Own the GPUs
Fleet operators renting their own accelerators, and research shops selling their spare cycles, hit the same wall. A fleet is only worth what you can keep busy, and keeping it busy is a power problem long before it is a demand problem.
Crossbridge Power builds behind-the-meter generation at industrial sites the hyperscalers skip — 1 to 20 MW per location, energized in months instead of interconnection years, on the same operating standard that scales to a 200 MW campus. On bridge fuels we refine ourselves, with the engineered step down to gas and grid from day one. We sell power, land and shells to the neoclouds; we do not compete with them for tokens.
Six different businesses are converging on one operating model. Rent out your own GPUs. Start next to power and build the data center. Aggregate other people's capacity. Do research and sell the spare cycles. Build your own silicon and sell inference on it. Route inference across everyone else's chips. Different starting points, one destination — and one binding constraint.
Compute gets cheaper every quarter. Energized industrial land does not. Crossbridge sits upstream of all six paths: we hold the site, the fuel and the generation, and we contract it to whoever is turning electrons into tokens.
Fleet operators renting their own accelerators, and research shops selling their spare cycles, hit the same wall. A fleet is only worth what you can keep busy, and keeping it busy is a power problem long before it is a demand problem.
Operators who began in generation or mining already know the answer. Their scarcity is not silicon — it is the next site with fuel, a permit and a customer. That site is the product we sell.
Aggregators and cross-silicon routers need many small pools in many places, not one gigawatt campus. A fabric of 1–20 MW sites is the supply shape they cannot buy from a hyperscale developer.
A 100 MW minimum order size is a hyperscaler artifact. The fastest-growing inference workloads are batch, background and agentic — they will take a megawatt at a time, in any jurisdiction, at 95% availability, if the price per MWh is right.
Compute load is the hardest version of the problem we already solve: it needs power before the utility date, and it cannot be left on rental diesel afterwards. The bridge model handles the fuel path; the availability classes handle the workload path. Together they let you place batch work on the cheap rung today and interactive serving on firm gas capacity later, without re-tendering.
| Rung | Fuel | What it delivers for compute | Typical availability class |
|---|---|---|---|
| 01 · Deploy | Renewable diesel / HVO | Load carried in weeks, at diesel speed, without committing the site to diesel | Flex ~95% |
| 02 · Displace | HVO, blends, LPG | Emissions and permit exposure drop without new iron, so run hours open up | Flex / Balanced |
| 03 · Convert | CNG / RNG / pipeline gas | Reciprocating gas prime power at ~1.9 MW granularity; the fabric block proper | Balanced ~99% / Firm 99.9%+ |
| 04 · Integrate | Gas, RNG, hydrogen-ready | Permanent microgrid, N+1, grid-parallel where it pays, heat recovered where there is a load | Firm / Campus N+1 |
A pure power developer buys its fuel from someone else. When a spec question or a supply squeeze arrives, that is a procurement cycle. For us it is an internal decision.
You are never re-tendering and never holding an asset that has run out of purpose. The bridge and the destination are the same contract and the same operator.
Clean bridge fuel keeps the air permit and your customer's reporting workable from the first hour, instead of becoming the thing you renegotiate in month four.
Inference platforms already price latency in tiers — interactive, background, best-effort. Power should be sold the same way. A background agent running for six hours does not need the availability of a checkout page and should not pay for it. Every Crossbridge block is contracted against a guaranteed number we can evidence, with an engineered path to the class above it.
Downtime budget is the arithmetic complement of the availability figure over 8,760 hours, drawn on a log scale. Class is assigned per block and per contract: it is what we guarantee, not the best case we have observed.
| Class | Guaranteed availability | Downtime budget | Redundancy | Workload fit | Product |
|---|---|---|---|---|---|
| FLEX | ~95% | 438 h/yr | N, single-train blocks | Batch inference, evals, RL rollouts, fine-tunes, checkpointed training | Fabric |
| BALANCED | ~99% | 87.6 h/yr | N, hot spare units on site | Background and agentic pipelines, subagents, long-horizon tasks | Fabric |
| FIRM | 99.9%+ | 8.8 h/yr | N+1 on the generation train | Interactive serving, control plane, storage and network tiers | Fabric & Campus |
| CAMPUS N+1 | 99.99%+ | <1 h/yr | N+1 minimum, phased blocks, BESS/UPS integration | Contracted campus load with liquidated-damage schedules | Campus |
Platforms built for background work already move jobs off a failed pool and fail over to more reliable compute when cheap capacity disappears. If your scheduler absorbs a site going down, you should not be paying a five-nines power premium to prevent it.
