There's a bias in tech toward distributed everything. Distributed computing, distributed energy, distributed manufacturing. The assumption is that smaller and local is inherently better. It's usually more flexible because you're relying on more generic instruments (3D printers versus injection moulding). But this shouldn't be conflated for inherent benefits of decentralized infrastructure.

A lot of the time when you actually run the numbers, centralized infrastructure wins pretty clearly in efficiency and economics.

Take solar energy. The popular argument for distributed solar is pretty straightforward. Put panels on every rooftop, eliminate transmission losses, minimize grid overhead, and democratize energy production. But it's worthwhile to break it down.

Typical rooftop solar panels max out at 22% efficiency. Solar cell tech can get higher but the commercial multicrystalline or lower-end monocrystalline silicon cells used in rooftop have to be mass produced so benchmark at lower than their theoretical lab research cell limits. They don't provide any internal storage so they need to delegate to an onsite battery if you want to charge anything. Most dump their excess capacity back into the grid.

Large-scale concentrated solar power (CSP) plants achieve efficiencies of 35-45% by using mirrors to concentrate sunlight and drive turbines. That underlying turbine tech is the same that we use in natural gas power plants. The largest CSP plant in Morocco uses 170,000 mirrors to generate 580MW and can store energy for up to 8 hours.¹ The only caveat is that Morocco is sunny almost year round.

Transmission losses are where the distributed argument usually gets made. The US electrical grid loses about 3-8% of electricity during transmission and distribution. But it's not enough to offset the gains of the centralized install. Even with transmission losses, the centralized system delivers a minimum of 46% more usable energy per unit of sunlight.²

Metrics are almost everything in AI. There seem to be two main ways that the energy sector thinks about cost:

• Economic cost (energy per watt), this just tracks the power bill rate that consumers will see when they pay
• LCOE (levelized cost of energy), which tries to model the total cost per MWh over the system's lifetime

Consumers care more about the economic cost. Utilities care about the LCOE. Here's how they breakdown.

Rooftop solar costs roughly $2.50-2.74 per watt installed in economic costs according to the latest DOE benchmarks. Large utility-scale solar farms cost $0.98 per watt. The economies of scale are a good 2.5-3x cost difference per unit of capacity.

Rooftop solar LCOE runs $80-120/MWh. Utility-scale solar LCOE runs $30-50/MWh. CSP with storage runs $70-100/MWh. Even accounting for transmission costs (roughly $5-10/MWh), centralized generation wins flat out. You can even integrate a separate storage system for the same overall LCOE as utility scale solar.

Economies of scale here apply to physical build outs in a way that they don't apply to systems design.

A 500MW solar farm has roughly the same interconnection costs as a 5MW distributed system. Centralized facilities can justify specialized equipment and expertise. CSP plants use optimized mirror arrays that would never be cost-effective at residential scale. The fixed costs like engineering, permitting, and grid connection get amortized across much larger capacity.

Large central plants can also provide grid services like frequency regulation and voltage support that you would otherwise need to pay a separate grid operator to provide.

There's a place and a time for distributed infrastructure:

• Transmission is impossible: remote locations where building transmission lines costs more than local generation
• Reliability is paramount: military bases or hospitals that need guaranteed power independent of grid failures
• Regulatory capture: markets where utilities prevent fair interconnection of central plants

But these are edge cases. For the vast majority of energy demand, centralized generation with smart transmission is more efficient.

The distributed argument seems mostly political. It seems easier - or at least faster - to acquire the permitting for distributed infrastructure in the US. When people own their homes they can largely do what they want. But centralized installs can take years of environmental reviews and public comment.

It seems that we don't see more centralized approaches because of the regulatory regime and not the inherent disadvantages of the technology. As engineers considering the actual numbers, I think we need to be more forceful advocates of the tech itself. The politics is a separate conversation.