--- title: "Why Modular AI Data Centers? | Aguasto" description: "Aguasto's approach to standardized engineering, phased development and independently operated AI facilities, with explicit power, permitting and capital trade-offs." canonical: "https://www.aguasto.com/infrastructure/why-modular" last-updated: "2026-10-13" --- # Why Modular AI Data Centers? | Aguasto Canonical HTML: [Why Modular](https://www.aguasto.com/infrastructure/why-modular) > Aguasto's approach to standardized engineering, phased development and independently operated AI facilities, with explicit power, permitting and capital trade-offs. [Resources](https://www.aguasto.com/resources) [GPU Library](https://www.aguasto.com/gpus) [Model Index](https://www.aguasto.com/models) [Inference Stack](https://www.aguasto.com/inference-stack) [Learn](https://www.aguasto.com/learn) [Insights](https://www.aguasto.com/insights) [Research](https://www.aguasto.com/research) [Engineering tools](https://www.aguasto.com/tools) The Aguasto thesis / Development strategy ## Why Modular AI Data Centers? An alternative path from infrastructure planning to operational compute. Standardized engineering and independently deployable facilities can improve execution flexibility where power, land-use and operating requirements are satisfied. Original Aguasto conceptual illustration. Not manufacturer geometry, an operating Aguasto asset, or a measured benchmark. Why now ### More capacity needs more deployment models. Aguasto is developing modular, high-performance data centers and distributed compute infrastructure for enterprise AI and next-generation inference. Traditional campuses will remain important. Independent inference replicas and dedicated enterprise workloads can also be evaluated in phased facilities, provided latency, reliability, power and customer requirements support the design. Modularity is a deployment strategy, not proof of readiness. LBNL's June 2026 report describes bottlenecks in connecting large loads. CBRE's H1 2026 evidence shows continued construction alongside power and approval constraints. Aguasto's response is a development thesis; these reports do not validate a particular Aguasto project. [LBNL Speed to Power](https://datacenters.lbl.gov/publications/speed-power-solutions-accelerating) · [CBRE H1 2026](https://www.cbre.com/insights/books/north-america-data-center-trends-h1-2026) Local water conditions, land-use scrutiny and equipment lead times can add constraints alongside electricity. Phased financing must still support qualified engineering, procurement commitments and customer demand. A modular enclosure changes selected delivery interfaces; it does not remove these obligations. [Review current Raleigh, FERC/PJM, LBNL and reliability evidence](https://www.aguasto.com/research/hyperscale-development-friction#current-evidence) , including unresolved source boundaries and continued hyperscale construction. ### Six principles. Real dependencies. 01 #### Industrialized deployment Prepare repeatable equipment assemblies, rack integration and selected enclosure systems off-site. Foundations, utility connection, heat rejection and final commissioning still need project-specific engineering. 02 #### Phased capacity Additional modules can divide development into executable stages. Each phase still depends on authorized electricity capacity, customer demand, equipment delivery and financing. 03 #### Distributed operations Independent sites can host inference replicas and route requests regionally. Geographically separated GPUs do not become one low-latency NVLink training fabric. 04 #### Repeatable engineering Standardized interfaces and acceptance tests can improve consistency across cooling, rack integration, networking and operations. Site-specific loads and conditions still require validation. 05 #### Infrastructure flexibility Containerized systems can be evaluated for different suitable industrial settings. Physical fit is only one part of land-use, utility, environmental and operating feasibility. 06 #### Capital deployment Phasing can limit the commitment needed for an initial stage. Repeated electrical infrastructure, networking and support teams can increase cost per unit of capacity. ### Factory work and site work share an interface. Illustration: Conceptual modular facility. Original isometric illustration for the module assembly development stage. A proposed enclosure shows equipment spaces and a central service aisle as systems are assembled. The enclosure is cut away to reveal its interior. Components are shown in their illustrative assembled positions. All systems are shown. No geographic intelligence layers are selected.No real property, equipment inventory, capacity, service right or operating Aguasto facility is represented. Physically installed electrical equipment does not establish usable