Follow the line, from the ground to the GPU.
Seam deploys private GPU infrastructure on natural-gas wells it owns. This is the whole path - how the gas is sourced, how it becomes power, where power becomes compute, and what a working site actually is. The compute itself stays dedicated and private; the rest, we will walk you through.
Time to power, not silicon.
The binding constraint in AI compute is not chips or capital - it is time to power. Grid interconnection queues run three to four years across ERCOT, MISO and PJM, and a handful of hyperscalers hold the front of the line. Generating power on-site skips the queue entirely.
Marginal gas, owned - not borrowed.
Seam acquires marginal-producing gas leases outright - multi-well pads where the gas reaches market only through a long, tolled route, earns too little at the wellhead to justify the operator's attention, and the operator wants it off their books. Owning the wellbore rather than renting flare gas makes the gas a primary, fixed-price commodity for the life of the lease. That ownership is the moat.
- Multi-well pads · 5-15 bores
- Known production decline curve
- P&A liability assumed at close
- ≈ $1 / MCF, contracted for term
Power, made at the wellhead.
Power is generated on-site by reciprocating gas engines - chosen over turbines for two reasons: they are available in months rather than on a 24-30 month backlog, and they are sized right for the hundreds-of-kW-to-low-MW volume a marginal lease yields. SLA-bearing sites run N+1, so any one engine can drop for service without dropping the cluster.
- 4 active + 1 standby × 325 kW gas gensets
- ≈ 1,235 kW gross → ~852 kW IT at 1.45 PUE
- Runs on field gas, minimal treatment
- Reciprocating, not turbine
Where power becomes compute.
This is the seam. Because the gas is contracted at a fixed price, the cost per kWh is fixed before the site ever turns on - raw energy in, precise dedicated compute out. The same line that carried power now carries compute.
Dedicated, and yours.
What the customer gets is dedicated and private: single-tenant, air-isolated facilities - no shared tenancy, no neighbor on the rack, full physical custody. Managed H100 clusters, or bring your own hardware; Seam is vendor-agnostic on silicon and earns power-and-uptime margin only. How the cluster is built is ours to operate - what it is, is yours.
- Single-tenant · air-isolated
- Managed H100 or BYOG
- Vendor-agnostic silicon
- Physical custody by default
And then you just address it.
Operating the wells, the gensets and the fiber is ours. What you touch is an endpoint - a base URL, a CLI, a cluster you address like any other. Hit managed H100 over an OpenAI-compatible API, or bring your own stack: a kubeconfig for your nodes, or the whole dark-fiber fabric as one logical cluster.
- OpenAI-compatible endpoint
- Kubernetes · Slurm · Terraform
- SSH · S3-compatible storage
- Dedicated tenancy · single-tenant
What a site actually is.
A Seam site is small by design - a ~1.3 MW off-grid build at the wells, not a hyperscaler campus. Because moving data is far cheaper than moving gas, we build compute where the gas already is and stitch sites together over dark fiber in a ring, so a handful of pads read as one logical cluster. Inter-site networking adds only ~5% to CAPEX. Closed-loop cooling, no potable water at scale, and a designed 99.5% SLA.
- ~640 H100s / site · ~3,200 per 5-site cluster
- Dark-fiber ring → one logical cluster
- 99.5% designed SLA · <4 hr/month unplanned
- Closed-loop / dry cooling
Designed for 16 weeks from close to compute.
The 16-week target runs from acquisition close to first revenue compute. The schedule is engineered around brownfield reuse - existing pads, tank batteries and road access carry a fraction of greenfield cost.
From a dollar of gas to a GPU-hour.
The all-in price works backwards from gas contracted at the wellhead. Because the input is fixed, so is the output - power that does not move with spot gas, grid tariffs, or capacity-market shocks. Indicative figures, not commitments.
Available, and priced predictably.
Seam competes for a specific dollar at the intersection of hyperscaler rentals, neocloud clusters and gas-fueled colocation. None is a pure substitute - honest comparison is more useful than blanket claims of cheapness.
The methane math, plainly.
Leaked methane has roughly 80x the 20-year warming impact of CO2. Combusting gas that would otherwise leak from a marginal well is materially better than venting it - not carbon-negative, not renewable, and we will not claim either. Seam's position is to publish per-site measurement rather than lean on industry-average emission factors.