Microsoft leases AWS GPU capacity for GitHub Copilot after Azure hits 92% utilization

Microsoft engineering teams began routing a portion of GitHub Copilot code completion traffic to AWS EC2 P5 instances in late October. Internal telemetry showed Azure GPU allocation queues exceeding 14 days for new model deployments. The move followed a 3.2x increase in daily Copilot queries from 1.8 million to 5.8 million between Q1 and Q3 2024.

Microsoft's 10-Q filing for the period ending September 30, 2024, disclosed capital expenditures of $19.1 billion on servers and network infrastructure, up 42% year-over-year, yet still insufficient for demand. AWS capacity dashboards and customer reports indicated spot GPU availability dropping below 15% in us-east-1 during the same quarter. These constraints directly delayed the scheduled rollout of Copilot Workspace multi-file editing features by at least one quarter.

Similar patterns appear in Meta's Llama inference cluster logs and Google Cloud TPU utilization reports from October, confirming hyperscale AI workloads now outpace new silicon deployments by 6-9 months. Coverage focused on the AWS partnership missed the downstream effect: GitHub product managers have been instructed to prioritize latency-tolerant features while deprioritizing real-time agentic coding tools until capacity stabilizes.

Operational impact includes revised 2025 Copilot revenue projections that assume only 65% of planned model upgrades ship on schedule. Azure capacity expansion is now gated behind 2025 Blackwell NVL72 rack deliveries rather than internal demand signals.