The Constraints: Data Gravity, Latency, and Replication

Spatial shifting sounds unbounded until you remember that compute is useless without its data, and data is heavy. This lesson is about the forces that pull against routing and decide which jobs can actually fly. Ignore them and a clever router will happily send a job to cheap power in a region where its dataset does not exist, then spend more on moving terabytes than it ever saved on electricity.

Data gravity is the governing concept. Large datasets exert a pull on compute: it is cheaper to bring the computation to the data than the data to the computation, so work tends to accumulate where its data already lives. A model-evaluation job that reads a hundred-terabyte corpus is effectively welded to the region holding that corpus, because shipping the corpus elsewhere would cost more in transfer fees, time, and network capacity than any power saving could recover. A job that reads a few gigabytes and writes a small result is nearly weightless and can run anywhere. The scheduler's routing freedom for a job is inversely proportional to the mass of data that job must touch. The first question of spatial shifting is therefore not where is power cheap but how heavy is this job.

The transfer math is explicit and belongs in the fitness function. Relocating a job incurs the cost and latency of moving its input data to the target region, and moving results back. Inter-region bandwidth is finite and often metered, so bulk transfer is neither free nor instant; a large dataset can take hours to move and cost real money in egress charges. The router must compare the power saving from running remotely against the transfer cost of getting there, and shift only when the saving clearly wins. For a small, long-running, compute-heavy job the transfer is a rounding error and shifting is easy. For a short job on a huge dataset, the transfer dominates and shifting is a loss. Between them lies a break-even the scheduler must estimate per job.

Replication changes the calculus by paying the data-movement cost in advance. If a dataset is pre-replicated to several regions, jobs that use it become mobile across all of them, because the data is already there when the router wants to send work. This is how mature fleets buy routing freedom: they identify hot datasets and maintain copies in multiple regions so that flexible work over those datasets can chase favorable power without per-job transfer. Replication is not free, it costs storage in every region and bandwidth to keep copies consistent, so it is applied selectively to the data that flexible workloads actually touch. The strategic move is to invest in replication where it unlocks the most shiftable compute, converting a one-time and ongoing storage cost into standing routing optionality.

Latency draws the hard line between what can shift and what cannot. Latency-critical inference must run physically near its users to meet millisecond budgets, because the speed of light imposes a floor on round-trip time that no scheduler can beat; a request served from another continent adds tens to hundreds of milliseconds that an interactive service cannot absorb. This work is pinned by physics and never enters spatial routing. Deferrable batch work, which no user is waiting on, is indifferent to its own location's distance from users and is free to fly, subject only to data gravity. The latency budget, in effect, sorts the fleet's work into the pinned and the mobile before the router ever runs.

Data residency and sovereignty add constraints that are legal rather than physical, and they are absolute. Regulations may require that certain data, personal, financial, medical, or national, remain within specific jurisdictions and never cross a border. A job touching such data cannot be routed out of its permitted region no matter how favorable power is elsewhere. The scheduler must treat these boundaries as inviolable placement constraints, encoded per dataset, and refuse any routing that would violate them. Efficiency never overrides compliance; the router optimizes only within the set of legally permitted regions.

The synthesis is a routing policy that scores each mobile job on power conditions minus data-movement cost, filtered first through hard constraints of latency, capacity, and legal residency. Weightless, deadline-slack, unrestricted jobs enjoy the full menu of regions and deliver most of spatial shifting's value. Heavy, latency-bound, or regulated jobs stay home. The engineering craft is recognizing which is which, and the investment craft is pre-positioning data through replication so that more of the fleet's valuable, flexible compute joins the mobile class. Spatial shifting rewards those who plan their data placement as deliberately as their power sourcing, because in the end the router can only move what the data will let it.