon Musk’s July interview with The Economist, where he argued that electricity and cooling can constrain AI expansion more than the supply of chips, particularly outside China.
That is his assessment; I wouldn’t treat it as a universal rule.
Still, if a computing campus needs its own generation and major electrical connections, where does the data centre end and the power system begin?
Worth watching: Lithuania’s open consultation
Litgrid’s consultation on regional isolated-operation services is worth reading if you want to go deeper. It is open until 10 December 2026.
I would be particularly interested in the technical requirements for restoration, sustained islanded operation, resource availability, and interactions between local control and the national system operator.
This is where an attractive architecture diagram has to turn into engineering requirements that can actually be tested.
The thread I see this week
I keep coming back to the different roles that network connections play in these stories. Lithuania wants to keep supplying electricity when some of them fail. Britain is spending billions because existing connections cannot always carry the electricity that is available. Norway needs to know how huge, software-controlled consumers will behave before connecting them. And Plexigrid is looking at how much additional capacity a network actually has as conditions change.
These are different problems, and I don’t think there is one solution for all of them.
Our software needs to represent the grid’s physical limits, operating modes and failure dependencies alongside its nominal capacity.
Better models and algorithms can improve planning, dispatch, coordination and utilization. Stronger infrastructure, local reserves and carefully designed protection are still essential.
The cloud analogies are useful to a point. Electrical networks bring constraints that software infrastructure rarely has to consider.
I still see useful lessons from distributed systems here: identify failure domains, avoid hidden shared dependencies, understand degraded modes, maintain observability, and test recovery.
Here, a failed recovery procedure can interrupt electricity supply, with consequences far beyond a failed API request.
See you next week.
— Dmytro
Software in the Grid is an independent publication. The views expressed here are my own.