capacity or utility authorization. Original integration study. Generic geometry and spatial arrangements explain interfaces, without a project specification or deployment timeline. #### What can progress in parallel? Selected enclosure systems, electrical assemblies, rack installation and factory checks can be prepared away from the site. Work should proceed against validated mechanical, electrical, cooling and transport interfaces. Site permissions, utility studies, foundations, access and heat rejection follow their own dependencies. Incomplete interfaces can move rework from the construction site to the factory rather than remove it. #### How does readiness become operational? Factory acceptance is an input to integrated site testing. Protection settings, loads, thermal behavior, network paths and operating handover must be assessed together before a facility supports a customer service. [Schneider Electric: prefabrication and parallel work](https://blog.se.com/datacenter/2022/05/17/supporting-network-edge-prefabricated-modular-data-centers/) · [OCP: cooling integration and logistics](https://www.opencompute.org/documents/ocp-liquid-cooling-integration-and-logistics-white-paper-revision-1-0-1-pdf) · [Vertiv: startup and commissioning](https://www.vertiv.com/en-us/services-catalog/services/project-services/startup-and-commissioning-services/) ### Continue the research [Inspect module assembly and rack interfaces](https://www.aguasto.com/infrastructure#infrastructure-transformation) [Define a local compute-cluster boundary](https://www.aguasto.com/gpu-cloud#dedicated-clusters) ### Factory preparation changes part of the sequence. Conventional and modular developments both need project approvals, electricity, construction interfaces and integrated commissioning. Selected modular assemblies may be prepared off-site while appropriate site work proceeds. This diagram shows dependencies, not guaranteed durations. Conventional and modular development · Conceptual dependencies, without duration estimates #### Conventional campus - **Approvals & power**Site permissions, utility studies and authorized service - **Equipment & construction**Site construction and equipment integration follow the project schedule - **Commissioning**Integrated protection, cooling, load and operating tests #### Modular development - **Approvals & power**The same permissions and electrical prerequisites - **Equipment & construction**Selected factory assemblies may proceed alongside suitable site work - **Commissioning**Equivalent integrated acceptance tests before service 1. Site & approvals Land use, environmental assessment, geotechnical design and local permits apply to both approaches. 2. Power & utilities Utility-authorized capacity, interconnection work and an energization plan remain on the critical path. 3. Factory & site works Modular equipment preparation may proceed alongside suitable site work, subject to completed interfaces and procurement. 4. Commissioning Integrated load testing, protection settings, cooling performance and operating handover establish readiness. **Potential advantage** Prefabricated equipment **Requirement that remains** Utility interconnection, land use and site-specific civil works. **Potential advantage** Phased investment **Requirement that remains** Repeat infrastructure costs, financing, utilization and customer demand. **Potential advantage** Independent locations **Requirement that remains** Regional staffing, network latency, software consistency and failover capacity. **Potential advantage** Existing electrical equipment **Requirement that remains** Utility-authorized capacity, protection studies and an approved operating use. **Potential advantage** Standard cooling interfaces **Requirement that remains** Actual rack loads, fluid quality, heat rejection, commissioning and maintenance. Containers do not bypass permits. An energized transformer does not prove immediate data-center capacity. No universal deployment timetable or operating Aguasto installation is represented. Develop capacity in parallel ### Bring the workstreams together. Keep their dependencies visible. Factory preparation, appropriate site work and compute-software preparation can progress separately where their accepted inputs allow it. Every branch still meets physical commissioning and workload acceptance. The intended benefit is a shorter path to usable compute; no Aguasto duration or deployment result is available. Parallel delivery / Dependency study Three workstreams. One accepted operating configuration. Illustration: Parallel workstreams: Engineering dependency study. Nineteen activities across infrastructure, modular equipment and compute. Highlighted lines identify prerequisites of the selected activity. Horizontal positions are not time or guaranteed concurrency. Original dependency