[Op-Ed] Toward “Intentional Multicloud” in the AI Era

An article by Dennis Smith, senior analyst at Gartner, published in IT Daily, examined how AI is changing enterprises’ multicloud approach. It noted that while multicloud in the past was a defensive strategy aimed at avoiding vendor lock-in and creating system redundancy, its status has now shifted to a proactive technology requirement for enterprises.

The article also pointed out that many existing multicloud strategies lack a solid foundation. In particular, it cited the absence of clear workload-based adoption rationale, underestimation of data egress costs, and underestimation of the complexity of cloud-to-cloud integration as key problems, explaining that these issues have led to a loss of validity in the traditional approach.

The article said this shift stems from changes in cloud decision-making driven by the specificity of AI infrastructure. Accordingly, it argued that companies need to move away from indiscriminate vendor expansion and proposed a transition to “intentional multicloud” as an alternative. It also identified three factors behind the change: first, differences in AI capabilities by vendor; second, tightening digital sovereignty and data residency requirements; and third, supply chain instability caused by shortages of GPU and specialized hardware.

AI workloads cannot be fixed to a single general-purpose technology stack, and workload placement should be based on available computing resources and the architectural fit of each model. Accordingly, it is important to determine workload deployment environments based on technical and economic requirements, which the article presents as the core of intentional multicloud. In this process, infrastructure and operations (I&O) leaders are required to shift roles away from a procurement-centered strategy and toward actively coordinating a distributed ecosystem.

In AI operations within such a distributed cloud environment, “Day 2 governance” is important, and operational requirements for continuously managing cost, performance, and security are emphasized. Without an integrated governance framework, operational complexity can increase and spending control can become difficult, while business units may attempt to use specialized resources by bypassing the IT department. As a result, the spread of “shadow AI” could follow, and the article presents as a challenge for I&O organizations the need to strengthen their strategic role so that multicloud becomes the foundation for AI innovation rather than a constraint.

In cloud environments, numerous structural changes are occurring, and multicloud is becoming an essential AI technology. Accordingly, changes in how and where workloads are deployed are underway, and companies are being asked to take a capability-based, intentional approach.

In this process, cloud service providers (CSP) are strengthening expertise across each layer of the AI stack, and the era of “all-in-one” cloud providers is fading. As a result, major data may exist only in one cloud, while certain large language models (LLM) and training clusters may exist only in another, creating an “AI gravity” phenomenon.

In such a distributed environment, it becomes more important to use each provider’s specialized capabilities without moving entire datasets, and high-performance intercloud connectivity is needed to do so. Regulatory changes are also accelerating this trend.

Countries are introducing sovereign cloud mandates that require data to be stored and processed within specific jurisdictions. As a result, global companies must meet different regional regulations, data residency requirements, and national security requirements, which makes multicloud essential for global enterprises.

With AI demand surging, cloud infrastructure supply constraints are widening and shortages of GPU and specialized hardware continuing, making it difficult for a single provider to reliably meet all workload requirements. Accordingly, the need for multicloud is growing as a means of securing computing resources and cloud bursting capabilities when needed.

The article identifies insufficient technological maturity as the reason multicloud was strategized in the past despite weak justification. But the situation has recently changed with the emergence of cloud networking technologies, integrated control plane technologies, and multi-data-center resource pooling technologies.

These technologies have made practical implementation of distributed architectures possible, improving the ease of moving workloads between different environments and applying management policies across different environments. They have also lowered the barriers to adopting integrated technologies, which are now serving as the foundation for smoother workload mobility and policy application across environments.

Executing an AI-based multicloud strategy requires more than execution conditions and basic capabilities; it requires the introduction of systematic and continuous management methods. In particular, cost management and optimization become more complex in multicloud AI environments, due to differences in cloud pricing policies, differences in cloud usage metrics, and data transfer costs. Accordingly, I&O organizations are tasked with forecasting data movement costs between AI services and primary storage, establishing granular data egress cost models, and supporting reductions in data transfer costs.

To address the cost area, securing automated FinOps tools and centralized cost visibility are cited as essential. Automated FinOps tools and centralized cost visibility help identify idle resources, optimize spending, and place workloads in the most economical and technically appropriate environments.

On the security side, fragmented security frameworks caused by distributing AI workloads across multiple vendors increase exposure to risk, making it necessary to evolve into an integrated identity-centered framework that applies consistently across all cloud environments. In addition, to ensure regulatory compliance and protect sensitive data, the article says identity and access management must be standardized on a federated model, and that vendor-independent encryption, cloud-to-cloud security posture management tools, proactive identification and management of third-party activity, and proactive identification and management of shadow AI activity are also required. These measures are intended to prevent new blind spots in risk coverage.

As conditions for improving operational efficiency, the article cites systematic management of AI assets and systems, along with comprehensive observability. To this end, a single information framework covering all AI resources is needed, as is an AI model lifecycle management policy to prevent idle resources from retired experiments from continuing to run. An integrated management approach is also needed to minimize the indiscriminate spread of agents, and telemetry from each vendor must be unified into a single observability layer. Key requirements include consistent performance metrics, distributed tracing, and a centralized alerting system, which help enable real-time visibility into complex cross-cloud AI environments and speed up problem resolution.

As changes driven by AI spread, multicloud is becoming essential, but simply adopting multiple clouds is not enough to secure competitiveness. Accordingly, the article presents workload placement in the optimal environment as an operational necessity, with cost, performance, security, and AI assets managed under a consistent framework. This also requires a change in the role of I&O organizations, expanding their responsibilities beyond cloud procurement to coordinating a distributed ecosystem. If this operational capability is secured, multicloud functions not as a source of complexity but as a foundation for AI innovation.

Source: IT DAILY · Yang Seung-gap
Original: https://www.itdaily.kr/news/articleView.html?idxno=241293

References

• https://www.itdaily.kr/news/articleView.html?idxno=241293