September 2026 Summaries

CoreWeave has introduced ARENA Pass, a program for qualified enterprise teams new to the company that provides a focused evaluation of AI workloads before production deployment, including access to CoreWeave Forge and engineering support. The program is intended to help customers assess factors beyond benchmark performance, such as latency, throughput, storage, scaling, model versioning, infrastructure economics, and operational support, using their own applications and success criteria. Examples cited include You.com’s evaluation of training, inference, quota elasticity, and tail latency; a voice AI team’s testing of 1,200 concurrent sessions; and Periodic Labs’ measurement of high-volume storage performance. Participants work with CoreWeave engineers through its Direct-to-Expert service and can use Forge’s workflow for deploying, observing, curating data, improving models, and evaluating results, while also testing CoreWeave Cloud infrastructure where relevant. At the end of an engagement, teams receive findings and recommendations that may support a production launch, further testing, or changes to their technical approach.

CoreWeave has announced limited availability of NVIDIA Vera Rubin NVL72 systems on its cloud platform, deploying hundreds of Rubin GPUs across multiple regions for selected production customers. Cognition, the developer of the autonomous coding agent Devin, is using the platform for inference, training, fine-tuning, reinforcement learning, and experiment tracking, and reports up to a 4.8-fold increase in total token throughput for SWE-2 inference and a 3.8-fold increase in reinforcement-learning output token throughput compared with NVIDIA GB200 NVL72 systems under matched interactivity. The liquid-cooled NVL72 architecture combines 72 Rubin GPUs, 36 Vera CPUs, NVLink 6 connectivity, and high HBM4 memory bandwidth to support large models and intensive agentic workflows involving long contexts, tool use, code execution, and iterative reasoning. CoreWeave says the systems integrate with its Kubernetes, inference, storage, and management services, allowing customers to run workloads soon after deployment without building underlying infrastructure, and is offering capacity-planning briefings and evaluation access.

AI is presented as a catalyst for discovery by helping people identify connections across vast, fragmented bodies of knowledge that would otherwise remain hidden across disciplines, languages, and generations. The passage argues that this capability can support both major advances in areas such as medicine and climate and everyday innovation by enabling more people, regardless of technical background, to turn ideas into practical work. It emphasizes that AI-generated insights still require human judgment, experimentation, evidence, and persistence. Because AI systems require continual observation, testing, iteration, and improvement after deployment, the author argues that traditional cloud infrastructure and disconnected tools are insufficient. CoreWeave describes its announcements at Fully Connected, including capabilities involving Weights & Biases, OpenPipe, marimo, NVIDIA hardware, partners, and production customers, as parts of an integrated platform intended to connect the AI development lifecycle and accelerate innovation.

CoreWeave has launched the CoreWeave Partner Network, a curated group of technology providers whose integrations have been tested on its cloud infrastructure under production conditions, aiming to reduce the complexity and risk of moving AI workloads across cloud environments. The network spans infrastructure, data services, independent software vendors and security, and models and inference, with partners including VAST Data, Dell, AMD, IBM, ClickHouse, Datadog, Grafana, CrowdStrike, Zscaler, Pinecone, Redis, Reflection, and Red Hat. CoreWeave says the initiative formalizes longstanding collaborations and gives customers visibility into tools already validated for use across training, data management, observability, security, and inference, rather than requiring them to assess large marketplaces or build integrations independently. The company cites deployments by IBM, Cohere, Mistral AI, and Decart as examples of partners and customers operating AI workloads at scale on its NVIDIA-based infrastructure. It also plans to add a shared web-search integration for AI agents through Exa, Parallel Web Systems, and You.com, designed to give open-weight model users access through a single CoreWeave integration without exclusive vendor lock-in.

CoreWeave has introduced Agent Lens, an observability platform within CoreWeave Forge designed to analyze AI agent production traces and identify user intents, recurring failures, capability gaps, regressions, and operational risks at scale. Using OpenTelemetry data, it automatically groups semantically related conversations and failures, links clusters to supporting trace evidence, and prioritizes issues by factors such as frequency, trend, workflow impact, safety, and compliance concerns. The platform also enables teams to create custom LLM-based evaluators aligned with business-specific quality standards, transform real production failures into reusable test cases, and compare candidate prompts, models, tools, or code changes against evaluation sets before deployment. Agent Lens can connect with coding agents through Skills and MCP to help automate investigation and testing while retaining human approval over releases. Built for enterprise workloads, it is available in public preview on CoreWeave’s hosted cloud, with dedicated and on-premises options planned, and its Insights and online LLM judge features free through the end of 2026.

