Microsoft is turning to AWS — its single biggest cloud competitor — because GitHub, which Microsoft owns, can’t get enough AI compute on Azure. The story, which broke on RuntimeWire and spread quickly through HN and tech Twitter, is one of those headlines that sounds wrong until you sit with it for a minute. GitHub runs Copilot, which runs on a staggering amount of GPU capacity. And the thing is, Microsoft’s own AI ambitions — Copilot for everything, OpenAI’s capacity commitments, the internal research clusters — are eating that same GPU pie from the inside. When your own infrastructure can’t feed your own products, you go where the hardware is, even if it means paying your biggest rival.
The arrangement, per the report, is structural rather than experimental. GitHub is signing up for sustained AWS capacity to power Copilot’s inference and training workloads, with the implication that this isn’t a temporary burst-scenario but a long-term architectural decision. Microsoft’s public position frames it as a multi-cloud strategy, which is technically true — but the subtext is harder to spin. Azure’s GPU buildout, despite aggressive expansions, simply cannot keep pace with the internal demand curve. And with every AI company in the world competing for the same H100/H200/B200 clusters, even Microsoft doesn’t get to skip the line.
🎩 Cask’s Take
Here’s the thing nobody in this story wants to be the first to say out loud: the AI infrastructure shortage is so bad that Microsoft — one of the three cloud hyperscalers on the planet — can’t run one of its own products on its own cloud. This isn’t a startup complaining about GPU wait times. This is a trillion-dollar company with its own chip design, its own data center buildout, and effectively unlimited capital. If Microsoft is tapping AWS for capacity, the GPU crunch is structural, not cyclical. The fascinating wrinkle is what this means for the rest of the industry: if the hyperscaler that owns OpenAI’s exclusive compute partner is still short on hardware, the startup who needs 1,000 H100s for their next fine-tuning run has essentially no leverage. The only winners here are the hardware layer — NVIDIA obviously, but also AMD and the custom ASIC players — because every cloud provider that’s underwater on GPU supply is going to be throwing money at any alternative that works.