Every business deploying high-performance computing infrastructure eventually faces the same fundamental question: how much of the stack do you own, and how much do you lease?
It sounds like an infrastructure procurement question. It is actually a risk allocation question. The answer determines your capital exposure, your vulnerability to hardware obsolescence, and your revenue or cost ceiling per megawatt of deployed compute. The two dominant models in the market today, traditional colocation and neocloud leasing, sit at opposite ends of that spectrum, and the financial gap between them is not subtle.
Here is how the math works, and what it means for the organizations on both sides of these arrangements.
Colocation: Maximum Protection, Minimum Upside
Traditional colocation is the most capital-light model available for HPC deployment. A service provider builds and operates the physical infrastructure: power, cooling, physical security, and network connectivity. The cost to build that powered shell runs between $8 million and $12 million per megawatt. Beyond that, the provider’s exposure ends.
The tenants bring their own GPUs. They own the hardware, they manage the hardware, and they absorb every cost and risk associated with it.
From the provider’s perspective, this is an attractive risk profile. There is no exposure to hardware procurement cycles, no balance sheet impact from GPU fleet depreciation, and no vulnerability when NVIDIA releases a new architecture that cuts the value of the previous generation. The provider builds power and cooling infrastructure that will be relevant for decades and collects rent from whoever happens to be running compute inside it.
The financial return on that risk posture is modest. Traditional colocation generates approximately $1.65 million per megawatt per year. The provider has offloaded almost all the technology risk and, in exchange, has capped its upside accordingly.
Neocloud Leasing: Maximum Upside, Maximum Obsolescence Risk
Neocloud leasing takes a fundamentally different position. The leasing company funds the GPU fleet itself, not just the facility. Rather than building a powered shell and waiting for tenants to fill it with hardware, a neocloud operator deploys the GPUs, manages the full compute stack, and leases access to that compute capacity as a service.
The revenue difference is significant. Neocloud models generate between $11 million and $25 million per megawatt per year, compared to $1.65 million for traditional colocation. That is a 7x to 15x difference in revenue per megawatt, driven entirely by the decision to own and operate the hardware layer rather than leaving it to tenants.
The risk that comes with that upside is hardware obsolescence. High-end AI GPUs have useful lives of four to seven years under normal conditions, but the effective competitive shelf life is considerably shorter. NVIDIA’s annual release cadence means a new architecture arrives roughly every 18 months. The transition from Blackwell to Rubin is the current example: operators who funded large Blackwell GPU fleets are now managing assets whose relative performance value is declining as Rubin deployments begin.
For a neocloud operator that has committed hundreds of millions of dollars to a specific GPU generation, an accelerated obsolescence cycle is not an abstract risk. It is a balance sheet event.
The Middle Ground: Where Most Enterprise Decisions Actually Live
The colocation-versus-neocloud framing is useful precisely because it clarifies the tradeoff, but most enterprise HPC decisions do not sit cleanly at either pole.
Full vertical integration, owning the facility, the power infrastructure, and the GPU fleet, costs approximately $45 million per megawatt. That is the high end of the capital intensity spectrum and is effectively reserved for hyperscalers and large sovereign AI deployments with long time horizons and the capital structure to absorb hardware cycles.
For most enterprises, the practical question is where on the spectrum between $8 million and $45 million per megawatt their infrastructure strategy should sit, and how much hardware obsolescence risk they are willing to carry in exchange for the revenue or capability upside that comes with owning more of the stack.
Consumption-based GPUaaS models from neocloud operators give enterprise tenants access to the high end of compute performance without carrying the obsolescence risk directly. The neocloud operator absorbs that risk and prices it into the per-token or per-hour access cost. The tenant pays a premium relative to owning hardware outright but avoids the balance sheet exposure of a GPU fleet that depreciates every 18 months.
What the Revenue Gap Tells You About Where Value Is Being Created
The $1.65 million versus $11 million to $25 million per megawatt comparison is not just a financial metric. It is a signal about where value is being created in the HPC infrastructure stack.
Physical infrastructure, power, cooling, and shell construction, is a commodity. It is essential, but it is not where the economics concentrate. The economics concentrate at the hardware and software layer, where GPU performance, orchestration quality, and inference optimization determine what a megawatt of compute can actually produce.
This has implications for enterprises evaluating their own HPC strategy. An organization that deploys GPU-accelerated workloads on top of leased colocation space is capturing value at the compute layer, not the facility layer. An organization that relies entirely on consumption-based cloud GPU access is paying the neocloud operator’s margin in exchange for avoiding the capital and obsolescence risk of owning the fleet.
Neither position is inherently correct. The right answer depends on the organization’s capital structure, its tolerance for hardware obsolescence risk, and the duration and predictability of its HPC workload requirements.
The Decision Framework: Risk Tolerance Before Procurement
Before choosing between colocation, neocloud leasing, managed GPUaaS, or vertical integration, the productive question is not “which model is cheapest?” It is “which risks are we willing to own?”
Organizations with predictable, long-duration HPC workloads and strong capital positions have a reasonable case for owning more of the stack and capturing the revenue or cost advantage that comes with it. Organizations with variable workloads, shorter planning horizons, or limited appetite for hardware obsolescence risk are better served by consumption-based models that transfer that risk to a specialized operator.
What the numbers make clear is that the decision has real financial consequences in both directions. Choosing colocation to minimize risk means leaving significant value on the table if the workload is large and durable. Choosing neocloud leasing to capture that value means accepting a hardware obsolescence exposure that needs to be actively managed.
The organizations generating the best outcomes from HPC investment are the ones that made this tradeoff explicitly, with clear eyes about what each model requires, before committing capital.
How CloudSyntrix Can Help
Navigating the colocation, neocloud, and managed infrastructure landscape requires understanding not just the financial tradeoffs but the integration requirements of each model. A colocation deployment that cannot connect cleanly to existing data pipelines and application infrastructure delivers its risk protection at the cost of operational utility. A neocloud arrangement that is not properly integrated into the enterprise stack adds cost without adding capability.
CloudSyntrix provides the systems integration expertise to make whichever model fits your organization actually work in production. From cable to cloud, CloudSyntrix delivers seamless systems integration with speed and precision. Their expert Strike Teams connect infrastructure, applications, and multi-cloud environments, integrating legacy systems, building data lakes, deploying wide-area networks, and training large language models.