Executive Summary: ISG Provider Lens™ Multi Public Cloud Services - U.S. 2025
15 Dec 2025
by Shashank Rajmane
The individual quadrant reports are available at:
ISG Provider Lens™ Multi Public Cloud Services - Managed Services - Large Accounts - U.S. 2025
ISG Provider Lens™ Multi Public Cloud Services - Managed Services - Midmarket - U.S. 2025
ISG Provider Lens™ Multi Public Cloud Services - SAP HANA Infrastructure Services - U.S. 2025
AI is being integrated into almost all public cloud engagements to improve productivity and efficiency
Public cloud platforms form the core of enterprise AI ecosystems, offering the scalability, elasticity and specialized infrastructure needed to train and deploy large models efficiently. ISG research shows that combining on-premises control with cloudbased acceleration enables organizations to integrate AI-powered intelligence into existing workflows and streamline their operations. Enterprises have been able to reduce development complexity, accelerate time to value and scale innovations from predictive analytics to autonomous operations by leveraging cloud-native AI services, pretrained models and GPU-optimized instances.
To support these growing AI demands, major hyperscalers such as AWS, Azure and Google Cloud have strengthened partnerships with semiconductor vendors and foundational model providers to deliver enterprises with AI-ready cloud services. For example, AWS integrates NVIDIA GPUs into high-performance instances to enable quick training and inference of large-scale AI models; Azure offers accelerated virtual machines (VMs) that allow enterprises to efficiently run complex AI workloads at scale; and Google Cloud combines its own TPUs and NVIDIA GPUs into a single ecosystem to deliver a powerful environment for enterprises to deploy and scale multi-modal AI applications seamlessly. Beyond the Big Three, other cloud service providers (CSPs) are also carving out distinct strategies to strengthen their AI-infrastructure portfolios. Oracle Cloud has deepened its partnership with NVIDIA to deliver GPU-accelerated OCI Superclusters purpose-built for GenAI training and inference at scale. Similarly, in Europe, OVH has progressed in establishing itself as an AI-ready CSP while delivering strong data sovereignty controls.
However, as enterprises scale AI initiatives across hybrid and multicloud environments, operational complexity has become the new frontier of transformation. Running AI workloads in distributed environments introduces challenges, where models are trained in one environment, fine-tuned in another and deployed across multiple inference endpoints, each with unique data residency, compliance and performance constraints. These challenges extend far beyond compute performance, encompassing security, cost governance, monitoring and orchestration.
Below are some of the key trends that ISG observed in the last four quarters:
Growing but mixed demand for cloud services driven by AI workloads: AI is now the engine behind overall cloud consumption, but growth is uneven and highly architectural. Training models produce seasonal peaks, inference workloads are often always-on and data preparation involves steady analytics with occasional spikes in cloud resource consumption. The practical approach is to segment these profiles up front, place heavy training close to curated data and run latencysensitive inference nearer to users or machines. Teams that codify these placement rules as policy avoid the slow bleed of egress, rebuilds and cross-region drift. The net effect is a clean operating model in which capacity planning, data adjacency and developer workflows align.
Cost optimization remains a top priority, and enterprises want it now: AI workloads are becoming table stakes for hybrid cloud architecture, increasing complexity and magnifying inefficiencies. The fastest outcomes come from treating cost as a design input rather than a month-end audit. To prevent waste before the product goes live, enterprises need to include budget as code in pipelines, tie rightsizing to deployment checks and map commitment plans to real usage patterns. On the model side, quantization, distillation and selective caching can reduce GPU minutes without hurting accuracy targets. For inference, moving preprocessing to the CPU and reserving accelerators for the tight loop improves utilization. The language of success becomes unit costs, which the business understands. This includes cost per answer or cost per transaction, reviewed alongside reliability and latency.
GenAI adoption remains in native stages: Incorporating GenAI into the existing mix of AI workloads remains nascent in enterprise adoption of AI, while scale requires discipline and high technology maturity. Many enterprises invest heavily in this space, with several funding multiple PoCs, but moving to production-grade GenAI initiatives only when the outputs are highly accurate. Naturally, these are very small in number. Enterprises that move fast set up a small PoC factory with strict guidelines and a simple rule for promotion to production: problem framing before model selection, red team testing before go-live and every use case carries an owner accountable for data, risk and budget. This structure enables teams to learn inexpensively, retire weak ideas quickly and concentrate investment where probability of success is high.
Focus on FinOps for AI and AI for FinOps: Sustainable FinOps and GreenOps are maturing and becoming integrated into FinOps platforms, while carbon currency is included in the same dashboard that drives infrastructure consumption decisions. Training workloads that can be deferred are scheduled into cleaner grids, while inference jobs use right-sized instances with power caps and storage teams clear duplicate or stale datasets on a cadence that optimizes AI workloads. Procurement departments now include energy disclosure in their vendor selection processes, while engineers receive a carbon budget just like a spend budget, with exceptions handled through the same approval flow.
AI application for cloud operations: Almost all providers have integrated GenAI and agentic AI technologies into their cloud management platforms, built with security and compliance guardrails. They now embed GenAI models to automate cloud governance, workload optimization and service orchestration. These AI systems enable predictive resource scaling, automated threat detection, drift monitoring and intelligent log analysis across environments, improving overall efficiency while automating the most mundane tasks and freeing engineers to focus on more critical tasks. Additionally, every action the agents take writes back into a knowledge base, so fixes become reusable runbooks, not just memory. The agents have also included confidence scoring, dual approvals for risky steps and automatic postincident reviews by keeping humans in the loop without slowing down the process.
The flip side of using AI: The primary downside of AI adoption is its potentially high and unpredictable infrastructure costs. Enterprises should avoid these surprises by making economics an explicit non-functional requirement. Before committing to a model, they should simulate traffic, concurrency and latency targets to estimate the accelerator share and memory pressure, then select an architecture and serving patterns that match the curve. Service providers are also guiding enterprises to control costs by keeping data close to compute to cut egress, trimming feature pipelines that add cost and preferring smaller models where possible.
Access to the full report requires a subscription to ISG Research. Please contact us for subscription inquiries.
Page Count: 14
Categories
QUESTIONS?
To purchase this product or for more information, please contact your account manager: