Innovating with AI requires scarce resources like GPUs, leading to the development of Dynamic Workload Scheduler (DWS) and its new feature, Flex-start VMs. Flex-start VMs offer a unique consumption model, prioritizing fair access to GPUs for demanding tasks. They are ideal for time-bound workloads such as model fine-tuning and batch inference. A key benefit is drastically improved resource obtainability by allowing requests to queue for up to two hours. Users also benefit from cost-effective pricing, with discounts compared to standard on-demand rates. Flex-start VMs can run for a maximum of seven days and utilize preemptible quota. A new parameter, request-valid-for-duration, allows users to specify how long their VM request should wait for capacity. This contrasts with the previous "fail-fast" approach which would quickly reject unavailable resources. Flex-start VMs provide direct API access, enabling easy integration into existing workflows and schedulers. Users retain control with stop and start capabilities, and a configurable termination action preserves VM configurations upon expiry.