AI Compute Procurement Becomes a Portfolio Problem, Not a Campus Race
The emerging value is not simply in securing the largest possible facility, but in matching workloads to capacity that can be powered, connected and operated when it is needed.
By Jonas Vale · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human field experience or credentials.
Key points
- Crusoe is funding both a large Texas data-center project used by OpenAI and transportable modular facilities, while retaining business lines that span leased space, GPU rental and inference compute.
Sources: S1
- CNBC reports that OpenAI and Anthropic are exploring materially smaller capacity deployments alongside their large infrastructure commitments, with speed to usable capacity a central consideration.
Sources: S2
- The practical split is workload-specific: tightly coupled training favors concentrated infrastructure, while many inference requests can be served through smaller clusters in more locations.
Sources: S2
The scarce product is usable capacity
The recent announcements point to a change in how AI infrastructure buyers may procure compute. Crusoe has raised capital for existing large projects and for its Spark modular data centers, while OpenAI and Anthropic are reportedly seeking smaller deployments in addition to very large commitments. These are not identical strategies or contracts. Together, however, they show that buyers and suppliers are treating capacity as a mix of facility types, locations and operating timelines rather than as a single bet on the largest available campus.
Crusoe’s model illustrates why that mix matters commercially. The company says it can lease data-center space to customers bringing their own GPUs, rent its own GPUs, and sell inference compute. Its new funding is intended to support a large Abilene, Texas site used by OpenAI as well as truck-transportable modular facilities that can connect to large power sources. A supplier able to serve different procurement routes may be useful to customers whose hardware ownership, workload needs and deployment schedules differ.
Sources: S1
Training scale does not erase the case for smaller sites
The largest AI workloads still create a strong case for concentrated infrastructure. CNBC reports that Anthropic has entered a cloud arrangement tied to capacity at a West Virginia development, while OpenAI has announced further infrastructure commitments in Georgia and Ohio after saying it had exceeded an earlier Stargate commitment. Such projects reflect the need for substantial, coordinated computing resources to train frontier models. Smaller deployments should therefore not be read as a replacement for megacampuses.
Sources: S2
The more consequential distinction is operational, not architectural fashion. According to Structure Research’s Jabez Tan, training a large model generally requires many chips to work closely together. Inference, the day-to-day serving of models, can often distribute separate requests across smaller chip clusters. That means a buyer can pursue a large, tightly integrated environment for training while placing inference capacity nearer to available power, existing sites or particular demand centers. Whether a workload can tolerate that distribution remains a technical and commercial decision for each operator.
Sources: S2
Sources: S2
A reported shift gives the portfolio thesis a demand base
The evidence includes a useful measure against which to judge the suppliers’ claims. JLL data cited by CNBC put inference at 9% of global data-center workloads in 2025, versus 14% for training, and project inference to reach 37% by 2030 while training reaches 13%. Those are projections rather than observed future outcomes, and they concern global capacity rather than any one company’s fleet. Still, they explain why a developer such as Crusoe would invest in facilities aimed at distributed inference while preserving exposure to much larger sites.
Sources: S2
OpenAI’s stated rationale for a diversified compute portfolio closely fits this workload split: its spokesperson told CNBC that requirements vary by performance, reliability, timing and cost. The company did not discuss specific commercial conversations. CNBC separately reports discussions involving smaller deployments in the Nordics and potentially the United States, while sources said Anthropic had sounded out opportunities in the U.K. and Nordics. These reports describe exploratory activity, not completed capacity contracts, so they are an early signal of procurement behavior rather than proof of a settled market structure.
Sources: S2
Sources: S2
Modularity addresses some dependencies, not all of them
Crusoe says manufacturing Spark facilities at its own sites can speed deployment and reduce dependence on large construction workforces. CNBC likewise reports that smaller allocations can offer speed to usable capacity because capacity at an existing powered site may be more practical than waiting for a larger block at one location. The common dependency is not merely land: it is the ability to obtain power and bring a functioning facility online. Modular construction may alter the building step, but it does not by itself create electricity, grid connections, GPUs, network connectivity or qualified operating support.
Community and regional constraints reinforce that point. TechCrunch says Crusoe sees smaller centers as a way to avoid at least part of the backlash directed at large complexes, while CNBC reports pressure from local communities around major projects and limited land and power in much of Europe. A distributed approach can reduce exposure to a single contested site, but it also spreads the operational burden across more locations. Each site still needs reliable power, physical security, maintenance processes and network performance appropriate to the service being delivered.
Inference: procurement now needs a workload map
Inference: the cross-source evidence supports a portfolio approach, but not a universal decentralization thesis. Large training commitments remain central, while smaller powered sites appear most compelling where workloads can be separated, capacity is needed sooner, and the operator can manage service consistency across locations. The original practical conclusion is that procurement teams should begin with a workload map—identifying which jobs need close chip coordination and which can be distributed—before choosing between a hyperscale build, a smaller allocation or a modular installation. That is an inference from the reported workload characteristics and deployment constraints, not a claim that any named company has adopted this exact decision process.
The limiting question is repeatability. A fast physical deployment has little value if it cannot deliver predictable availability, networking and operating support at the required site. Conversely, waiting for a single large build can leave deployable demand unserved when smaller capacity is available. The reported appeal of existing powered sites and Crusoe’s plan to manufacture modular facilities both make time-to-operation important, but neither source provides comparative uptime, network-latency, power-cost or service-quality results for these designs. Those missing operating measures prevent a firm conclusion that distributed capacity is cheaper or more reliable.
What could change the assessment
The assessment would strengthen if the reported exploratory smaller-site discussions turn into disclosed contracts with workload scope, delivery timing and operating performance. It would weaken if inference workloads prove to require more centralized infrastructure than expected, or if smaller sites cannot obtain dependable power and network access quickly enough to offset their construction advantage. Updated evidence on the projected balance between inference and training capacity would also matter, because the portfolio case rests heavily on the expectation that serving models becomes a larger share of demand.
For now, the meaningful comparison is between a supplier building both giant and modular facilities and major model developers reportedly looking beyond giant facilities for some of their needs. The connection is not that every deployment will move to smaller sites. It is that compute procurement increasingly has to reconcile workload physics with the real-world availability of power, sites, equipment and operations. The firms that can make those pieces work together, rather than merely announce the largest capacity figure, are better positioned to turn infrastructure commitments into usable AI service.
Why it matters
The AI buildout is often measured in headline capacity, but the operational bottleneck is matching the right compute, power and site conditions to the workload. A portfolio model can diversify that risk, while introducing new demands for distributed operations and service consistency.