Marin.01 / 15
August 2026

Marin.

AI capacity behind the meter.

Scroll

The Problem

The next unit of AI supply must be distributed.

  • Demand for AI inference is exploding
  • Datacenter construction is gated by power, politics, and siting, not by GPUs
  • Gigawatt campuses take years, and have limited location options
  • Nearly all new datacenter supply is already accounted for
  • Buyers are increasingly latency-sensitive about where inference runs

“I don’t have warm shells to plug into.”

Satya Nadella · microsoft ceo

What is Marin?

We turn underused commercial power into AI infrastructure.

  • Modular data centers deployed behind the meter in commercial buildings
  • An orchestration fabric turns many small sites into one schedulable network
  • We buy the GPUs, fund the install, pay the power, run the site
  • Reserved capacity contracted to labs, clouds, and enterprises
  • Idle capacity sold into burst markets like OpenRouter

The Hardware

Capacity is sized to the building, not the other way round.

GB300 NVL72 rack NVIDIA NVL72 72 gpus · 48u Eight-GPU B300 node, stackable NVIDIA NVL8 8 gpus · 8u
Cost$3.7–4.0Mplus $300K install$400–500Kplus $40K install
Rev @ 75%$270/hr$2.4M a year$30/hr$263K a year
Rev @ 100%$360/hr$3.2M a year$40/hr$350K a year
Power132–142 kW155 kW peak14 kW18 kW peak
Compute1,080 PFLOPSfp4 dense · 20 TB HBM112 PFLOPSfp4 dense · 2.3 TB HBM
CoolingLiquidAir

how buildings become one product

The Fabric

Reliability is a property of the network, not the building.

routerofficelight industrialtelco colabshaded = one model, replicated per site
  • Replicas, not shards. Every site serves whole models, nothing splits across buildings
  • Routing. Requests reach the nearest warm replica in milliseconds
  • Placement. A slower loop moves models by demand, price, and SLA
  • Failover. If a site drops, traffic reroutes rather than fails
  • One endpoint. Buyers get an API, not forty buildings

The Host Offer

A new NOI line. No capex, no utilization risk, just like cell towers.

  • The host provides space and access to power it already holds
  • Marin funds hardware, install, electricity, and operations
  • A guaranteed $50/kW per month, or 5% of post-power revenue if greater
  • About $125K a year from an NVL72 rack in 200sqft, or $14K from an NVL8 in 20sqft
  • Recurring AI revenue, no GPU exposure, no datacenter staff

GTM

Revenue on day one, contracts after proven performance.

  1. Launch at 5 pilot sites in Los Angeles (already committed, est. 1.25MW)
  2. Negotiate anchor LOIs before hardware ships, finalize once we prove performance
  3. Install 20 racks and prove latency, uptime, thermals, ops
  4. Earn aggregator revenue from day one, not just on idle hours
  5. Repeat city by city

how we fill the fleet

Demand

Supply is scarce. Utilization is a credibility problem, not a demand problem.

  • Routers. Listed on OpenRouter from day one. Traffic arrives based on price and latency at zero acquisition cost
  • Platforms. Fireworks, Together and Baseten buy regional capacity wholesale - one contract, large volume
  • Marketplaces. SF Compute, Vast.ai and others buy capacity on-demand
  • Anchors. Two or three per metro, focusing on latency-bound and regulated workloads
  • Proof. Head-to-head latency and performance benchmarks, not just a deck

per gb300 nvl72 rack · annual

Illustrative Economics

Sell capacity like a utility. Finance it like real estate.

$2.68Mgross · 1.8-yr payback
Customer compute revenue$2.68M
Direct electricity−$193K
Distributable revenue$2.49M
Host · floor or 5%$125K
Marin$2.36M

Why now?

Marin can bring hundreds of mini-datacenters online in weeks, without any politics.

  • Datacenter buildouts are backlogged for 10 years
  • GPU's can be easily bought, gigawatts cannot
  • Inference demand requires more localized processing
  • Anti-datacenter rhetoric is at a fever pitch

Competitive Landscape

SegmentScaleSpeedLocalPrivateSLA
Hyperscalersaws · azure · gcp✓——~✓
GPU cloudscoreweave · lambda✓~—~✓
DePIN computeakash · io.net~✓✓——
Marinbehind the meter~✓✓✓✓

Founding Team

Our team built the largest decentralized physical network in the world. Marin solves the same problem, except supply constrained.

  • Amir Haleemceo

    Founded Helium and built the world’s largest decentralized wireless network, with more than a million distributed nodes, and millions of daily cellphone users.

  • Frank Mongcro

    Founding member and COO of Helium, running sales and business development. Before that, 25 years in security leadership at Palo Alto Networks, Hortonworks, and HP.

  • Charles Kimcfo

    Previously 20 years as EVP at Wells Fargo Commercial Capital. Sourced, syndicated, and purchased nearly $50 billion of senior bank debt.

  • Mark Phillipscdo

    EVP of business development at Helium, where he drove the AT&T, T-Mobile, and Telefónica carrier partnerships and built the off-grid disaster-relief hotspot program.

Ask

A repeatable, financeable metro playbook in 12–18 months.

  • Raise $20M seed to bring up the first metro
  • Secure three anchor LOIs and three to five host partners
  • Deploy 20 nodes across Los Angeles, equity plus equipment financing
  • Live aggregator revenue first, reserved scheduling layered on top
  • Prove uptime, cross-metro routing, margin, and refinanceability
  • Amir Haleem · founder & ceo
  • amir.haleem@gmail.com