【Microsoft lets AI calculate the power grid first—can models solve power shortages?】
When talking about AI and electricity, people tend to think only about building more power plants. But once power has already been generated and fed into the grid, the question is whether it can be delivered to where it’s needed at low cost.
In its September 30 grid planning presentation, Microsoft Research Institute showcased GridSFM: an approximate computation method that uses models to accelerate alternating optimal power flow. This research was previously made public in May—not something that was suddenly invented overnight as a “power supply button.”
In plain terms, what it studies is: when load, lines, and generation conditions change, how should the grid arrange operations more appropriately. The model can help filter scenarios faster, but the predicted results still need to satisfy physical constraints and real operational requirements.
The business significance of this approach is that it may improve the utilization efficiency of existing facilities. But it can’t create transformers out of thin air, nor can it infinitely expand the capacity of constrained transmission lines. Software optimization and hardware expansion address problems in different parts of the system.
When it comes to AI infrastructure, I would ask where the savings actually happen: in computation time, in dispatch costs, or in the real-world ability to supply more electricity? The three cannot be mixed into a single lump-sum revenue. When discussing AI narratives related to RENDER, FET, and NEAR, you also need to separately consider compute demand and the underlying ability to provide power—you can’t directly attribute Microsoft’s research to token gains.
Calculating faster is about finding solutions; only delivering real power is the proof of capability.
The image is a stock photo of a Microsoft office building.
$RENDER $FET $NEAR
Click on my avatar to view trades with single-account live orders
When talking about AI and electricity, people tend to think only about building more power plants. But once power has already been generated and fed into the grid, the question is whether it can be delivered to where it’s needed at low cost.
In its September 30 grid planning presentation, Microsoft Research Institute showcased GridSFM: an approximate computation method that uses models to accelerate alternating optimal power flow. This research was previously made public in May—not something that was suddenly invented overnight as a “power supply button.”
In plain terms, what it studies is: when load, lines, and generation conditions change, how should the grid arrange operations more appropriately. The model can help filter scenarios faster, but the predicted results still need to satisfy physical constraints and real operational requirements.
The business significance of this approach is that it may improve the utilization efficiency of existing facilities. But it can’t create transformers out of thin air, nor can it infinitely expand the capacity of constrained transmission lines. Software optimization and hardware expansion address problems in different parts of the system.
When it comes to AI infrastructure, I would ask where the savings actually happen: in computation time, in dispatch costs, or in the real-world ability to supply more electricity? The three cannot be mixed into a single lump-sum revenue. When discussing AI narratives related to RENDER, FET, and NEAR, you also need to separately consider compute demand and the underlying ability to provide power—you can’t directly attribute Microsoft’s research to token gains.
Calculating faster is about finding solutions; only delivering real power is the proof of capability.
The image is a stock photo of a Microsoft office building.
$RENDER $FET $NEAR
Click on my avatar to view trades with single-account live orders