map, not a delivery calendar. Activity numbers identify the controls below. “Separate” means no direct or indirect dependency on the selected activity in this study; shared resources and local requirements can still constrain parallel work. Infrastructure01 Site qualificationprerequisite 02 Utility authorizationNo dependency in this study 03 Engineeringselected 04 Permittingdownstream 05 Site worksdownstream 06 Electrical & cooling integrationdownstream Modular equipment07 Module specificationprerequisite 08 Factory manufacturedownstream 09 Factory equipment testingdownstream 10 Shipmentdownstream 11 Equipment installationdownstream 12 Physical commissioningdownstream GPU & compute platform13 GPU/OEM architecture validationprerequisite 14 Hardware procurementdownstream 15 Cluster networkingdownstream 16 Infrastructure softwareNo dependency in this study 17 Model-serving systemsNo dependency in this study 18 Performance & service validationdownstream 19 Customer provisioningdownstream Selected activity #### Engineering Qualified specialists reconcile electrical, structural, thermal, fire, network and transport interfaces. Repeatability does not replace professional site-specific design. Direct prerequisites Site qualification · GPU/OEM architecture validation · Module specification Evidence required Accepted interfaces, design responsibilities and a configuration-specific engineering package. Downstream dependencies Permitting · Site works · Electrical & cooling integration · Factory manufacture · Factory equipment testing · Shipment · Equipment installation · Physical commissioning · Hardware procurement · Cluster networking · Performance & service validation · Customer provisioning Distributed by architecture ### Keep clusters local. Route eligible inference workloads. Independent facilities can host replicated model-serving configurations. A unified operating platform would monitor health, model versions, capacity allocation and data-location policy before directing a request to an eligible replica. Cross-region routing is an application function; local training and tightly synchronized inference have different network requirements. Distributed inference / Concept study Independent clusters, conditional workload continuity. Illustration: Independent inference replicas: One site isolated. A separate eligible replica may support continuity after site A is isolated. B needs compatible model versions, tested routing, approved data location and sufficient spare capacity. Three abstract replicas illustrate architecture, not actual Aguasto sites or a live routing service. Original architecture model. The three replicas are abstract failure domains. This study makes no claim about Aguasto site count, operational availability or an enterprise SLA. Normal routing One site isolated Shared dependency fails No spare capacity Model versions differ Data location restricts routing One site isolated #### A separate eligible replica may support continuity after site A is isolated. B needs compatible model versions, tested routing, approved data location and sufficient spare capacity. ##### Rack-local connectivity The configured accelerator/server/rack interconnect. Manufacturer system boundaries and supported parallelism determine its role. ##### Intra-site scale-out The local cluster’s network, storage and runtime topology. Multi-node execution requires its own measured acceptance. ##### Cross-region routing Application requests to eligible independent model replicas. It does not create a geographically distributed low-latency training fabric. #### Single-site reliability Local power, cooling, hardware, network and operating controls determine a facility’s behavior under the tested fault scope. #### Fleet-level availability Multiple facilities require independently assessed failure domains, shared-service dependencies and maintained operating capacity. #### Workload continuity The customer’s model, spare resources, data policy, routing and accepted recovery behavior must all support a tested transition. This proposed architecture needs operational testing. More sites alone do not establish better uptime, customer continuity or an enterprise SLA. Design priorities ### Designed for more resource-conscious AI infrastructure. Adaptability and appropriate reuse can reduce some avoidable work. Actual safety, transport limits, replacement needs and lifetime impacts determine whether reuse is suitable. Modular equipment is not indefinitely reusable. - Evaluate appropriate reuse and avoidable construction waste against actual condition and safety. - Specify adaptable equipment interfaces and future replacement/transport limits. - Select qualified cooling and heat-rejection systems for the real climate and hardware envelope. - Consider local freshwater conditions alongside electricity-related and