Lean-burn reciprocating gas engines in roughly 1.9 MW increments give granular maintenance, low forced-outage impact per unit, fast start and a heat rate that holds at part load. Adding one unit moves a block from Flex to Firm without redesigning the plant.
Availability and capacity are measured as delivered load at the POD, at design summer ambient, after parasitic load and step-up losses — never nameplate. We would rather sign a lower number we can prove than a headline number we cannot.
Most developers pick one. Micro sites are too small for the people who can finance campuses, and campus discipline is too heavy for the people who move fast. We run both off the same fleet, the same fuel desk and the same operating team.
| Crossbridge Fabric | Crossbridge Campus | |
|---|---|---|
| Block size | 1–20 MW per site | 100–200 MW net, phased |
| Capacity basis | Net delivered at the POD, design summer ambient | Net delivered at the POD after parasitics and step-up losses |
| Generation | Lean-burn reciprocating gas, ~1.9 MW units, redeployable | Large reciprocating fleet, N+1, many smaller units over big frames |
| Heat rate design | ≤ 8,500 Btu/kWh HHV | ≤ 8,500 Btu/kWh HHV at full load, site conditions |
| Air permitting | Minor-source path targeted where the site allows | Major-source PSD where triggered, plus public notice |
| Time to first power | Months from notice to proceed, equipment-led | Phased blocks: first power, then substantial completion |
| Availability class | Flex / Balanced / Firm | Firm / Campus N+1 |
| Commercial form | Power tariff or dry lease; short and medium tenor | Capacity charge ($/kW-month) plus energy charge ($/MWh), fuel passed through |
| Ownership | Crossbridge owns and operates; residual and relocation rights retained | Build-own-operate preferred over lump-sum turnkey EPC |
| Downside case | Units redeploy to another site or back to field service | Site tenure, credit support, partial-COD capacity payments |
Small blocks stay under the thresholds that turn a project into a multi-year regulatory program, and they sit on land that is already industrial.
At campus scale the customer is buying an annuity with a credit story: firm capacity, a measured heat rate, redundancy, liquidated damages and a payment schedule.
Indicative engineering arithmetic, not a quotation. Gas burn is shown at the design heat rate and energy at a 90% capacity factor. Rack count uses a commonly cited reference of roughly eight liquid-cooled racks per megawatt.
Drag to see how the product, the permitting path and the generation fleet change with size.
Every site in the fabric sits inside an operating industrial position — a refinery, a gas pad, a former terminal — where fuel, yard, water, road access and workforce already exist.
Generation, heat recovery and fuel production on one footprint. Data hall provisioning is the gating item for compute load — not power.
Sited for gas access, data center demand and the fuel logistics corridor. First block energizing on new reciprocating gas gensets.
Industrial-zoned land with city water and sewer and interstate access. Gas connectivity, parcel work and permitting are in development; no capacity is offered yet.
Crossbridge does not sell GPUs, tokens or managed workloads, and it does not want your customers. The proposition is upstream: controlled land, fuel reliability, bridge and permanent generation, site development, and a contract structure a lender will accept.
Capacity in the places and sizes the majors will not build, with an availability class matched to each pool in your fleet.
Background, batch and agentic serving where cost per token dominates and the scheduler already tolerates a site going away.
Land, power and shell delivered as one package, phased to the arrival of your hardware rather than a developer's pro forma.
The original business: microgrids, bridge power and fuel logistics for operators who cannot wait for a utility interconnection.
Send the megawatts, the jurisdiction, the required energization date and the availability class your workload actually needs. We come back with a block, a delivery model and the permitting path that governs it. If we do not have a site that fits, we will say so.