Technology · Platform

RunPod

Rented cloud GPU spot capacity is one provisioning target behind my substrate's fail-closed daily budget; always available, never uncapped.

FAFO Core · Inference FabricRented spot GPUsBehind a fail-closed daily budget
Spend in the critical path
spend, counted per secondpods past the cap are not started
  • 0dollar cap on a day with no budget set; paid provisioning is refused
  • 1/100 ¢resolution of spend, counted per second
  • 5 minprice cache, which can list cheaper GPU types under a ceiling

Rented GPUs behind a compute trait

FAFO Core and Function Forge needed to burst onto rented GPUs. RunPod is one implementation of a provider-neutral compute trait, next to local Docker and remote Docker on LAN hosts over mutual TLS, so location is configuration and one function binary runs in any of the three.

Local Docker
never budget-limited
LAN host
mutual TLS
Rented spot pod
typed GraphQL client
The outlined cell is the injected runtime binary. The provider client adds typed errors, spot and on-demand modes, GPU-type priority lists with datacenter and CUDA filters, and network volumes.

A daily cap only a person sets

Paid provisioning passes a fail-closed daily dollar cap. Only a person can set it, through a command-line tool I kept off the MCP surface.

One day's spend against the cap
Warnings fire at each mark. Spend is counted in Redis first and in PostgreSQL for durability, and at teardown each container reports its final cost.

I let agents drive cloud GPU capacity without letting them spend money nobody approved, by putting the budget in the critical path.

Inference capacity on demand

Inference Fabric uses the same provider to add inference capacity. One call starts a TensorRT-LLM consumer node on an interruptible spot GPU, reaching the platform through outbound connections only, and the node pulls work like a local card. See A message broker that understands tokens.

Local card
Local card
Rented spot node, outbound connections only
An engine that does not declare its token budget

Fill illustrative. The provisioner reads the engine's configuration, derives batch limits and mounts from it and publishes a capability entry for the batch planners; an engine that does not declare its limits never gets a lane.

Datacentre Blackwell through the same path

The GPU type is a priority list passed straight to the provider, so the same call can ask for consumer cards or datacentre Blackwell parts, with a CUDA-version filter so the pod matches the CUDA 13 toolchain my engines need. The provider's hourly rate, GPU and disk together, is stored at deployment and spend accrues from power-on. The daily cap governs new provisioning; pods already up are stopped by draining and terminating them.

Provisioning path
Budget gate and spend accounting
Broker and its four-number contract
Hardware check, widened
Engine tier for the part
Kernels compiled for SM 100
Datacentre Blackwell part

RunPod's platform and API are theirs; the client, the runtime injection, the budget gate and the spend accounting are mine.