Building an AI data center is becoming a race against more than GPU availability. Power, cooling, networking, site readiness and software orchestration can all determine how quickly a workload reaches production. RAVEL and Strata Expanse are now trying to shorten that cycle with the first commercial deployment of their AMPHIX AI Infrastructure Platform, with RT-One using an AMPHIX Center of Excellence in California to validate the infrastructure configuration it plans to replicate across its growing AI footprint in the Americas.

The AI infrastructure market has spent the past several years optimizing for one headline metric: how many accelerators can be deployed.

That metric is becoming less useful on its own.

As AI clusters become more power-dense and deployments grow geographically distributed, getting GPUs into a building is only one part of the problem. Operators also need power, cooling, networking, physical sites, security and workload orchestration to work together before a production customer can actually use the infrastructure.

RAVEL and Strata Expanse are targeting that gap with AMPHIX, a deployment model designed to create a production-oriented environment where AI infrastructure can be tested before permanent facilities are completed.

The companies have named RT-One as the first commercial customer of the platform.

The model is built around Centers of Excellence (COEs) that combine site-ready land, resilient power, secure networking, certified compute configurations and orchestration software. The goal is to give AI infrastructure operators a place to validate their intended production stack under realistic power, thermal and workload conditions before committing to a large-scale rollout.

That approach addresses a growing problem in AI infrastructure: the infrastructure itself is becoming a bottleneck.

Gartner estimates global data-center electricity consumption will reach 565 terawatt-hours in 2026, up 26% from 2025. AI-optimized servers are expected to account for 31% of data-center electricity consumption this year, with their power consumption projected to exceed that of conventional servers in 2027.

Gartner has also identified power availability as a constraint on AI scaling, making access to electricity increasingly important when deciding where new AI capacity can be built.

In that environment, a conventional pilot can be an expensive experiment.

RAVEL says production-grade pilots can take as long as nine months and cost up to $4 million, while still failing to reproduce all the conditions of the eventual production environment. Those figures come from the companies rather than an independent industry benchmark, but they illustrate the problem AMPHIX is designed to address: organizations can spend heavily proving an architecture that does not necessarily translate cleanly to the next data center.

AMPHIX’s proposition is essentially test once, then replicate.

RT-One will use an AMPHIX COE near Colusa, California, to validate its high-performance technology stack, tune operational policies for performance, energy and cost, train operations personnel and establish customer environments before its permanent campuses are finished.

The configurations proven in California are then intended to become a reference architecture for RT-One’s wider deployment model.

That is particularly important for RT-One because the company is developing a federated AI infrastructure strategy spanning the Americas. RAVEL and RT-One announced their strategic relationship in June, with RAVEL providing technology and expertise for managing geographically distributed and hybrid infrastructure as a unified platform.

The idea is to establish a standardized deployment block rather than independently engineer every location.

That could have significant implications for sovereign AI infrastructure.

AI operators increasingly need to deploy compute in different jurisdictions while maintaining control over data, infrastructure and operational policies. A standardized reference architecture can provide consistency, while allowing individual sites to adapt to local regulations, power conditions and physical constraints.

RT-One’s approach is particularly relevant to Latin America, where the company is developing AI technology parks and data centers. Instead of waiting for each campus to be completed before testing its operational model, the COE allows RT-One to work through the infrastructure and operational questions earlier.

RAVEL supplies the orchestration layer through Orchestrate AI, which is intended to manage and optimize workloads across the infrastructure. Strata Expanse provides the physical foundation, including land, power, cooling and secure connectivity.

The division of responsibilities is important because AI infrastructure is becoming increasingly multidisciplinary.

A GPU cluster can be technically operational while still being commercially inefficient. Workloads may not be scheduled optimally, power may be underutilized, cooling capacity may constrain deployment, or network performance may prevent expensive accelerators from reaching expected utilization.

Workload orchestration is therefore becoming a critical layer between physical infrastructure and AI applications.

Research into power-flexible AI data centers is also exploring how workload scheduling can respond to grid conditions, potentially shifting workloads geographically or reducing consumption during periods of grid stress. Recent research demonstrated power flexibility on a 130-kilowatt GPU cluster while maintaining service levels for priority workloads.

