National Compute Opens B300 Access

This is a meaningful attempt to democratize frontier compute, but the “free GPUs” framing needs qualification: $100 buys experimentation, not sustained training, and capacity is metered by the GPU-hour.

• –Public-research users receive an entire eight-GPU node rather than a shared slice.
• –The B300 configuration targets CUDA-heavy workloads, while AMD MI355X nodes offer a ROCm alternative.
• –Four-hour leases, queue-based access, preemption, and single-node limits make this better for focused experiments than long-running training.
• –Local node storage is erased on release, so checkpoint and artifact management is essential.
• –The $100 credit lowers the barrier to entry but does not eliminate ongoing billing after the starter balance is exhausted.