lifecycle impacts. - Improve useful GPU work through evaluated software, utilization and operating practices. - Measure operating outcomes and report verified results when facilities become operational. ### Continue the research [Cooling, water and efficiency evaluation](https://www.aguasto.com/infrastructure/cooling-resource-efficiency) [Explore transparent electricity and token-cost assumptions](https://www.aguasto.com/tools) ### From infrastructure to inference. A conceptual operating architecture. Aguasto's compute and API services remain in development. Conceptual request path: Request → Routing → Prefill → Decode → Response. Topology varies. Independent sites host replicas; routing does not create a geographically distributed NVLink fabric. Electrical service → Facility → Rack → Runtime → Model → API → Customer - **Electrical service**Utility authorization and protection - **Modular facility**Site works, heat rejection and approvals - **GPU rack**Power, cooling and fabric integration - **Serving software**Pinned runtime and hardware compatibility - **Model deployment**License, weights, cache and evaluation - **Inference API**Authentication, quotas and observability - **Customer**Workload, latency and reliability requirements A conceptual engineering chain for planned services. Each interface requires project-specific verification. Conceptual distributed inference · Independent facilities and model replicas **Enterprise request → Service routing**Choose an eligible region using latency, policy, health and available capacity. #### Region A Independent facility Local GPU racks & runtime Model replica & local cache #### Region B Independent facility Local GPU racks & runtime Model replica & local cache #### Region C Independent facility Local GPU racks & runtime Model replica & local cache Geographic request routing does not join GPUs into one NVLink fabric. Failover needs compatible replicas, spare capacity and a defined data-handling policy. Electrical service supports the facility. The facility supports GPU racks, network paths and serving software. Model deployment and application interfaces depend on tested operations, security and capacity planning at every layer. [Read the full architecture](https://www.aguasto.com/learn/from-power-to-inference) ### Capital and operating trade-offs Smaller phases can reduce the capital committed before the next development decision. They can also repeat equipment, networking and operational support that a larger campus shares. The appropriate design depends on measured workload needs and a project-specific cost model—not modularity alone. ### Continue the research [Read the flagship modular research publication](https://www.aguasto.com/research/why-modular-ai-infrastructure-matters-now) [Review jurisdiction-specific development evidence](https://www.aguasto.com/research/hyperscale-development-friction) [Dedicated and shared inference economics](https://www.aguasto.com/insights/dedicated-and-shared-inference-economics) [Discuss potential OEM and strategic collaboration](https://www.aguasto.com/partners) [About Aguasto](https://www.aguasto.com/company) ## Company navigation - [Home](https://www.aguasto.com/) - [Infrastructure](https://www.aguasto.com/infrastructure) - [GPU Cloud](https://www.aguasto.com/gpu-cloud) - [AI Inference](https://www.aguasto.com/ai-inference) - [Development & Investment](https://www.aguasto.com/development) - [Company](https://www.aguasto.com/company) - [Contact](https://www.aguasto.com/contact) - [Privacy](https://www.aguasto.com/privacy) - [Resources](https://www.aguasto.com/resources) - [GPU Library](https://www.aguasto.com/gpus) - [Model Index](https://www.aguasto.com/models) - [Inference Stack](https://www.aguasto.com/inference-stack) - [Learn](https://www.aguasto.com/learn) - [Insights](https://www.aguasto.com/insights) - [Research](https://www.aguasto.com/research) - [Industry cases](https://www.aguasto.com/research/cases) - [Why Modular](https://www.aguasto.com/infrastructure/why-modular) - [Cooling & Resource Efficiency](https://www.aguasto.com/infrastructure/cooling-resource-efficiency) - [Partners](https://www.aguasto.com/partners) - [GPU comparison](https://www.aguasto.com/gpus/compare) - [Model comparison](https://www.aguasto.com/models/compare) - [Engineering tools](https://www.aguasto.com/tools) - [Development friction](https://www.aguasto.com/research/hyperscale-development-friction) - [NVIDIA H100 SXM](https://www.aguasto.com/gpus/h100) - [NVIDIA H200 SXM](https://www.aguasto.com/gpus/h200) - [NVIDIA B200 SXM](https://www.aguasto.com/gpus/b200) - [NVIDIA B300](https://www.aguasto.com/gpus/b300) - [NVIDIA GB200 NVL72](https://www.aguasto.com/gpus/gb200-nvl72) - [NVIDIA GB300 NVL72](https://www.aguasto.com/gpus/gb300-nvl72) - [NVIDIA Vera Rubin NVL72](https://www.aguasto.com/gpus/vera-rubin-nvl72) - [AMD