XWind: A Cross-site Router for Large Language Model Inference Serving at Renewable Energy Farms

AI power demand is growing while grids struggle to keep up; grid expansion carries high capital cost and transmission losses, yet abundant renewable energy exists at the source, just not matched to demand.

Proposes 'AI Greeninferencing' - bringing modular AI compute to renewable energy sources (wind) to create behind-the-meter demand and ease strain on utilities - plus XWind, a lightweight reactive workload-agnostic inference router using only real-time signals (inference latency, KV-cache utilisation, queue depth) to configure sites and distribute requests under variable wind power.

Reduces P99 end-to-end latency by up to 52% over the strongest contender - which is also the authors' own idea - and by up to 98% over baselines SUCH AS POWER-CAPPING AND GPU IDLING. Evaluated on a real 64-GPU A100 testbed emulating three wind-powered sites using Azure production traces, with consistent gains across workload types, load levels and GPU generations. IMPORTANT CAVEAT: the 98% figure is against the authors' own chosen baselines, not against the absence of the technique, so it measures how much better routing is than static power management - it is not a general claim about cross-site routing performance.

INCLUDED BECAUSE IT INVERTS THE USUAL SIGN AND IS THEREFORE A TRAP FOR THE READER WHO SKIMS. Power-capping and GPU idling are this paper's BASELINES TO BEAT, not its recommendations - so it should not be cited as evidence that power-capping is a good idea. Its actual distinctive contribution to this slice is geographic: it treats the interconnector and transmission constraint recorded in supply/grid-interconnection-queue-power-not-chips as something to engineer around by co-locating compute with generation, and its feasibility analysis asserts 890+ GW of wind capacity lies within 50 ms network round trip time of Azure data centers. That 50 ms figure is the operative constraint and is a latency budget, not a power figure - the interesting inversion is that a POWER problem gets solved by a NETWORK-LATENCY argument. Note the tension with supply/grid-interconnection-queue-power-not-chips: behind-the-meter co-location is precisely the case the LBNL dataset excludes from its queue statistics, which is why that record flags its scope limits so explicitly. Affiliations are unverified - the author list spans multiple organisations and no affiliation block was captured. AFFILIATIONS BACKFILLED 2026-10-04 (paper-fields pass, agent w3p): PDF title block states 'Microsoft' for all ten authors; OpenAlex returns none for this arXiv ID. Filled from the API/PDF rather than from memory, because an affiliation is a fact about a specific paper's author block and inference from the author's employer history is how wrong affiliations get shipped. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.