What It Takes to Run Enterprise AI in Production and Why We Are Collaborating with AMD

The conversations we are having with CIOs, CTOs and chief data officers are changing. In regulated industries across healthcare, financial services and the public sector, the focus is shifting from how to build AI to where it should run and who is accountable once it is in production. That shift is what led us to a Memorandum of Understanding with AMD today. 

Access to models and compute has improved quickly, allowing teams to stand up new use cases with little friction. Once those systems move into production, the requirements change. Risk tolerance drops, oversight increases and the environment, where it runs and who is accountable, becomes just as important as the model itself. 

Enterprise AI adoption often slows at exactly this point. The technology is capable, but the challenge is running it in environments where governance, uptime, data residency and accountability are required from the start. 

Accelerated compute infrastructure is now a critical decision in enterprise architecture. What has changed is how directly it shapes where AI workloads can run, how they are governed, how they scale and what they cost to operate over time. 

With this MOU, we are working to integrate AMD Instinct^TM GPUs, AMD EPYC^TM CPUs and the ROCm^TM software ecosystem directly into the governed AI infrastructure Rackspace Technology operates, so those decisions can be made deliberately and executed in production. 

In regulated environments, the infrastructure requirements are non-negotiable 

Data cannot move freely across jurisdictions. Decisions require auditability. Systems must be resilient, observable and aligned to policies that exist outside the application. What works in a development or test environment does not hold up under these conditions. 

The decision is moving from model selection to workload placement, governance and operational accountability. In healthcare, financial services and the public sector, that responsibility increasingly sits with an operator. One of the largest Epic environments anywhere runs on Rackspace private cloud, not on a hyperscaler. That is the kind of regulated, mission-critical workload our governed AI infrastructure is built to operate. 

Operating the full stack 

Over the past two years, we have assembled a governed AI stack that reflects how enterprise AI runs across private, hybrid and sovereign environments, with governance built in from the start. This spans the control plane, the data and AI operations layer and the operating model required to run it, from infrastructure through to agents and applications. 

VMware Cloud Foundation 9 is the control plane that unifies compute, storage, networking and security across private, public, edge and sovereign environments. Palantir supports governed data and AI operations. Uniphore enables agent-based workflows. Rackspace will operate across these layers as a single system, with operational responsibility for performance. 

AMD is our anchor partner for accelerated compute 

We are not a hardware vendor building toward managed services. We are a governed operator. What makes AMD the right anchor for that model is not a single component. It is the architecture. 

AMD brings both Instinct^TM GPUs and EPYC^TM CPUs inside a single integrated compute architecture, with ROCm^TM tooling that orchestrates workloads across both. That matters because production enterprise inference is heterogeneous. Frontier models run on GPU. Small language models, classical ML, embeddings, retrieval and many domain-specific workloads run more efficiently on CPU.  

Customers should not be forced to put every workload on the most expensive silicon. Our governed operating model routes each workload to the right compute, which is what production economics actually requires. 

That operating flexibility is critical in sovereign AI deployments, where data residency and jurisdictional requirements dictate where and how systems operate. It also shapes total cost of ownership, as enterprises make deliberate decisions about workload placement, performance and long-term infrastructure efficiency. 

The market today is dominated by commodity GPU rental, where capacity is sold by the hour and the customer carries the burden of integration, security and accountability. We are building the opposite. Dedicated accelerated compute embedded inside a governed operating model where Rackspace owns the stack from silicon to outcomes. Once in place, the full stack, from compute through inference and agents, is designed to operate under a single model of accountability. 

For enterprise buyers, the decision is shifting from purchasing hardware with services attached to selecting an operator that can run AI reliably in production, with accountability for governance, resilience, performance and long-term outcomes built into the model from the start. We aim to deliver an ongoing operating relationship where responsibility continues well beyond deployment as systems scale, workloads evolve and enterprise requirements change over time. 

Forward-deployed engineers keep AI accountable in production 

Forward-deployed engineers are central to how AI systems are operated in production. They run systems inside the customer environment and stay with them as those systems scale, taking responsibility for runtime performance, availability and stability under real- world conditions. 

Accountability stays live after deployment because the engineering team does not step away. Responsibility extends beyond delivery into how the system performs day to day, particularly in environments where uptime, governance and data residency are non-negotiable. 