CoreWeave plans to offer NVIDIA’s upcoming Vera CPU, an Arm-based processor designed for AI-agent workloads such as sandboxed execution, reinforcement-learning environments, post-training pipelines, tool calls, and data preprocessing, which increasingly rely on CPU capacity alongside GPUs. Vera includes 88 custom Olympus cores and is paired with Rubin GPUs in NVIDIA’s Vera Rubin NVL72 systems, while standalone bare-metal nodes on CoreWeave will feature two CPUs, 1.5 TB of RAM, 15.36 TB of NVMe storage, and an 800 Gbps BlueField-4 DPU. CoreWeave says its testing showed more than threefold faster agent sandbox startup than an alternative x86 CPU and that a rack can support more than 11,000 concurrent environments using 128 Vera CPUs and 11,264 cores. The company also highlights predictable performance through spatial multithreading, DPU offloading, integration with its Kubernetes, serverless, and management services, and flexible committed, serverless, and spot-capacity purchasing options. Vera will join AMD EPYC and Intel Xeon offerings in CoreWeave’s bare-metal CPU portfolio, enabling customers to select hardware based on changing workload requirements without moving to a separate platform.

CoreWeave reports earning a Platinum rating in SemiAnalysis’ ClusterMAX assessment for the third consecutive major evaluation, making it one of two Platinum-rated providers in ClusterMAX 3.0 and the only cloud provider to maintain that status across all three evaluations. The assessment evaluated 77 providers across 10 categories, testing current-generation AI infrastructure, including accelerators, high-bandwidth networking, storage, monitoring, Kubernetes and Slurm support, stress performance, and fault recovery. CoreWeave attributes its result to an integrated operational stack featuring automated provisioning, hardware validation, fleet monitoring, failure remediation, and GPU straggler detection, which identifies degraded components that can slow distributed workloads without producing conventional errors. The company says it provisions roughly 10,000 GPUs weekly and uses operational data, custom diagnostics, predictive analysis, and upstream collaboration with hardware manufacturers to improve reliability before systems reach customers. It presents these capabilities as reducing customers’ infrastructure-management burden, improving GPU utilization, accelerating deployment, and supporting reliable performance as AI hardware and workloads evolve.

ARIA is a research agent embedded in Weights & Biases that uses existing experiment data, production traces, page context, and selected runs to investigate questions, create durable reports or workspaces, and prepare follow-up experiments subject to researcher approval. Its development shifted from a chat-centered tool toward a persistent research process designed to preserve evidence, support collaboration, recover from failures, and distinguish between the agent’s decisions and the platform’s authorization controls. ARIA represents work as durable, resumable “Turns” that retain prompts, history, configuration, and sandbox state, allowing long-running investigations to continue after a browser session closes while separating service operations, agent reasoning, tools, and isolated execution environments. Production failures are replayed and evaluated using W&B Weave for tracing and W&B Agent Factory for controlled ARIA-specific comparisons, although exact reproduction also requires stable external project data. The team is exploring a constrained form of agent-assisted improvement in which ARIA can help diagnose failures and propose focused changes to prompts, tools, scripts, or skills, but humans retain responsibility for defining correctness, reviewing evidence, approving consequential actions, and deciding whether changes ship.

CoreWeave announced the general availability of Mission Control MCP, a hosted read-only Model Context Protocol server, and expanded its Mission Control Agent within the CoreWeave console to help teams investigate and optimize AI workloads. Built into CoreWeave Cloud, Mission Control combines application behavior with detailed infrastructure telemetry spanning GPUs, nodes, racks, networking, cooling, and hardware events, while using fleet operations data and historical performance patterns to identify likely bottlenecks and recommend changes. MCP enables compatible AI clients such as Claude Code, Cursor, and Codex to query metrics, logs, alerts, dashboards, documentation, and resource states using an organization-scoped API token, reducing the need to manually collect evidence from multiple tools. The console-based Agent analyzes deployment telemetry to identify performance constraints and estimate the value of remediation, illustrated by an inference example in which growing context lengths created KV-cache pressure, reduced cache-hit rates, and lowered throughput; it recommended memory-saving quantization and reconsidering GPU model distribution. CoreWeave also describes integrations with existing observability tools and Weights & Biases, aiming to connect research, development, and infrastructure operations so teams can determine whether performance issues stem from workloads, configurations, or underlying systems.

