AI Infrastructure
If you have capital and you're looking at data centers as an asset, the hard part is never the servers — it's power, land and time. We help you decide what to build, find and secure the site, get it connected to transmission, build it modularly and hand it over running.
For investors and owners
Most people who want to own compute infrastructure stall at the same place: which configuration actually pays back, where to put it, and who is accountable when the grid study comes back at 30 months. We take that whole chain. Below are the three structures we usually work in — the right one depends on how much of the asset you want to own and how much operating risk you want to carry.
We design and build the facility to your spec and hand over the keys. You hold the land, the asset and the depreciation.
We build against a signed offtake — an AI company, a cloud operator or an enterprise — so the facility is contracted before it's energized.
You bring land, power position or capital; we bring engineering, supply chain and delivery. We build and operate together.
Tell us the budget, the geography and the time horizon. We come back with a straight answer on what that capital actually buys in megawatts, how long it would take, and where the money goes.
No deck, no pitch. If the numbers don't work, we say so on the call.
How it works
Nothing here is exotic. It's a long chain of unglamorous work where the expensive mistakes happen early — buying land whose interconnection queue is six years deep, or sizing cooling for a density the racks will exceed in eighteen months.
Before any land is bought, we size the thing: capacity, workload profile, density per rack, redundancy level, and what the total cost of ownership looks like over ten years.
Training clusters, inference fleets and enterprise HPC want different buildings. Getting this wrong is the most expensive error available.
We screen land against the criteria that actually gate a data center — available transmission capacity first, then water, fiber, zoning, climate and tax regime.
Most parcels that look perfect on a map die on the interconnection study. We check that before you make an offer.
Purchase or long-term ground lease, plus the paperwork that makes the megawatts real: interconnection agreements, facility extension agreements, substation capacity.
A site without a signed power position is farmland with a good story.
Substation, high-voltage interconnect, transformers and switchgear — engineered, procured and coordinated with the utility, in phases if the load ramps.
This is usually the longest pole in the tent, which is why we start it in parallel with design rather than after it.
Detailed design of power, cooling and network, then factory production of the modules while site works and foundations proceed in parallel.
Building the box and preparing the ground at the same time is where most of the schedule saving comes from.
Delivery, assembly, integration, testing and handover — power on, cooling balanced, network live, load proven.
Pre-configured modules arrive tested, so commissioning is verification rather than discovery.
Run it, or have it run: colocation, hosting, GPU-as-a-service, reserved clusters, or wholesale capacity to a single large operator.
A megawatt sold as electricity has one value. The same megawatt sold as AI compute has a considerably higher one.
Anatomy
Strip away the marketing and an AI data center is four systems that all have to be sized together: something to bring power in, something to take heat out, something to move data, and a structure to hold it. Get one wrong and the others are wasted.
The gating constraint on every project. Everything upstream of the rack, from the transmission line to the busbar.
AI racks run at densities that air alone can no longer handle. Cooling design follows the chip, not tradition.
The part everyone pictures — and the part that is easiest to change later.
The unglamorous half of the CAPEX, and the half that determines the schedule.
Why modular
Traditional data center construction was designed for a market where demand was predictable and you could plan three years out. AI demand does not behave that way. Modular construction moves most of the build into a factory and runs it in parallel with site preparation.
| Traditional build | Modular build | |
|---|---|---|
| Time to first megawatt | 24–36 months | 3–6 months |
| CAPEX | Rising 30–50% on materials and labor | 20–30% lower via factory production |
| Land footprint | 700+ acres per gigawatt | Dense vertical and stacked layouts |
| Scaling | Fixed at design, expensive to extend | Phased from 10 to 100+ MW |
| AI readiness | Retrofit needed for GPU density | Liquid cooling designed in from the start |
| Renewables | Bolted on, if at all | Solar, storage and heat recovery integrated |
| Supply chain | Local contractors, local prices | Established international sourcing |
Modular is not automatically the right answer. For very large single-tenant campuses on cheap land, conventional construction can still win. We tell you which case you're in before you commit capital.
How it gets built
This is what the schedule actually looks like on site. The trick isn't working faster — it's that the modules are being built in a factory the whole time the ground is being prepared, so two of the longest jobs happen at once instead of one after the other.
Stage 1Access roads, drainage, survey, compaction. Unglamorous and easy to underestimate — this is where a cheap parcel starts costing money.
Stage 2Concrete pads in the module grid, conduit and earthing cast in, water and fiber brought to the boundary. Module fabrication is already underway in the factory.
Stage 3Modules arrive tested and are craned onto their pads, then coupled: power, cooling loops, network. A row goes down in days rather than months.
Stage 4Commissioning, cooling balanced, redundancy tested, load ramped. Because the modules were tested in the factory, this stage is verification rather than discovery.
Stage durations depend on the site: grading and foundations move fast on flat, dry ground and slowly on anything else. Grid works run in parallel throughout and are usually what sets the finish date.