Forward-deployed engineers extend responsibility into production and keep it there, which is how enterprise AI systems remain reliable when they are treated as infrastructure rather than short-lived projects. 

Rackspace capabilities aligned to how enterprises buy and use AI 

The AMD collaboration is intended to position Rackspace to advance its curated enterprise AI stack and introduce four integrated capabilities. Together, these capabilities span bare metal compute, developer-ready inference tooling, a fully operated inference runtime with defined SLAs, and a governed Enterprise AI Cloud. This gives enterprises a single operator accountable for every layer, aligned to the sovereignty, performance and compliance requirements of their workload. 

The four integrated capabilities include: 

Enterprise AI Cloud provides a fully managed private and hybrid AI environment, launching with AMD InstinctTM accelerators and the Rackspace operating model, with one operator accountable across the stack from accelerated compute to AI inference and agents in production. 

Enterprise Inference Engine is a fully operated, context-aware inference runtime for domain-specific enterprise AI workloads, routing each workload across CPU and GPU based cost, latency and accuracy requirements. Unlike stateless compute rental, the Enterprise Inference Engine retains domain knowledge, session history and enterprise-specific data context across queries, enabling AI agents and large language models to perform with the consistency and institutional memory that production enterprise environments require. Rackspace would own the SLA and take responsibility for availability, scaling and platform performance, with full auditability and cost accountability. This is a production-ready inference capability, not compute by the hour. 

Inference as a Service provides dedicated accelerated compute, launching with AMD InstinctTM GPUs and EPYCTM CPUs, for infrastructure and platform teams. The customer brings their own model, orchestration and engineering. Rackspace would provide reliable bare metal capacity across both CPU and GPU, matched to workload requirements, with operational discipline and performance SLOs. 

Bare Metal AMD Instinct provides dedicated, high-performance, unmanaged bare metal AMD InstinctTM compute for demanding training and inference workloads that require deterministic performance and direct hardware access. 

These capabilities reflect how enterprise AI is deployed and operated in production, with each aligned to a different buyer and level of responsibility, from infrastructure access to fully operated AI systems. 

Building and operating enterprise AI in production 

Together, Rackspace and AMD aim to build and operate AI systems that run reliably under real-world constraints, with clear accountability for how they perform over time. 

The next phase is already emerging. As enterprise AI evolves toward agentic workflows, where machines act on regulated data and processes run continuously without a human in the loop, the demands on governed infrastructure become even more acute. Training will largely sit with specialized providers. Inference, particularly context-aware inference on regulated data, is where production enterprise AI lives, and that is the workload Rackspace is built to operate. 

Inference itself is a portfolio. A frontier model serves a few critical use cases. Small language models, classical ML, retrieval and embeddings serve many more. The cost difference between routing each to the right silicon and running everything on GPU is material at production scale. The AMD collaboration will enable us to operate that heterogeneous stack as a single governed system. 

We intend to expand this capability across our broader ecosystem, with AMD as our anchor compute partner, demonstrating how governed AI systems perform in production at scale.   

The question is not whether organizations can build something that works, but whether they can run it reliably under real-world constraints and stand behind the outcome. That is the goal we are working toward.  

We expect to share that progress over the coming year, including planned live demonstrations of these systems running in real deployment scenarios at AMD Advancing AI in San Francisco this July. We hope to see you there. 

Learn more about how we are working with AMD to run enterprise AI in production here.

Forward Looking-Statements 

This post contains forward-looking statements within the meaning of the Private Securities Litigation Reform Act of 1995. Actual results may differ materially from those anticipated. For a discussion of risks and uncertainties, see Rackspace Technology's filings with the SEC, including its most recent reports on Form 10-K and 10-Q. For additional important information regarding forward-looking statements regarding the non-binding memorandum of understanding (MOU) with AMD, please refer to our press release issued on May 7, 2026, available here. The MOU is a framework for potential collaboration and does not constitute a binding commitment by either party to complete any specific transaction, financing or other commercial arrangement or transaction. No definitive agreements have been reached, discussions remain preliminary, and there can be no assurance that any such arrangements will be entered into or that the parties will reach definitive agreement on terms, or that the anticipated benefits of the collaboration will ultimately be realized. Any third-party financing required to implement the transactions contemplated by the MOU is subject to the availability of financing on acceptable terms. Rackspace assumes no obligation to update these statements except as required by law.