Axel Rietschin, a former Azure Core engineer, has published a detailed account of organizational dysfunction at Microsoft that nearly cost the company its relationship with OpenAI and damaged trust with the US government. The story centers on Azure Boost, an internal project code-named Overlake, where management allegedly pushed to port complex Windows features onto a tiny ARM SoC with just 4KB of memory. The chip was fanless and roughly fingernail-sized. Rietschin, who had previously worked on early Overlake designs and held multiple patents from his time on the Windows kernel team, walked into his first meeting in May 2023 and found leadership seriously entertaining this technically impossible plan.
The deeper problems were even stranger. Rietschin discovered that Azure nodes were running 173 different management agents, and nobody could explain what most of them actually did or why they existed. These agents were constantly hammering the hypervisor with WMI calls, creating instability that got worse when Azure increased VM density from 32 to 48 per node. Crash rates jumped 50%. The platform was already straining on 400-watt Xeon processors, yet the plan was to somehow squeeze this bloated stack onto hardware that could barely blink an LED. Noisy neighbors can create similar resource contention issues in cloud environments.
The contrast with AWS Nitro is stark. Amazon built custom hardware with a lightweight hypervisor that cleanly offloads virtualization work. Azure tried to replicate this by cramming Windows components onto underpowered chips while keeping the WireServer metadata service on the host OS, where guest VMs could potentially compromise the entire machine. Rietschin describes a culture of fear that blocked necessary refactoring, with sales contracts taking priority over software quality. Microsoft hasn't publicly responded to these claims, but if accurate, they raise real questions about the infrastructure running a big chunk of the world's AI workloads.