Sites and projects
A representative selection of the sites in our pipeline — one in advanced diligence, three under evaluation. They give a sense of what a viable data center site looks like in practice: power position first, everything else second.
In diligenceEast Texas, United States
A large parcel with an existing private substation and an on-site high-voltage grid interconnect. The regional transmission utility has approved a facility extension for 70 MW, delivered in two phases as the substation expands. Natural gas runs a few miles out, there are five water wells on the property, and fiber is being installed now.
Key parameters
Under evaluationWest Texas energy corridor, United States
A development-ready site with a direct grid connection and substation, switchgear and transformer infrastructure already installed. Structured for step-down and containerized deployment, which means megawatts can be brought online in increments rather than all at once.
Key parameters
Under evaluationMidwest United States, MISO grid
A fully energized multi-site portfolio — three co-located facilities within an hour of each other — with an operations team already in place and a track record above 99% uptime. Substation capacity for a further 20 MW is already approved, which makes it a candidate for redeployment to AI and HPC loads.
Key parameters
Under evaluationStockholm, Sweden
Conversion of an existing steel-frame industrial building into a vertical high-density compute facility, aimed at inference and fine-tuning workloads in a supply-constrained in-city power market. Industrial zoning and building permits are already in place, which removes the slowest part of a European build.
Key parameters
Counterparty names, exact addresses and commercial terms are withheld — these materials are covered by confidentiality agreements. Status reflects where each site stands today, not a claim of ownership.
On the ground
Photographs from site diligence and live deployments. This is what most of the work actually looks like — substations, switchgear rooms and steel boxes on gravel, long before anything resembles a server room.
Site selection
We screen sites in this order. A site that fails the first test is not a data center site, no matter how good the price is.
Not proximity to a power line — actual deliverable megawatts, with a utility willing to commit to them on a timeline. This single factor eliminates most candidate sites.
A queue position measured in years turns a good site into a bad investment. We check the study status before anyone signs anything.
Cool climates cut cooling OPEX for the life of the asset. Water rights matter wherever evaporative or hybrid cooling is on the table.
Diverse routes and real capacity. Training workloads move data volumes that make a single provider a single point of failure.
Correct zoning, buildable ground, room to expand, and a permitting authority that has done this before.
Investment tax credits, sales tax exemptions on equipment and local abatements can move the returns more than a better price on the land.
Who does the work
Data center development is not one discipline. It's engineering, procurement, energy, construction, project finance and law, and it fails at the seams between them. We assemble the whole chain under one point of accountability.
Specialists with delivery experience on US data center projects, including facilities in Texas — electrical, mechanical, cooling and modular integration.
Established international sourcing for the long-lead items that stall projects: transformers, switchgear, cooling plant, batteries and network hardware.
Interconnection strategy, utility negotiation, on-site generation and storage, and the tariff structures that determine whether the economics work.
Structuring the holding and project companies, modelling the returns and bringing in institutional capital alongside yours.
International corporate structuring, land and power contracts, offtake agreements and the regulatory process in the host jurisdiction.
Our own discipline — the operations, monitoring and AI systems that run on top of the facility once it's live. See our AI engineering work for what that looks like.
Where we work
Power is local. The right jurisdiction depends on what you're optimizing for — cheap electricity, speed to market, tax treatment or proximity to your customers.
Texas and MISO — deep transmission capacity, 30% investment tax credit on integrated solar and storage
Cool climate, clean grids and in-city power positions for latency-sensitive inference
Hydropower economics and sovereign AI programmes, with India as the logistics and equipment hub
Government-backed sovereign data center initiatives opening genuinely new markets
FAQ
There is no honest single number, and anyone who gives you one on a website is guessing. The cost is driven by land price, how much grid work the site needs, rack density and therefore cooling architecture, whether you integrate generation and storage, equipment lead times and local labour. We price it properly once we know the site — a bad site can cost more in grid works than the entire IT load.
Modular deployment runs 3–6 months from a prepared site, against 24–36 months for conventional construction. The honest caveat is that securing land and a firm power position usually takes longer than building the facility, which is why we start the grid conversation first.
A conversation. Concretely: roughly what you want to invest, whether you want to own the asset or contract it out, whether you have a target geography or an existing site, and what your time horizon is. From there we can tell you fairly quickly whether the project makes sense.
No, but it changes the risk profile substantially. Build-to-suit against a signed offtake finances far more easily than speculative capacity. If you'd rather not carry the leasing risk, we structure it that way from the start.
Whoever the structure says. In a turnkey build you own the land and the asset outright. In a joint venture, ownership splits according to what each side contributes — land, power position, capital or delivery. We set this out before any money moves.
We build and deliver the facility, and we work with the supply chain for the compute hardware. Whether GPUs sit in the same contract depends on your model — some owners lease shell and power, others want the whole stack including the compute revenue.
Optional. Some clients take the keys and run it themselves, others want us on a long-term operations contract. Since we also build the AI and automation software, we can take on the monitoring and operations layer as well.