01
MANAGED
Inference
Send OpenAI-compatible requests. nafnaf routes each one to healthy capacity and streams the response back.
Three ways to run
PRODUCT
02
INTERRUPTIBLE
Spot Jobs
Run checkpointable batch workloads on the lowest-priced available hardware, with retries when capacity moves.
03
PROTECTED
Protected Jobs
Keep a GPU allocated for work that cannot be marketplace-preempted, with clear runtime and budget controls.
HARDWARE
Two architectures. One control plane.
Workloads declare what they need. nafnaf selects compatible capacity without pretending every accelerator is interchangeable.
NVIDIA · CUDA
RTX 3090 to B300
- 24 GB or more VRAM
- Managed inference and containerized jobs
- Consumer cards through datacenter clusters
APPLE · METAL
M3 Ultra and newer
- 96–512 GB unified memory
- Native MLX, Metal, and approved runtimes
- Large-model inference and memory-heavy jobs
How the market works
PROTOCOL
01
Verify
Providers connect eligible hardware. We benchmark performance, memory, thermals, storage, and network quality.
02
Route
Builders choose a workload contract. The control plane matches architecture, region, price, and reliability.
03
Settle
Usage is metered consistently across providers. Builders pay for execution; providers earn for useful compute.
FOR BUILDERS
Run work, not infrastructure
- One API for inference and scheduled jobs
- Spot or Protected execution per workload
- Explicit hardware, region, and budget constraints
FOR PROVIDERS
Put idle compute to work
- Connect NVIDIA GPUs or an eligible Mac Studio
- Control availability, workload modes, and price floors
- Earn across inference, Spot, and Protected execution
EARLY ACCESS
Be first on the grid.
Join the founding cohort of builders and compute providers. We’ll send your invite as soon as access opens.