Instinct MI300X](https://www.aguasto.com/gpus/mi300x) - [AMD Instinct MI325X](https://www.aguasto.com/gpus/mi325x) - [AMD Instinct MI355X](https://www.aguasto.com/gpus/mi355x) - [AMD Instinct MI455X](https://www.aguasto.com/gpus/mi455x) - [AMD Helios](https://www.aguasto.com/gpus/helios) - [GLM-4.5](https://www.aguasto.com/models/glm-4-5) - [Qwen3-235B-A22B](https://www.aguasto.com/models/qwen3-235b-a22b) - [DeepSeek-R1](https://www.aguasto.com/models/deepseek-r1) - [Llama 3.3 70B Instruct](https://www.aguasto.com/models/llama-3-3-70b-instruct) - [Gemma 3 27B IT](https://www.aguasto.com/models/gemma-3-27b-it) - [gpt-oss-120b](https://www.aguasto.com/models/gpt-oss-120b) - [Kimi-K2-Instruct](https://www.aguasto.com/models/kimi-k2-instruct) - [GLM-5.3](https://www.aguasto.com/models/glm-5-3) - [Qwen3.8-27B](https://www.aguasto.com/models/qwen3-8-27b) - [DeepSeek-V4.1-Flash](https://www.aguasto.com/models/deepseek-v4-1-flash) - [Llama 4 Scout 17B-16E Instruct](https://www.aguasto.com/models/llama-4-scout-17b-16e-instruct) - [Gemma 4 31B IT](https://www.aguasto.com/models/gemma-4-31b-it) - [Kimi K3](https://www.aguasto.com/models/kimi-k3) - [vLLM](https://www.aguasto.com/inference-stack/vllm) - [SGLang](https://www.aguasto.com/inference-stack/sglang) - [TensorRT-LLM](https://www.aguasto.com/inference-stack/tensorrt-llm) - [NVIDIA Dynamo](https://www.aguasto.com/inference-stack/dynamo) - [NVIDIA Triton Inference Server](https://www.aguasto.com/inference-stack/triton) - [AMD ROCm](https://www.aguasto.com/inference-stack/rocm) - [Ray Serve](https://www.aguasto.com/inference-stack/ray-serve) - [Kubernetes GPU orchestration — NVIDIA device plugin](https://www.aguasto.com/inference-stack/kubernetes) - [Model weights are only the beginning of the memory budget](https://www.aguasto.com/learn/model-weights-and-runtime-memory) - [Why tokens per second is not enough to benchmark inference](https://www.aguasto.com/research/inference-benchmark-methodology) - [Prefill, decode and the cost of separating them](https://www.aguasto.com/learn/prefill-decode-and-disaggregated-serving) - [Choosing between vLLM, SGLang and TensorRT-LLM](https://www.aguasto.com/learn/choosing-an-inference-serving-framework) - [From Power to Inference—Building a Distributed AI Compute Network](https://www.aguasto.com/learn/from-power-to-inference) - [Dedicated compute and shared inference: compare the whole workload](https://www.aguasto.com/insights/dedicated-and-shared-inference-economics) - [Distributed inference reliability begins with the failure boundary](https://www.aguasto.com/learn/distributed-inference-reliability) - [Core Scientific: reading contracted, energized and billed capacity](https://www.aguasto.com/research/cases/core-scientific-capacity-disclosures) - [Why Modular AI Infrastructure Matters Now](https://www.aguasto.com/research/why-modular-ai-infrastructure-matters-now) - [What Is a Modular AI Data Center?](https://www.aguasto.com/learn/what-is-a-modular-ai-data-center) - [How Modular AI Data Centers Are Built and Commissioned](https://www.aguasto.com/learn/modular-data-center-factory-to-commissioning) - [Liquid Cooling for High-Density GPU Systems: Follow the Interfaces](https://www.aguasto.com/learn/liquid-cooling-gpu-rack-interfaces) - [Why Electricity Availability Shapes AI Infrastructure Deployment](https://www.aguasto.com/research/electricity-availability-and-time-to-power) - [What Is KV Cache? Capacity, Reuse and Isolation](https://www.aguasto.com/learn/kv-cache-capacity-and-policy) - [Continuous Batching: Throughput Without Hiding Latency](https://www.aguasto.com/learn/continuous-batching-and-latency) - [Applied Digital: Reading the Polaris Forge 1 Delivery Milestones](https://www.aguasto.com/research/cases/applied-digital-polaris-forge-capacity-milestones) - [Crusoe Abilene: Separate Delivered Infrastructure from Expansion Plans](https://www.aguasto.com/research/cases/crusoe-abilene-delivery-and-expansion) - [Data Center Cooling: Choosing Air, Liquid, Dry and Hybrid Architectures](https://www.aguasto.com/learn/data-center-cooling-air-liquid-dry-hybrid) - [Can Data Centers Reduce Water Consumption? Define the Water Boundary First](https://www.aguasto.com/research/data-center-water-consumption-and-withdrawal) - [Measuring PUE, WUE and AI Compute Efficiency at the Right Boundary](https://www.aguasto.com/learn/measuring-pue-wue-and-ai-compute-efficiency) - [The Environmental Trade-offs of High-Density AI Infrastructure](https://www.aguasto.com/research/environmental-tradeoffs-and-infrastructure-reuse) - [Official LinkedIn](https://www.linkedin.com/company/aguasto) This is a generated representation of the public HTML page. Diagrams are conceptual. GPU and inference services are planned and are not currently available to book or call. The contact page describes the available inquiry flow.