The AI Data Center Business Model: GPU-as-a-Service vs. Powered Infrastructure

Every AI data center business model reduces to one of two bets. Sell GPU compute by the hour and carry utilization risk, or lease powered, cooled space to someone else’s hardware and carry real-estate risk instead. The GPU-as-a-service route (the neocloud model) generated over $25 billion in revenue in 2025 and is forecast to approach $400 billion by 2031, but debt-financed clusters only break even around 70% utilization. The powered-shell route trades that volatility for a colocation market growing at roughly 9-19% CAGR through 2030, paid on multi-year leases that collect rent whether or not a single GPU ever spins up.

This post covers the two routes into the AI infrastructure business → the stack both require → where they diverge on software and contracts → the economics and risk each one carries → how to pick the one that fits your balance sheet.

What Is the AI Data Center Business Model, Really?

Most people who say they want to “get into AI data centers” haven’t actually decided which business they’re building. That’s the mistake worth naming before anything else, because the two options sit on opposite sides of the balance sheet.

Route one is GPU-as-a-service, also called the neocloud model. You buy or lease the GPUs. You operate (or contract someone to operate) the facility around them. You sell compute by the GPU-hour to AI labs, enterprises, and hyperscaler overflow demand. You carry utilization risk and GPU depreciation risk, plus the technology-refresh risk of every new Nvidia generation undercutting the last one’s rental price. We’ve written the full mechanics of this model, including the take-or-pay contracts and debt structures that fund it, in how the neocloud business works.

Route two is powered shell, the infrastructure-only play behind traditional colocation. You build or lease the site, the power, the fiber, and the cooling; the tenant fits out the interior and brings their own GPUs, their own servers, their own software stack, a structure explained in detail in dgtlinfra’s powered shell guide. You never touch depreciation on a single chip. Your revenue is rent, typically quoted per kilowatt per month on a long lease, and your risk is that the building sits half-empty, not that a GPU generation goes stale on your balance sheet.

This is the real neocloud vs. colocation split, and it’s a capital-structure decision before it’s an engineering one.

GPUaaS behaves like a leveraged trading business wrapped around hardware that loses value every day it sits dark. Powered shell behaves more like industrial real estate with an unusually demanding tenant improvement spec attached. Nothing about picking one requires being wrong about the other — it requires being honest about which balance sheet you can actually run.

The Infrastructure Stack Both Routes Must Build

Strip away the business model and the physical stack looks identical for the first four layers. Site, power, connectivity, cooling. Every AI infrastructure business, whichever revenue model sits on top, has to solve these in roughly this order.

Site selection starts with power, not land. A parcel with cheap land and no substation nearby is worthless on any AI timeline. Buyers now screen sites by proximity to transmission capacity and by whether an existing interconnection agreement can be assigned, because starting from zero is brutal. Lawrence Berkeley National Laboratory’s Queued Up: 2025 Edition research found that of all generation capacity that entered US interconnection queues between 2000 and 2019, only 13% had reached commercial operation by the end of 2024, with median time from interconnection request to commercial operation now exceeding four years for recent project cohorts. That statistic applies to generation, but the same congestion is now colliding with data center load requests in the same study windows.

Power comes next, and it splits into two paths: wait for the grid, or generate on-site. With interconnection queues now measured in years, a growing share of both neoclouds and powered-shell developers have turned to gas turbines and on-site solar-plus-storage, or to direct deals with independent power producers, rather than waiting in line behind everyone else’s data center. We’ve mapped the mechanics of that shift, including tolling agreements and behind-the-meter generation, in why data centers are done waiting for the grid.

Fiber and connectivity are the layer everyone underrates until it fails. AI clusters need low-latency, high-bandwidth paths to both the internet backbone and, increasingly, to sister facilities for multi-site training or inference failover. Carrier-neutral meet-me rooms and diverse physical fiber routes aren’t optional at this scale; a single-path facility is a single point of failure that no SLA can paper over.

Cooling architecture is where rack density decides the answer. Air handles roughly 5-20 kW per rack. Above that, direct-to-chip liquid cooling becomes structural rather than optional, and dense AI inference racks at 40 kW and up are designed around it from the start. The specific engineering of that transition (CDUs, manifolds, leak detection) is its own discipline, covered in our AI data center build guide.

Every operator, in either model, is solving these same four problems before a single customer signs anything. Past that point, the resemblance ends.

Where GPUaaS and Powered Shell Diverge: Software and Metering

Once the physical stack is up, the two models need almost entirely different software, and this is the layer most first-time operators budget for last, if at all.

A GPUaaS business needs a multi-tenant orchestration layer sitting on top of the hardware. That means Kubernetes or Slurm (or both; CoreWeave built “SUNK,” Slurm-on-Kubernetes, and Nebius built a comparable Slurm operator) to schedule jobs across shared GPU nodes, tenant isolation so one customer’s workload can’t see or starve another’s, role-based access control, and quota management. Billing has to fire at job boundaries, not on a clean hourly clock, because a customer who reserves a node for an hour but only runs a 40-minute job expects to see that reflected. That whole stack sits closer to cloud platform engineering than to facilities management, and most powered-shell operators have nobody on the bench who has ever built it.

