How to invest in AI data centers, honestly.

A guide for people sizing up the AI infrastructure boom and trying to figure out where they fit. We will not tell you what it returns. We will tell you what the business actually is, where money is made and where it gets lost, the realistic options across different capital sizes, and what we think the next few years look like. Written by the people building underneath it, no returns promised on this page, just the mechanics.

The landscape, in numbers

The five biggest US cloud and AI players will collectively spend somewhere between $660 billion and $690 billion on capex in 2026, close to double the year before, pushing the wider industry into its first trillion-dollar year of compute build-out. Single campuses are now spec'd at five gigawatts. The largest model labs are signing multi-year, multi-gigawatt infrastructure deals, the kind of commitment that, ten years ago, only the US Department of Defense made.

The sceptics are loud too. The head of a $2.1 trillion sovereign wealth fund recently warned that an AI bubble could erase a third of its value. A lot of the build is financed off balance sheet through capacity vehicles. Some of it is circular: hyperscaler A funds model lab B that pre-leases capacity from hyperscaler A. We have written about this in detail; the short version is that even past bubbles destroyed the financiers, not the underlying assets. The tracks, the fibre, the grid survived their respective manias and quietly powered the next era.

So the useful question is not whether there is froth. It is whether your position is froth or foundation. Two attributes have always survived a correction: secured demand (capacity sold before it is built) and a low cost floor (power so cheap you are the last operator still profitable when prices fall). Hold both and you do well in either a boom or a bust. Hold neither and the cycle decides your fate.

What an AI data center actually is, as a business

Strip away the AI mystique and a data center is a piece of specialised real estate. You own (or lease) land. You build a shell with serious electrical and mechanical infrastructure inside it. You sell that shell, or the compute that runs inside it, to someone else. The mechanics are closer to a power plant or a self-storage business than to a tech startup. The math rewards patience and discipline more than cleverness.

The capital stack, roughly, breaks down into four pieces:

• Land and shell. Buying or leasing the site, building the structure, securing permits and grid interconnect. Around 30 to 40 percent of total project capex on a greenfield. The grid interconnect is often the slowest piece, not the construction.
• Electrical and cooling. Transformers, switchgear, UPS, backup generation, then liquid cooling distribution units, dry coolers, leak detection. Roughly 30 to 40 percent. For AI-density work (100 kW per rack and up), liquid cooling is no longer optional, it is the only way the math closes.
• Networking and security. Fibre, transit, perimeter, biometrics, surveillance. Around 5 to 10 percent. Small in dollars, large in dependency.
• IT load (the GPUs themselves). If the operator owns the compute, this can easily be the largest line. A single NVL72 rack of Blackwell GPUs runs $3 to $4 million. Many AI colocation operators avoid this line entirely by letting tenants bring their own hardware, which is the model Liwa runs.

On the income side, operating costs are dominated by one thing: power. Sixty to seventy-five percent of operating cost in a well-run AI hall is electricity. Everything else (maintenance, payroll, security, financing) is secondary. This is why people in this business talk about power deals the way oil traders talk about crude. It is the commodity that decides whether the business is profitable.

And on the revenue side, there are two layers, and it matters which one you are buying into:

• Wholesale colocation. Own the shell and the power, lease it by the megawatt to tenants on 5 to 15 year contracts. Predictable, capital-intensive, real-asset returns. This is what most data center REITs (Equinix, Digital Realty) do, and what Liwa does at the operator layer.
• Retail GPU hosting and AI compute. Own the GPUs too, sell compute by the hour or the token. Higher revenue per kilowatt, but exposed to GPU price cycles, utilization risk, and customer concentration. CoreWeave, Lambda, and a growing wave of "neoclouds" sit here. The DesertGrid API venture we run is a layer above this again.

The customers, in order of size: hyperscalers taking gigawatt-class wholesale contracts; neoclouds aggregating GPUs and reselling; AI labs (Anthropic, OpenAI, xAI) taking direct multi-year capacity; enterprises and sovereigns for private inference and training; and the long tail of small builders and researchers paying by the hour.

Where money is made (and lost)

Margins in this business have three structural sources and one fragile one.

The structural sources. First, the arbitrage between cheap power and dense rack space: if your fully loaded power cost sits below the prevailing wholesale colocation rent expressed per kilowatt, the spread is yours, structurally, for the life of the lease. Second, pre-leased capacity: anchor tenants signing 10 to 15 year deals before the slab is poured de-risk the project enough that financing terms improve, which improves equity returns. Third, operational excellence: PUE (power usage effectiveness), uptime, and density. A facility running at PUE 1.2 with 150 kW per rack will outearn a facility running at PUE 1.6 with 30 kW per rack, on the same lease, every month, for decades.

