GigaCompute launches sub-$0.01 GPU-hour instances for burst AI training

Tech · 4 min read

GigaCompute revealed a new pricing tier for GPU-burst instances aimed at short-duration training and hyperparameter sweeps. The offering, backed by custom re-purposed hardware and predictive scheduling, undercuts traditional cloud GPU pricing for workloads under five minutes.

The company emphasized the cost-benefit for people doing large-scale parameter sweeps or gradient-checkpoint experiments where many tiny runs are needed. To ensure fairness and reliability, GigaCompute uses a task scheduler that prefers spot-like capacity but provides soft guarantees on job completion for burst workloads.

GigaCompute also introduced orchestration primitives that integrate with popular ML frameworks and experiment-tracking tools. The product includes a credits program for early-stage AI teams and academic researchers to test large sweeps without significant upfront cost.

Investors see this as a response to demand for affordable compute during the prototyping stage of model development. GigaCompute will expand region availability and add higher-duration GPUs for production training in the second half of 2026.