CoreWeave presents new identity and encryption capabilities intended to help enterprises deploy sensitive AI workloads while retaining existing security controls across multi-cloud environments. CoreWeave IAM federates with customers’ identity providers through SAML and OIDC, continuously synchronizes users and groups, and applies familiar RBAC and ABAC policies across CoreWeave services, reducing manual account provisioning and supporting centralized audit trails. Its SUNK User Provisioning feature extends this approach to Slurm-on-Kubernetes clusters by automatically creating POSIX, SSH, and Slurm identities from federated user data. Remote Key Encryption, planned for availability later in the year, is designed to encrypt CoreWeave AI Object Storage data using keys that remain in customers’ existing KMS, HSM, or secrets-management systems, with CoreWeave holding only ciphertext and not accessing plaintext keys. The company argues that integrating with established identity and key-management infrastructure can simplify compliance reviews, preserve least-privilege access, and accelerate production deployment of regulated AI applications.

CoreWeave reported MLPerf Inference v6.1 Datacenter Closed, Available-category results across NVIDIA HGX B200, HGX B300, GB200 NVL72, and GB300 NVL72 systems for Qwen3-VL-235B-A22B, DeepSeek-R1-671B, GPT-OSS-120B, and Llama 2 70B, emphasizing that inference throughput and infrastructure requirements differ substantially among multimodal, mixture-of-experts reasoning, and conventional language models. The company said its GB300 NVL72 submission delivered the highest cloud-provider server throughput for Qwen3-VL, while its DeepSeek-R1 result showed a 19.8% derived per-GPU server-throughput increase over its prior v6.0 submission. It also claimed leading GPT-OSS-120B per-GPU throughput, including more than 1.16 million tokens per second in the server scenario on a GB300 NVL72 rack, and the highest cloud-provider throughput for Llama 2 70B, with 944,902 tokens per second in server mode and 1,136,100 in offline mode. CoreWeave attributed the results to production infrastructure, ongoing software and scheduling optimization, topology-aware workload placement, and operational tools, while noting that per-GPU throughput is a derived metric not verified by MLCommons.

CoreWeave AI Object Storage has introduced cross-region write acceleration and an Archive storage class to reduce operational overhead for AI workloads whose compute and data may reside in different regions. Its Local Object Transport Accelerator (LOTA) caches object reads on local NVMe within GPU and CPU nodes, enabling high-throughput, low-latency reads from buckets in any CoreWeave region without requiring application changes or duplicate datasets. Cross-region write acceleration allows standard S3 writes to be durably acknowledged in the local compute region while data migrates asynchronously to a selected remote bucket location, maintaining a single namespace, unified policies, and consistent object visibility. Archive complements existing Hot, Warm, and Cold tiers by providing lower-cost long-term retention for infrequently accessed data such as older checkpoints and datasets, without retrieval, request, early-deletion, tiering, or egress fees, though it has longer time-to-first-byte and additional write latency. Together, the capabilities are intended to let organizations run jobs where GPU capacity is available, preserve a unified data environment, and automate retention policies across the AI training and evaluation lifecycle.

CoreWeave AI Object Storage offers Local Object Transport Accelerator (LOTA), cross-region write acceleration, and Archive storage as configurable features intended to improve GPU read performance, reduce write latency, and lower costs for inactive data without requiring application rewrites. LOTA uses node-local NVMe caching through a dedicated endpoint, with multipart uploads and pre-staging via ranged HEAD requests helping improve first-epoch read performance for large datasets, model weights, and checkpoint restores. Cross-region write acceleration acknowledges writes locally while synchronizing metadata across regions and migrating object data asynchronously, enabling a shared namespace while reducing checkpoint commit delays. Archive can be selected directly for long-term objects or applied automatically after a configurable inactivity period of at least 60 days; archived objects remain accessible through standard GET requests but have slower initial response times and lower throughput, and can be copied back to STANDARD before planned bulk reads. The guidance proposes combining these features into checkpoint-retention policies and monitoring storage tiers, usage, and costs through CoreWeave Mission Control.

