Your next idea.
More compute.
Private inference. Fine-tuning.
On GPUs reserved for you.
Run and adapt open models on dedicated hardware, instead of sending your prompts and data to a shared model API.
Check availabilityPre-launch orders open · Rentals start 12 OctoberYour own models.
On your own terms.
Inference and fine-tuning on hardware
reserved for you, not a shared API.
Keep it private.
Make it yours.
Keep sensitive prompts and datasets off third-party model APIs.
Serve open models privately.
Adapt them to your domain.
Private inference
Serve open-weight LLMs behind your own endpoint. Your prompts and outputs stay on a machine dedicated to you, not a shared model API.
Fine-tuning
Adapt open models to your domain with your own datasets, from LoRA and QLoRA to full fine-tunes where memory allows.
Evaluate and prototype
Benchmark models, test retrieval pipelines and agents, and validate a workload before scaling it up.
Starting with DGX Spark.
Our first machine, with more GPUs to follow.
We match the hardware to your workload.
Unified system memory
CPU and GPU share one coherent memory pool, giving large models more room to work.
AI compute performance
NVIDIA GB10 Grace Blackwell brings a new generation of AI compute to your experiments.
Parameter model support
Explore inference with large models on a single system, depending on model and precision.
*Up to 1 PFLOP at FP4 with sparsity. Model capacity depends on quantization and workload; these are hardware specifications, not hosted performance guarantees. NVIDIA specifications
Your workload.
Your dates.
Pick the days you need the NVIDIA DGX Spark and tell us what you want to run, inference or fine-tuning. We’ll review your request and confirm fit, pricing, and access by email before anything is reserved.
Before you build.
When can I start?
Pre-launch orders are open now, and rentals start on 12 October 2026. Rentals are booked by the week, with a one-week minimum. The machine is yours alone for the whole booking, and storage is wiped before the next customer. Pick your dates in the booking calendar below and we’ll confirm them with you by email.
What will a rental cost?
$145 per week, or $480 per 4-week block. Pre-launch orders placed before 12 October get $110 per week. Need a longer or custom arrangement? Tell us about your workload and we’ll work it out with you.
Where does my data go?
Your workload runs on a machine dedicated to you, rather than a shared model API. Tell us about your data requirements in your enquiry and we’ll walk you through access, storage, and isolation before you commit.
Which GPUs are available?
We’re starting with NVIDIA DGX Spark and plan to add other GPUs over time. Tell us what your workload needs so we can match the right hardware.
Is DGX Spark the same as renting an H100?
No. DGX Spark uses the NVIDIA GB10 Grace Blackwell Superchip with 128 GB of unified system memory. It is designed for AI development, prototyping, and inference. Its memory, bandwidth, and performance characteristics differ from datacenter GPUs.
Will my software run on DGX Spark?
DGX Spark runs an Arm-based CPU and NVIDIA’s AI software stack. Your containers and dependencies need compatible Arm64 builds and GPU support. Include your framework and model in your enquiry so we can discuss compatibility.
Is Lycompute affiliated with NVIDIA?
Lycompute is an independent service. NVIDIA, DGX Spark, and related marks belong to NVIDIA Corporation. No partnership or endorsement is implied.