September 2026 Summaries

Stealthium has joined the Vultr Cloud Alliance to provide GPU-level runtime observability and security for organizations running AI workloads on Vultr Cloud GPU infrastructure. As GPU adoption grows, the partnership addresses limitations in conventional monitoring and CPU-focused security tools by offering telemetry, workload context, health monitoring, utilization insights, and threat detection within GPU environments. Stealthium, an NVIDIA Inception Program member, can be deployed as an optional single agent per host without changes to the Vultr platform, enabling infrastructure, platform, and security teams to investigate issues faster, identify potentially malicious activity, optimize GPU use, and manage AI operations at scale. The partnership was announced September 30 at Cloud & AI Infrastructure Singapore, where the companies are presenting their combined capabilities.

Vultr has added Wasabi Technologies to its Cloud Alliance, combining Vultr’s compute infrastructure with Wasabi’s S3-compatible object storage through direct private connectivity. The partnership lets organizations independently select and scale compute and storage while aiming to lower data-access costs through Wasabi’s pricing model, which includes no egress or API request fees, and Vultr’s included bandwidth and predictable egress pricing. Customers can retain data in Wasabi while running applications on Vultr Cloud Compute, Bare Metal, Kubernetes Engine, or Cloud GPU services, supporting workloads such as backup and disaster recovery, long-term retention, cloud applications, media libraries, and AI data processing. Wasabi complements Vultr’s existing object storage offerings and expands multicloud storage options for customers.

Agentic AI is increasingly being deployed with access to enterprise systems and the authority to take multi-step actions, creating a need for security controls that operate independently of an agent’s reasoning or prompt-level safeguards. NVIDIA’s Agent Safety Platform addresses this through OpenShell, an open-source sandboxed runtime that governs agent access to files, networks, credentials, tools, and model endpoints; BlueField-4 hardware for isolated threat detection and policy enforcement; and Vera CPUs for orchestration and processing workloads. Vultr is evaluating these components for its GPU cloud infrastructure to provide customers with runtime-enforced least privilege, isolation, authorization, and auditable logs, which could support compliance reviews and allow compromised agents to be quarantined without disrupting broader systems. The company also supports the Open Secure AI Alliance’s call for transparent, inspectable security boundaries and argues that trustworthy infrastructure can enable both greater AI autonomy and stronger enterprise security.

Vultr and VAST are expanding their Cloud Alliance partnership to address the infrastructure requirements of production AI, combining Vultr’s global cloud compute and GPU resources with VAST’s AI data platform. The collaboration aims to support workloads such as large-scale model training, real-time inference, and data-intensive AI pipelines by placing scalable compute and data infrastructure closer together. As agentic AI adoption grows, the partnership emphasizes timely, controlled, secure, and accountable access to enterprise data, allowing customers to independently scale computing, GPU, and data layers while maintaining responsive data access and enterprise governance.

The State of AI in Platform Engineering 2026 report examines how AI-driven and agentic development are affecting software delivery, platform operations, infrastructure investment, and organizational strategy. Drawing on survey data and perspectives from platform engineering leaders, it evaluates whether faster development output is producing measurable business value and return on investment. The report also explores how platform teams are adapting to support AI-enabled development, the approaches used by leading organizations, the areas where AI is delivering value or exposing gaps, and the priorities and challenges shaping platform engineering in the coming year.

Cycle.io has introduced a Model Context Protocol server for its Vultr integration, allowing MCP-compatible AI assistants and coding tools to access and manage applications and infrastructure through natural-language requests. The capability can retrieve operational data such as workload health, events, metrics, logs, networking details, DNS information, and load balancer traffic, helping teams investigate production issues without moving among separate dashboards and tools. Subject to Cycle.io’s existing capability controls and API key permissions, assistants can also perform approved actions including deployments, container and environment restarts, configuration updates, DNS and ingress management, and workload migrations across servers, data centers, or cloud providers. At launch, the server provides 18 tools, consisting of 10 read-only tools, 7 mutation tools, and one command-execution tool, while enabling platform teams to maintain governance over developer self-service.