That broader direction strengthens the case for infrastructure software that understands more than compute utilization.

AMPHIX is also entering a market where infrastructure providers are increasingly trying to package multiple layers together. RAVEL and Strata Expanse launched AMPHIX in March as a modular AI infrastructure platform combining land, power, secure connectivity, certified compute and orchestration. The initial model was announced across 30 U.S. sites.

The competitive landscape includes hyperscalers such as Microsoft Azure, Amazon Web Services and Google Cloud, specialized GPU cloud providers, colocation companies and emerging neocloud operators. Many of these companies already provide portions of the same stack.

AMPHIX’s difference is its emphasis on pre-production validation and repeatable physical deployment, rather than simply selling access to cloud GPUs.

For enterprises, the model could be useful when AI workloads have requirements that do not fit neatly into public-cloud environments. Regulated industries, sovereign AI programs and companies with predictable high-performance workloads may want dedicated infrastructure while avoiding the risk of designing every deployment from scratch.

But the model also introduces its own questions.

A reference architecture that works in California may not perform identically in Brazil, Paraguay or Colombia. Power quality, climate, regulations, network connectivity and supply chains differ by region. The value of a “test once, scale many times” strategy will therefore depend on how much of the validated configuration can genuinely be standardized.

There is also the question of economics. Consumption-based access to a COE can reduce upfront commitment, but customers still need to establish that the cost of validation is justified by faster revenue generation and lower deployment risk.

For RT-One, the commercial logic is straightforward: if it can validate a configuration once and reuse that knowledge across multiple campuses, every subsequent deployment should require less engineering effort and fewer surprises.

That is a different way of thinking about AI infrastructure.

Instead of treating a data center as a one-off construction project, operators can treat infrastructure as a repeatable product—with a validated architecture, documented operating policies, trained personnel and software-defined orchestration.

As AI clusters grow more expensive and power constraints become more severe, that shift could become increasingly important.

The next competitive advantage in AI infrastructure may not belong solely to the company that can secure the most GPUs. It may belong to the operator that can prove its entire stack works before the concrete is finished—and then reproduce that result across multiple sites.

Market Landscape

The AI infrastructure market is moving from GPU procurement toward integrated AI factories. Compute, power, cooling, networking, storage and orchestration increasingly have to be designed together because failures in any one layer can limit the utilization of the others.

AMPHIX sits between traditional data-center development and cloud infrastructure. Its proposition resembles a hybrid of a reference architecture, production test environment and infrastructure deployment service.

That makes it relevant to the expanding neocloud and sovereign AI markets, where operators are building specialized GPU infrastructure outside the traditional hyperscaler model.

RAVEL’s orchestration layer is particularly relevant because software can determine how efficiently expensive compute is used. As AI workloads become geographically distributed, orchestration can potentially optimize around performance, energy availability, cost and governance.

For enterprise buyers, however, the critical evaluation criteria will extend beyond GPU availability. They will need to assess power guarantees, cooling architecture, network performance, deployment timelines, workload portability, security controls and the degree to which a validated configuration can actually be reproduced at another site.

The underlying market constraint is increasingly physical. AI demand can grow faster than power and data-center capacity, making infrastructure deployment speed a strategic issue rather than merely an engineering concern.

Top Insights

• RT-One is AMPHIX’s first commercial customer, using a California Center of Excellence to validate AI infrastructure before permanent campuses begin production.
• AMPHIX combines physical infrastructure with orchestration software, connecting power, cooling, networking, compute and workload management into a repeatable deployment model.
• Power availability is becoming a major AI constraint, with Gartner forecasting global data-center electricity consumption of 565 TWh in 2026.
• RT-One aims to certify one reference configuration and replicate it, potentially reducing engineering work and deployment risk across AI campuses in the Americas.
• Sovereign AI operators could benefit from the model, but local power, regulatory, networking and climate differences will determine how portable each validated architecture really is.

Power Tomorrow’s Intelligence — Build It with TechEdgeAI