CoreWeave describes its approach to deploying NVIDIA Vera Rubin NVL72 systems across multiple racks, where hundreds of GPUs, networking components, power systems, and liquid-cooling infrastructure must operate reliably as a unified distributed AI platform. Its process automates hardware onboarding, firmware and configuration management, and the monitoring of compute, cooling, and power through tools including Mission Control, Racky, and Valvey. Validation progresses from individual nodes to full racks and multi-rack clusters, using diagnostics and realistic training and inference workloads to detect underperforming GPUs, connections, thermal issues, and other potential stragglers before production. For inter-rack communication, CoreWeave uses a two-tier, non-blocking RoCE fabric with multi-rail and multi-plane connectivity, designed to provide high bandwidth, redundancy, and scalability up to roughly 128,000 GPUs. By testing the entire system under sustained and variable workloads, the company aims to provide customers with a consistent, observable pool of Vera Rubin compute whose rack boundaries are effectively invisible to applications.

Production agentic inference requires teams to decide what latency outcome to optimize for, whether to route requests by availability, session, or reusable prefix cache, how to scale capacity, and whether to measure cost per token or per completed task. The appropriate choices depend on workload characteristics such as prefix stability, chain length, burstiness, interactive versus asynchronous use, traffic predictability, and the team’s willingness to operate infrastructure. Stable and long shared prefixes can benefit from KV cache-aware routing, while mutable prompts reduce caching value, and warm capacity can avoid cold-start and cache-rebuild penalties during bursts. The piece compares fully managed endpoints, which prioritize consumption-based pricing and minimal operational responsibility for evolving workloads; managed configurable deployments, which provide control over models, routing, and scaling for established latency-sensitive traffic; and self-managed systems, which offer maximum flexibility but require full operational ownership. It argues that OpenAI-compatible interfaces make movement among these models easier as an agent’s workload becomes better understood.

AI cloud platforms differ from simple GPU-rental businesses by combining computing capacity with the software, operations, networking, security, orchestration, observability, and technical expertise needed for customers to develop and deploy AI products. While large hyperscalers and frontier labs may manage raw hardware internally, most enterprises require integrated environments tailored to AI workloads and supported across connected data centers. The passage argues that a diversified customer base across industries can indicate platform depth and durability, citing companies such as Jane Street, Zonos, and Mercado Libre as examples of organizations building production systems on CoreWeave. Although GPU supply, power availability, and access to NVIDIA chips remain essential for market entry, the long-term value of an AI cloud provider is presented as residing in the tools and services built above the hardware rather than the hardware alone.

Published by Weights & Biases and CoreWeave, the post examines OpenAI’s GPT-6 Astra, introduced in September 2026 as a successor to GPT-5.6 Sol, comparing its coding, agentic-task, token-efficiency, and cost performance with Sol and Anthropic’s Claude Fable 5.1. It reports that Astra leads on several vendor-reported benchmarks, including Terminal Bench 4.0, DeepSWE, and Terminal-Bench Science, while Fable leads on Humanity’s Last Exam with tools and the Artificial Analysis Intelligence Index; Astra’s near-perfect ARC-AGI-3 result also illustrates concerns that established benchmarks are becoming saturated and less able to distinguish frontier models. Although Astra and Fable have identical listed standard token prices, the post emphasizes that cost per successful task depends on total tokens, retries, and reasoning used rather than benchmark scores alone. Using Weights & Biases Weave, the authors provide a reproducible evaluation setup that prompts all three models to generate Python solutions for the same 30 CodeContests problems, executes their code against public test cases, and records correctness, execution latency, token usage, and estimated API costs. In that limited experiment at low reasoning effort, Astra passed 97% of tasks at an average 4.48 cents each, Sol passed 90% at 3.17 cents, and Fable passed 77% at 8.64 cents, while code execution latency was similar; the post concludes that broader evaluations should account for end-to-end reliability, iterations, time, and expense on real-world projects.

As open-weight models such as Kimi K3, DeepSeek V4, Qwen 3.8, and GLM-5.3 approach the performance of leading closed systems, enterprises are increasingly evaluating models based on inference cost, scalability, and operational control rather than benchmark rankings alone. The text argues that self-hosting appropriately sized open models can substantially reduce per-token costs compared with frontier APIs, with break-even estimates around 5 to 10 million tokens daily, while allowing organizations to use techniques such as quantization, batching, specialized hardware, and model distillation. Open weights also provide greater control over data handling, fine-tuning, regulatory guardrails, model versioning, and long-term availability, whereas closed models can offer simpler initial adoption and embedded protections. However, these potential benefits depend on capable infrastructure that maintains GPU utilization, networking performance, autoscaling, and latency targets. CoreWeave presents its dedicated inference and Kubernetes offerings as options for deploying open models, while noting that model benchmarks, pricing, licensing, and available serverless catalog models may change.