SoftwareOne has joined the Vultr Cloud Alliance to combine Vultr’s enterprise cloud infrastructure with SoftwareOne’s advisory, FinOps, migration, optimization, and managed-services capabilities. The collaboration aims to help multicloud enterprises reassess where workloads run, addressing situations in which applications remain on their original infrastructure despite higher costs or weaker performance. Vultr’s VX1 Cloud Compute is highlighted as an x86-compatible option intended to improve price-to-performance without requiring application rearchitecture, while Vultr Cloud GPU supports demanding AI workloads. Both companies’ participation in the FinOps Foundation ecosystem supports a shared emphasis on incorporating workload economics and infrastructure choices earlier in deployment planning, extending optimization beyond rightsizing and discounts.

Vultr has joined the Linux Foundation’s FinOps Foundation and Tokenomics Foundation to collaborate on open practices for managing and maximizing the value of cloud and AI investments. The company plans to contribute its perspective as an independent cloud provider, emphasizing high-performance infrastructure, customer choice, predictable costs, and reduced dependence on proprietary cloud architectures. The FinOps Foundation focuses on cloud financial operations and technology business value, while the Tokenomics Foundation extends similar principles to assessing AI infrastructure, production, consumption, and investment value. These memberships also support Vultr’s work with financial services organizations, which must balance innovation with cost control, resilience, regulatory compliance, and infrastructure governance, including requirements such as the EU Digital Operational Resilience Act.

OpenFold3 supports prediction of 3D structures for proteins, antibodies, nucleic acids, complexes, and small-molecule ligands, but deploying it in production can require substantial AI infrastructure expertise. MindWalk independently validated AMD Inference Microservices for OpenFold3 on Vultr Kubernetes Engine using AMD Instinct MI325X GPUs, demonstrating a Kubernetes-native approach to simplifying protein structure prediction workloads. Available through the VKE Marketplace, the offering is intended to reduce manual work involving deployment, networking, HTTPS configuration, and model serving for healthcare and life sciences AI applications.

Enterprise AI infrastructure is shifting from GPU-dominated clusters built for model training and inference toward integrated CPU and GPU systems designed for agentic workloads. Unlike chatbot-style requests, AI agents must decompose tasks, coordinate other agents, access databases and APIs, execute applications, manage context and permissions, validate results, and operate within security policies, creating substantial CPU requirements alongside GPU-based model reasoning. The text argues that compute ratios may move from roughly one CPU per four to eight GPUs in generative AI deployments toward a near 1:1 balance, reflecting a broader data-center architectural change. CPUs support agent orchestration, data preparation and streaming, secure sandbox environments for untrusted code, and the increased demand AI agents place on conventional cloud services such as ERP, SaaS, analytics, microservices, and API backends. It presents Vultr VX1 Cloud Compute, powered by AMD EPYC processors, as a platform for these needs through high core density, AVX-512 data-processing support, nested virtualization for isolated microVM sandboxes, and high-throughput networking.

Modelplane v0.3 adds Vultr Kubernetes Engine as a supported inference cluster provider, allowing users to define scalable GPU clusters through YAML manifests that specify regions, node pools, instance types, and scaling limits while automatically provisioning networking, Kubernetes components, GPU infrastructure, and the inference layer. The open-source platform aims to provide a uniform operating model across cloud providers including Vultr, GKE, EKS, and AKS, enabling model deployment manifests to remain unchanged when underlying hardware changes. The release also adds end-to-end support for Anthropic’s Messages API through vLLM, allowing compatible clients such as Claude Code to use models hosted on users’ own Vultr GPU clusters by mapping Anthropic model tiers to locally served models. Users can follow Modelplane’s Vultr getting-started materials or review release notes, while unsupported providers and instance classes can be requested or contributed through GitHub.

Qdrant has released Qdrant-Fineweb-10B, a public benchmark dataset containing dense and sparse embeddings for 10 billion FineWeb documents and exact ground-truth results for 120,000 queries up to a depth of 1,000, enabling developers to evaluate vector search systems under production-scale conditions. Intended for applications such as retrieval-augmented generation, semantic search, recommendation engines, and AI agents, the dataset addresses limitations of smaller or synthetic benchmarks by supporting measurement of retrieval accuracy, indexing, memory, sharding, and scaling behavior on real-world data. Qdrant generated the embeddings using Alibaba’s gte-multilingual-base model on Vultr infrastructure, combining parallel CPU tokenization, GPU processing, and object storage to sustain roughly 19,000 rows per second per node. The full workflow processed the corpus in about five days, producing approximately 25 TB across 500,000 files, and highlights the importance of coordinating compute, preprocessing, and storage for billion-record AI workloads.