Start with the building.
Size capacity to the space, power, and cooling available. Grow through distributed sites.
A new geography of compute
AI capacity.
Behind the meter.
The next generation of AI infrastructure is already powered.
It’s in the buildings around us.
Local power. Distributed compute. One fabric.Follow a prompt through Marin
01 / The opportunity
The power is there.
Let’s put it to work.
Commercial buildings hold a resource AI needs: access to power. Marin’s model brings modular GPU infrastructure to those sites and connects them into one network.
Connect the capacity.
The Marin fabric coordinates independent locations, routes inference, and manages the fleet.
Deliver one service.
Applications get a familiar API. Building owners get a new use for existing infrastructure.
02 / The fabric
Many locations.
One place to build.
A coordination layer turns independent GPU sites into a single inference endpoint. Each site runs complete models. Marin chooses a healthy replica and streams the response back.
Your app→API & access→Marin router→GPU replica
One request goes to one complete model. Response tokens stream back through Marin to your application.
Illustrative architecture · not live traffic
Prompt Response
01 / Your application
A familiar API. A new kind of network.
Your application or agent sends a prompt to Marin using an OpenAI- or Anthropic-compatible API.
PROMPT / EXAMPLE
“What can we build together?”Whole models, local execution.
Models are replicated across sites. A single generation stays within its selected site.
Routing with context.
Health, load, geography, and conversation affinity help requests find the right replica.
Managed as one fleet.
Monitor sites, manage model availability, and apply signed updates from a central control plane.
03 / For builders
Familiar API.
New possibilities.
Bring your application, SDK, or agent harness. Marin exposes OpenAI-compatible chat completions and Anthropic-compatible messages, with streaming and tool calling.
Use native conversation and cache signals where available. The fabric handles placement behind the endpoint.
Talk to us about inference accessBuilt forApplicationsAgentsOpen models
curl https://fabric.marin.ltd/v1/chat/completions \
-H "Authorization: Bearer $MARIN_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen/Qwen3.5-4B-Fast",
"messages": [{
"role": "user",
"content": "What can we build together?"
}],
"max_tokens": 512,
"stream": true
}'For the agents building next
USDC / Solana / x402Agents that can build.
And pay their own way.
Agentic stablecoin payments, built into Marin. Fund prepaid inference credits with USDC on Solana through x402 over HTTP. Your agents, your wallet infrastructure, the same inference APIs.
- 01 / Discover
Find capacity and pricing.
Agent-first documentation and machine-readable model and price information.
- 02 / Fund
Authorize a bounded top-up.
An x402-capable client pays with USDC. Confirmed settlement funds inference credit.
- 03 / Build
Use familiar inference APIs.
Scoped credentials, usage limits, and receipts—with prepaid credit covering subsequent requests.
04 / For property owners
Your building.
A new source of value.
Have space and power available? Marin’s hosting model turns them into productive AI capacity.
You provide the site and access to power. Marin’s model covers the GPUs, installation, electricity, and operations—with host compensation structured around a guaranteed floor and revenue participation.
Explore hostingSite suitability and commercial terms are evaluated together.
A closer look
Built on a
different foundation.
What makes a distributed site part of the fabric?
A local agent reports model readiness and machine health, and maintains an encrypted outbound connection to Marin. The control plane uses that connection to route inference and stream results. A separate management channel coordinates the fleet.
Does a model run across multiple buildings?
No. Each eligible site holds a complete model replica. A generation runs at one site; the fabric distributes requests between replicas. This avoids making model execution dependent on communication between buildings.
What happens if a site goes offline?
Marin stops routing new requests to unavailable replicas. A request that cannot establish a connection may retry another eligible site, subject to placement policy. An interrupted generation must be retried by the client; it cannot move between sites mid-stream.
Do I need a Marin-specific SDK?
No. Use a compatible OpenAI or Anthropic client with Marin’s endpoint and an issued API key. Streaming and tool calls are supported. Model availability and supported features depend on the selected runtime.
What kind of building can host Marin?
We’re interested in commercial sites with available power, suitable space, and operational access. The right installation depends on electrical capacity, cooling, connectivity, and the building’s requirements. Contact us to evaluate your site.
Let’s build the next layer.
More places.
More possibilities.
Building with AI? Have power to put to work?
We’d like to hear from you.