// Case studies
Real workloads. Measurable outcomes.
How deep-tech teams build on Gensparkly.
LLM Training · Northwind AI
Training a 70B model on 512 × H100.
Reduce pre-training wall-clock from months to weeks without exceeding budget.
Stack
512 × H100 SXM5 · NDR InfiniBand · 3 PB parallel NVMe
Result
3.4× faster convergence · 42% cost reduction
Inference at Scale · Meridian Labs
Serving 12M inference requests per day.
Deliver sub-40 ms p99 latency for a multimodal model across three regions.
Stack
L40S fleet · Triton Inference Server · TensorRT-LLM
Result
38 ms p99 · 99.99% availability
HPC & Simulation · Helion Dynamics
Computational fluid dynamics at scale.
Run large-scale CFD simulations previously constrained by on-prem capacity.
Stack
H100 bare metal · NVLink islands · low-latency scratch
Result
6× cost reduction · 4× simulation throughput