Follow-the-Sun and Follow-the-Price Routing
Spatial shifting is the routing of workloads across a geographically distributed fleet so that each job runs in the region with the most favorable power conditions at that moment. Where temporal shifting exploits how the grid changes over the day at one site, spatial shifting exploits how the grid differs across sites at one instant. Combine the two and the scheduler shops across both time and space for the best cell in a grid of options.
The canonical patterns have names. Follow-the-sun routing sends flexible compute toward regions currently rich in solar generation, chasing midday troughs of cheap, clean power around the globe as the earth turns. When California's solar peaks, batch work flows west; hours later, as panels in Europe or Asia hit their own midday, the flow follows. Follow-the-price routing generalizes this to any signal: send work to whichever region has the lowest marginal power price or carbon intensity right now, whatever the source. Solar is one driver; surplus overnight wind, abundant hydro, or simply slack demand in a region are others. The scheduler does not care why a region is favorable, only that it is, and that it has spare capacity to accept the work.
The mechanism is a global scheduling layer sitting above the regional clusters. Each region reports its live state: available accelerator capacity, current and forecast power price, carbon intensity, and grid headroom. A deferrable job entering the system is not bound to its origin region. The global scheduler scores every region that can serve it, on a combined function of power conditions, available capacity, and the cost of getting the job's data there, then dispatches to the winner. A latency-critical job, by contrast, is pinned near its users and never enters this competition. As with temporal shifting, only the deferrable class is mobile, and the classification done in the queue is what makes the routing possible.
The economic case rests on the spread between regions being large and persistent. At any given instant, power prices across a continent's grids can differ severalfold, and carbon intensity can differ by an order of magnitude between a coal-heavy region and a hydro- or nuclear-heavy one. A job that would run on expensive, carbon-intense local power can instead run on cheap, clean remote power, and the saving per unit of compute is often far larger than anything temporal shifting alone can reach at a single dirty site. The multiplier is geography: more regions means more of the curve to shop across, and the best cell in a wide grid of regions and hours is better than the best hour at one location.
Capacity coupling is the constraint that keeps this honest. A region is only a valid target if it has spare accelerators to run the job. Cheap clean power in a region whose GPUs are all busy is useless to an incoming job. So the global scheduler solves a joint problem: it matches mobile work to regions that are simultaneously power-favorable and capacity-available, and those two conditions do not always coincide. This is why spatial shifting works best across a fleet deliberately built with headroom and diversity, spread across grids that peak at different times and run on different fuel mixes, so that when one region is expensive and full, another is cheap and open.
Routing also interacts with reliability and congestion. Concentrating too much flexible load into one favorable region can saturate its compute, its network ingress, or even stress the local grid the scheduler was trying to treat kindly, recreating the herding problem across space. Mature routing therefore spreads work across several good-enough regions rather than piling it all onto the single best one, and it respects transmission and network limits as hard constraints. The objective is not to find the theoretical optimum for one job but to keep the whole fleet running near the efficient frontier without any one region tipping over.
The reframe spatial shifting offers is that a distributed fleet is not a collection of independent buildings but a single, planet-scale computer with a power-aware dispatcher. That dispatcher treats the global grid as a heterogeneous, ever-shifting supply of energy and places each mobile job where the energy is best and the capacity exists. The ceiling on this technique is not the scheduling logic, which is well understood. It is the cost of moving the data the job needs, which the next lesson takes head-on.