How to Choose AI Infrastructure Across Cloud, Sovereign Systems, and Agent Platforms

What Happened

Recent announcements emphasize where AI workloads run and how they are governed more than new GPU specifications. Microsoft and Intel describe a sovereign deployment stack spanning Azure Local on Intel Xeon 6, Foundry Local, Intel AMX and OpenVINO. Supported Azure Confidential VM configurations can use Intel TDX and hardware-backed attestation to help protect workloads while in use [5]. The Barcelona Supercomputing Center’s operational AI Factory Accelerator Programme gives startups access to sovereign computing resources; many participants also belong to NVIDIA Inception [4].

On cloud platforms, Google has put Claude Opus 5.5 and Sonnet 5.5 in Antigravity alongside its own models, with model switching across its development tools. Partner-model usage is billed separately from pooled Gemini Enterprise token quotas [2]. Google’s Gemini for government services is in preview, combining agents, connectors, approvals and governance controls for public-sector workflows [3]. Databricks has introduced agent memory on Agent Bricks, although the available announcement does not specify its behavior or operating controls [1]. These announcements do not establish comparable new hardware or service capabilities for AMD, AWS, Snowflake or Cloudflare.

Why It Matters to Businesses

Infrastructure selection is no longer just a GPU price comparison. Buyers must decide whether data and inference can leave a controlled environment, how agents access existing systems, and whether model choice changes costs or audit obligations. Google’s government preview illustrates a useful pattern: connectors respect existing permissions, while consequential actions can require human approval [3]. Its Antigravity pricing also shows why a model picker needs budget controls, not just developer convenience [2].

Kimbodo Engineering Perspective

We would separate compute, model access, agent orchestration and data governance in the architecture. Local deployment may improve control but adds capacity planning, patching and operational responsibility. Cloud services can reduce that burden, but require careful treatment of access, data movement and usage charges. Hardware-backed isolation is valuable for a defined threat model; it does not replace application authorization or review of agent actions [5]. We would evaluate Databricks agent memory only after testing its retention, deletion, access and retrieval behavior against the application’s requirements [1].

How We Would Implement It

• Classify workloads by data sensitivity, residency, latency and required human oversight; place each on cloud or customer-controlled infrastructure accordingly [3][5].
• Put models behind an application-level routing layer with per-model budgets, logging and evaluation tests. Test higher-capability models on complex work and lower-cost models on bounded tasks rather than assuming one model fits every step [2].
• Connect agents to systems of record through scoped identities. Require approval before writes, disclosures or other consequential actions, and retain citations and decision logs [3].
• Benchmark candidate GPU and CPU deployments on the actual inference workload, including throughput, latency, utilization and total operating cost. Do not infer comparative performance from these announcements.

Risks, Costs and Security

Watch for partner-model charges outside existing quotas and charges from background model activity [2]. For sovereign deployments, verify the exact supported confidential-computing configuration and attestation process rather than treating a product label as proof of isolation [5]. For agents, test permission propagation, prompt-injection resistance, approval boundaries and memory lifecycle before production use. Preview status and unspecified feature details warrant a pilot, not an assumption of production readiness [1][3].

Sources

1. [1] Introducing Agent Memory on Databricks
2. [2] Claude Opus 5.5 and Claude Sonnet 5.5 are now available in Google Antigravity
3. [3] Introducing Gemini for government services
4. [4] Spain-Based Startups Advance Regional Innovation With Barcelona Supercomputing Center
5. [5] Designing for sovereign AI: A shared foundation from edge to cloud