The fragile one is leverage on top of speculative capacity: off-balance-sheet structures that pre-finance gigawatts of build against demand that has not yet materialised. They look brilliant in a year when AI capex doubles. They look like the 2001 telecoms fibre glut in a year when it does not. More on that here.

Where money gets lost, in our experience and observation, follows a few patterns:

• Wrong location. Building in a market with expensive industrial power locks an uncompetitive cost base for the life of the asset. There is no operational improvement that fixes an extra $0.04 per kilowatt-hour.
• Wrong density. Designing for 20 kW per rack on a build that comes online in 2027 is building obsolescence into the slab. The current generation of accelerators already pushes 130 kW per rack; the next generation goes higher.
• Wrong cooling. Air-cooled halls cannot serve the current frontier, full stop. Operators retrofitting air to liquid spend a fortune doing it later instead of cheaply at the start.
• Wrong contract horizon. Signing a 5-year tenant lease on hardware that obsoletes in 3 leaves the operator carrying GPU depreciation risk the customer should bear.
• Speculative build. Pouring capital into capacity with no pre-lease, on the hope that demand turns up. In a soft cycle the bank takes the building.

The asymmetric insight is that cheap power compounds quietly. Every dollar per megawatt-hour below the market rate flows to your bottom line on every token the tenant serves, every minute of every year, for fifteen to twenty years. It is not a flashy advantage. It is a permanent one. Anyone telling you otherwise is selling you a story that depends on staying expensive.

Your options, by capital and expertise

There is no single "AI data center investment." There is a spectrum, from passive public market exposure to building your own campus. Each tier has different liquidity, capital requirements, expertise demands and risk profiles. We have grouped them roughly by entry capital so you can find your row.

The plain mapping: if you have under $100k and want exposure without operating risk, go to row A. If you have time and curiosity as well as capital, row B is where you learn the most per dollar. If you are a serious operator or want to be one, row C is where the actual spread lives. Rows D and E are for people who have either already done rows B and C, or have partners who have. Skipping straight to E is how greenfield projects become cautionary tales.

The Gulf thesis, briefly

Most of the AI infrastructure money over the next decade will sit in three places: the US, the Gulf, and a handful of European and Asian hubs. The US is the obvious centre of gravity for capital and demand. We think the Gulf is the most under-appreciated structural opportunity, and not by a small margin. A few reasons.

Power is the bottleneck, and the Gulf has it in abundance. UAE free-zone tariffs sit at the low end of any major-economy benchmark; Liwa serves at $0.10 per kilowatt-hour, roughly 30 percent below typical European hubs. Sovereign capital in the region is allocating tens of billions specifically to AI infrastructure (Stargate UAE, Humain, MGX), which means anchor tenancy and policy alignment for new operators. Geography matters: the Gulf is within low-latency distance of Africa, South Asia and Southern Europe, three of the largest populations of next-billion AI users. And the regulatory environment is sympathetic to private and sovereign workloads (PDPL data residency, free-zone tax treatment, fast permits) in a way many other jurisdictions are not.

Where Liwa fits. We are a white-label, liquid-cooled colocation operator in a UAE free zone. Capacity at $0.10/kWh, up to 150 kW per rack, pre-sold via reservations held in escrow. Operators and investors can take space under their own brand and resell to their own customers. We run a separate venture, DesertGrid, on top, which turns the same infrastructure into an OpenAI-compatible API for builders who do not want to think about colocation at all. The two together let a partner choose the layer of the stack that fits their capital, expertise and time horizon.

The math, without the pitch deck

You will notice we have not put a single return number on this page. That is deliberate. The internet is full of pitch decks promising specific IRRs on AI infrastructure investments. Most of them are guessing, and the rest are selling you something. Returns in this business depend on power cost, lease terms, density, financing structure and tenant credit, and a serious diligence exercise needs all five inputs to your specific deal.

What we will say, on the public record:

• The biggest line item is power. Get that cheap and most other things forgive.
• Pre-lease before you spec. Build what is already paid for in escrow, not what you hope to fill.
• Match the GPU cycle to the lease length. Do not sign 5-year tenant deals on hardware that obsoletes in 3.
• Liquid cooling is mandatory above 50 kW per rack. It is not an optional upgrade.
• PUE matters quietly but enormously. Halving non-IT power consumption is worth more than any single negotiated discount.

Specific numbers (capex per megawatt, lease pricing, payback under various scenarios, downside cases) live in a private brief we share after a short conversation. If you are seriously evaluating an entry point, request it below.

A reading list, organised by question

We write everything else we know about this business on the Insights blog. If you are running diligence on a specific angle, these are the places to start:

Common questions, straight answers