By 2026, the big cloud players (Google Cloud, AWS, Microsoft, Databricks) are aligning on the same message: the hard part of enterprise AI is no longer model performance, but how you operate AI agents at scale.

In 2024, most discussions were about which model to choose. In 2026, the decisive questions are:

• Who controls the business context the agent uses?
• How are permissions and identities managed?
• What traces and observability do we have on agent behavior?
• How do we manage and optimize unit inference cost over time?
• How easily can we switch models or providers if needed?

Each provider is reimagining its platform around these operational concerns:

• Google Cloud launched the Gemini Enterprise Agent Platform with managed runtime, memory, identity, logging, tracing and monitoring.
• Microsoft states that the bottleneck is no longer model power, but the shared enterprise context agents need to act inside business systems.
• AWS emphasizes continuous improvement from production traces and highlights the risk of “silent failures” that only appear later as customer complaints.
• Databricks notes that the visible “agentic loop” is only about 1% of the work; the remaining 99% is deployment, security, evaluation, observability, context and sharing.

The result is a shift from classic MLOps (versioning models, deploying endpoints, tracking a few metrics) to what many now call AgentOps: managing runtimes, memory, tools, permissions, traces, quality and costs as a coherent operating chain. In practice, this moves AI from a data science experiment into the operational core of the information system.

Running AI agents introduces a set of hidden operational and cost layers that go well beyond the model API price.

On the operations side, the 2026 stack now includes:

• Agent runtime management (sessions, multi-step chains, latency).
• Short- and long-term memory (storage, retrieval, governance).
• Action permissions and identity (who is the agent, acting on whose behalf, with what rights?).
• External tools and gateways (connecting to CRM, ticketing, databases, workflows).
• Tracing and observability (tokens, latency, sessions, invoked tools, error patterns).
• Continuous evaluation (quality scoring, human feedback, behavioral compliance).

Vendors are building around this:

• Google documents agent identity based on the SPIFFE standard, so agents can authenticate securely to cloud resources and other agents.
• AWS AgentCore Gateway turns existing APIs and services into tools with fine-grained access control and Model Context Protocol compatibility.
• Databricks MLflow 3 unifies tracking, evaluation and observability with real-time traces and automated scorers on production samples.
• CNCF Inference Gateway routes traffic by model, LoRA adapters and endpoint health to improve accelerator utilization.

On the cost side, AI spend becomes clearly composite. According to Flexera’s State of the Cloud 2026:

• 58% of organizations already use public cloud GenAI services; 45% say they use them extensively.
• 73% operate in hybrid mode.
• 49% now use unit economics to connect cloud spend to business outcomes.
• Estimated IaaS/PaaS waste has climbed back to 29%.
• 64% measure cloud more by business value delivered than by cost efficiency alone.

In this context, AWS’s own AgentCore pricing illustrates that costs accumulate around the model: gateway calls, short-term memory, long-term storage, retrieval, observability and more. The practical budgeting question becomes:

What is our full cost per useful agent?

That full cost typically includes:

• Model usage (tokens, fine-tuning, adapters).
• External tools and API calls.
• Memory (storage and retrieval).
• Logging, tracing and observability.
• Security, guardrails and identity management.
• Context data preparation and governance.
• Human time for evaluation, remediation and continuous improvement.

This is why FinOps is evolving. Flexera now highlights an AI Cost Management layer that spans applications, agents, models, data platforms and compute. For CIOs and CFOs, the conversation is moving from “how much does a prompt cost?” to “what is the cost per service, per use case, per workflow, per team, per customer?”

In Europe, AI cloud strategy is increasingly shaped by sovereignty, regulation and provider mix, not just technical features.

Two regulatory milestones matter:

• On June 3, 2026, the European Commission adopted a proposal for a Cloud and AI Development Act to strengthen Europe’s cloud and AI ecosystem, investments and infrastructure.
• The AI Act becomes fully applicable from August 2, 2026, with transparency rules and a framework that increases responsibilities for both providers and deployers.

Reuters reports that European groups such as Siemens, Renault, Orange and ChapsVision are already:

• Multiplying providers to reduce dependency risk.
• Reacting to access restrictions on some U.S. services.
• Paying close attention to token costs as agents automate more tasks and budgets are consumed faster than expected.

In this context, sovereignty is less about isolation and more about flexibility and fallback:

• Which components must be operable in-house if needed?
• Which tools and runtimes must remain substitutable across providers?
• Which context data must not be locked into a proprietary runtime?
• How quickly can a critical agent switch models or clouds if access rules change?

OVHcloud’s plan to train frontier models, with an estimated cost of €150–200 million for this technology cycle (well below the often-cited €1 billion), illustrates how AI cloud sovereignty is moving from policy discussions into concrete product and infrastructure strategies.

For enterprise architecture teams, this leads to a practical design question: do we standardize on a single agent runtime from one hyperscaler, or do we build a portable layer across multiple clouds and frameworks? The evaluation criteria shift from “who has the best model?” to:

• Context portability (can we move our knowledge and prompts?).
• Observability quality (do we see behavior across providers?).
• Control granularity (permissions, identity, guardrails).
• Cost visibility (clear unit economics per agent and per use case).
• Fallback capability (time and effort to migrate a critical agent).

In short, once agents are embedded in business processes, vendor dependency becomes a risk variable. Organizations that treat AI operations as part of their broader cloud steering, financial control and governance strategy will be better positioned to adapt as the regulatory and provider landscape evolves.