A powered-shell business needs a different kind of precision entirely. Power has to be measured at the branch-circuit or PDU level because the lease is priced in dollars per kilowatt, plus a per-kilowatt charge for critical-load draw that commonly runs around $35/kW on top of the base rate. DCIM software tracks capacity commitments against actual draw across dozens of tenants in a colocation hall. SLAs are written around uptime percentage, temperature and humidity bands, and increasingly PUE, since EU tenants now need that number reported regardless of who owns the servers inside. None of this requires a scheduler. All of it requires metering accuracy a court would accept in a lease dispute.

The two software stacks don’t transfer to each other. A team that’s spent years building GPU job schedulers has no particular advantage building lease-administration and metering systems, and vice versa. Pick the model, then hire for the layer it actually needs.

GPU-as-a-Service vs Colocation: Economics and Risk Compared

The numbers make the divergence concrete. The table below compares the two routes on the dimensions that decide whether either one survives a downturn.

Read that table as a risk transfer, not a ranking. GPUaaS operators absorb technology risk and demand risk in exchange for a much larger addressable margin if utilization holds. Powered-shell operators hand both of those off to the tenant and collect a flatter, more bond-like return, in exchange for construction and pre-leasing risk instead. Grand View Research sized the global colocation market at $91.1 billion in 2025, growing to $184.4 billion by 2033 at a 9.3% CAGR, while JLL’s 2026 Global Data Center Outlook puts the broader data center sector at a 14% CAGR through 2030, with colocation the fastest-growing segment inside JLL’s Asia-Pacific breakdown specifically, at 19%. Synergy Research Group tracks the neocloud side, and its numbers show the opposite shape: about $25 billion in revenue in 2025, forecast to approach $400 billion by 2031.

Both things are true at once. The colocation growth holds regardless of whether this quarter’s GPUs are fully booked; GPUaaS is smaller today but compounding faster — and it’s also the number that can go negative on a bad utilization month. Which is why this is a risk-appetite question before it’s a “which market is bigger” question.

How Each Route Acquires Its First Paying Customers

The customer-acquisition motion differs as much as the software does.

GPUaaS operators sell compute directly to three buyer types: hyperscalers absorbing overflow demand they can’t build fast enough themselves, frontier AI labs running training and inference workloads too large or too intermittent to justify their own fleet, and enterprises that need dedicated capacity without the capital outlay. Analysys Mason projects total GPUaaS revenue climbing from $21 billion in 2024 to $134 billion by 2030 on the back of exactly this buyer mix. The anchor deals are the ones that make headlines: Microsoft, Meta, and OpenAI have all signed multi-billion-dollar capacity commitments with neocloud operators, and those relationships are the backbone of the take-or-pay financing structure described in our neocloud unit economics breakdown. Below the anchor tier sits a broker and spot market for shorter commitments, where price competition is fiercest and margin is thinnest.

The powered-shell sale barely resembles that motion. The audience is real estate and capital markets as much as it is technical buyers. The customer is often a hyperscaler leasing an entire shell to fit out itself (a 36-megawatt shell in Ashburn and a 24-megawatt shell in Manassas, both leased directly to cloud tenants, are documented examples of this structure), or it’s a neocloud that needs power and space faster than it can build and is willing to lease rather than own the shell. That second pattern matters: a meaningful share of neocloud capacity today sits inside leased powered shells, which means the two business models aren’t strictly competitors. One is frequently the landlord of the other. Acquisition here runs through brokers and capital markets desks (securitized single-tenant leases are now a standard financing tool), and through hyperscaler real estate teams directly. Nobody in this channel is pitching GPU-hours.

Which AI Data Center Business Model Fits Your Capital and Risk Profile?

There’s no universal right answer, but there is a reasonably clean way to sort operators into the model that fits them.

1. Capital-light operators with land, power access, or real estate experience, and no GPU technical bench, should default to powered shell. If your team has never run a Slurm cluster or negotiated a take-or-pay GPU contract, don’t let a hot market talk you into learning that discipline under pressure. Powered shell rewards exactly the skills real estate and utility-adjacent operators already have: site control, power procurement, permitting, and long-lease structuring. The revenue is lower-margin but far more predictable, and the tenant, not you, eats the GPU depreciation curve entirely.
2. Well-capitalized operators with access to debt markets and an actual bench in scheduling, billing, and ML infrastructure should consider GPUaaS. This is not a side project. It requires building or buying a multi-tenant orchestration stack, negotiating GPU supply at scale, and being comfortable borrowing against contracted cash flow the way CoreWeave and Nebius do. The margin ceiling is much higher. So is the downside once utilization slips below the breakeven band, a risk we’ve detailed at length in so you have GPUs, then what?
3. Regional integrators and mid-market operators often do best starting with powered shell and graduating later. Build the site, power, and cooling discipline first, generate stable lease income, and use that cash flow and credibility to eventually build or acquire the orchestration and billing layer GPUaaS demands. Several of today’s neocloud giants started as colocation or crypto-mining infrastructure operators before pivoting; the physical competence came first, the compute-selling business came second.

Why Factory-Built Modular De-Risks the Infrastructure Layer Either Way

One thing gets missed in almost every version of this debate. The business-model decision and the infrastructure-delivery decision are separate questions, and the second one has gotten a much clearer answer no matter which model you pick.

Sell GPU-hours or lease powered shells, the underlying risk doesn’t change: time-to-energization. Every month a facility isn’t live is a month of dead capital, sitting in depreciating GPUs on one side or an unrented lease on the other. The cost of that delay, worked out in detail for GPU clusters specifically, is in our GPU data center deployment guide, and the underlying math applies just as directly to an empty shell collecting no rent.

Factory-built modular data centers compress that exposure on both sides of the model. A prefabricated module arrives with power distribution and cooling pre-integrated, fire suppression too, all of it factory-tested before it ships, which cuts the commissioning risk a stick-built facility carries into its first months of operation. It’s designed to meet Tier III/Tier IV principles from the factory floor rather than having them bolted on during a site punch-list. And because the module itself doesn’t care what’s inside it, whether that’s a neocloud tenant’s GPUs or a colocation customer’s own servers, an operator can commit to modular capacity before fully deciding which business model sits on top, then configure the software and contract layer once the physical risk is already retired. The full case for that approach, including cost comparisons against conventional builds, is in our modular data center cost guide and the broader modular data center guide.

The infrastructure layer is where both business models are most exposed and least differentiated. Solving it with a factory-built approach doesn’t pick your business model for you. It just means the choice you eventually make won’t be undone by a late substation or a botched commissioning schedule.

The Bottom Line: Pick the Business, Not Just the Building

GPUaaS and powered shell aren’t two flavors of the same business. One sells a depreciating asset by the hour and lives or dies on utilization; the other rents infrastructure by the month and lives or dies on occupancy and lease terms. Confusing them is how operators end up with a facility built for the wrong risk profile, or a compute business with no orchestration layer to run it.

The market is large enough for both, and getting larger for both at once. What decides which one you should build isn’t ambition. It’s whether your balance sheet, your technical bench, and your appetite for a 70%-utilization tightrope match the model you’re about to sign up for.

FAQ

What is the difference between GPUaaS and colocation?

GPU-as-a-service (the neocloud model) means the provider owns or leases the GPUs and sells compute by the hour, carrying utilization and depreciation risk. Colocation, including the powered-shell variant, means the provider supplies power, cooling, and connectivity while the customer brings and owns their own GPU hardware, so the provider carries real-estate and occupancy risk instead.

Is GPU-as-a-service more profitable than powered shell colocation?

GPU-as-a-service can be more profitable, but only above roughly 70% utilization on debt-financed clusters; its bare-metal gross margins run 55-65% before depreciation. Powered shell’s margin ceiling is lower but far more predictable, since triple-net colocation leases pay whether or not the tenant’s GPUs are fully booked.

What is a powered shell data center?

A powered shell is a finished building with power and fiber connectivity already brought to the site, but with the interior left unfinished. The tenant fits out the interior with cooling, backup generation, and IT racks to their own specification, which lets the landlord limit scope while still capturing rent on the power and land.

Can a company operate both a GPUaaS and a powered-shell business at once?

Yes, and many do. A significant share of neocloud GPU capacity today sits inside leased powered shells rather than operator-owned buildings, meaning one company is frequently the landlord and another runs the compute business inside its walls. Operators sometimes start in powered shell for stable cash flow and add a GPUaaS line once they’ve built the orchestration and billing capability it requires.

What contract structures are used in each AI data center business model?

GPUaaS operators typically sign 2-5 year take-or-pay GPU capacity contracts with customers, then borrow against that contracted revenue to finance hardware purchases. Powered-shell operators sign 10-15+ year triple-net leases, where the tenant pays a base rate per kilowatt plus a share of operating costs and, often, a critical-load charge on top.

What software does a GPU colocation or GPUaaS business need that traditional colocation doesn’t?

GPUaaS requires multi-tenant orchestration (Kubernetes or Slurm-based schedulers), tenant isolation, quota management, and billing systems that meter usage at the job level rather than a flat hourly rate. Traditional powered-shell colocation instead needs precise power metering at the branch-circuit level, DCIM capacity tracking, and SLA administration built around uptime and PUE rather than compute scheduling.

Which AI infrastructure business model should a new operator choose?

Operators with real estate, land, or power development experience and no GPU technical bench generally fit better with powered shell, since it transfers utilization and obsolescence risk to the tenant. Operators with access to debt markets and a bench in scheduling and billing infrastructure are better suited to GPUaaS, which carries a higher margin ceiling alongside materially higher utilization risk.

Does modular construction favor one AI data center business model over the other?

No. Factory-built modular capacity is agnostic to what runs inside it, so it reduces time-to-energization risk for GPUaaS operators racing GPU depreciation and for powered-shell operators racing to start collecting rent, without committing either one to a specific business model in advance.