Bare-Metal ServerA physical server used exclusively by one customer.CUDANVIDIA's platform for general-purpose GPU computing.DDoS (Distributed Denial of Service)An overload attack from many sources that knocks services offline.DPA (Data Processing Agreement)The contract required under Article 28 GDPR for processing data on your behalf.Fair-Use TrafficTraffic without per-gigabyte billing.Inference vs. TrainingThe two phases of an AI model — with very different resource needs.KVM VirtualizationHardware-assisted virtualization built into the Linux kernel.NVMeFast SSD storage over the PCI Express bus.OverprovisioningSelling more resources than physically exist.Preemption (Spot Instance)The forced shutdown of a cloud instance by the provider.Snapshot vs. BackupTwo ways to preserve data — with very different protection.TDP (Thermal Design Power)A GPU's power envelope — and what the power limit means for performance.Tensor CoresSpecialized compute units for AI matrix operations.vCPUThe compute unit assigned to a virtual machine.VRAM (Video RAM)The dedicated graphics memory on a GPU.