August 2026 Summaries

A storm-related failure at the us-central1 data center on August 19, 2026 disabled both the building management system and chilled-water cooling, causing temperatures to rise rapidly and triggering thermal shutdowns across servers, network equipment, and rack power systems. The resulting regional outage disrupted compute, VM networking, storage, Managed Kubernetes, Token Factory, managed platform services, and regional console and API endpoints, while other regions remained largely unaffected. Cooling was restored and the facility stabilized within about six hours, but recovery continued into August 20 because the regional control plane required manual restoration, VM recovery mechanisms were rate-limited and did not automatically retry failures, stale disk attachments blocked some resources, and a disk hot-plug bug left Kubernetes control-plane data disks unattached. Object storage recovered by 13:27 UTC, external VM connectivity by 12:50 UTC, Kubernetes clusters by 21:58 UTC, Token Factory by 20:25 UTC, and reserved GPU capacity by 08:38 UTC the following day, although some on-demand capacity remained reduced during hardware repair. The provider identified gaps in independent facility monitoring, environmental alerting, full-region recovery procedures, automation for mass failures, regional observability, and traffic health checks, and plans resilience reviews, independent alerts, emergency procedures, regular recovery drills, improved storage reconciliation, and cross-region monitoring.

Nebius has announced that its Token Factory inference platform will be the first AI cloud to adopt NVIDIA Groq 3 LPX alongside NVIDIA Vera Rubin NVL72, targeting faster token generation for latency-sensitive agentic workloads. The company argues that as agents make many sequential model calls, generation speed increasingly determines overall task completion time, particularly when processing long contexts and maintaining multi-turn state. Each Groq 3 LPX rack combines 256 LPU accelerators, 128 GB of on-chip SRAM, and 640 TB/s of scale-up bandwidth in a liquid-cooled NVIDIA MGX architecture. Artificial Analysis reportedly measured Groq 3 LPX running Gemma 4 31B at 3,400 output tokens per second for one user, while NVIDIA projects Vera Rubin NVL72 could deliver up to 35 times greater inference throughput per megawatt than GB200 NVL72 for long-context, low-latency 2T-parameter workloads. Nebius plans to provide the new hardware through existing Token Factory capabilities, including serverless and dedicated endpoints, autoscaling, observability, function calling, structured outputs, and safety guardrails, allowing current customers to use it through model selection rather than new infrastructure, SDKs, billing arrangements, or vendor relationships.

Nebius announced the results of its Serverless AI Builders Challenge, in which 35 qualifying public projects across AI and machine learning, healthcare and scientific AI, and robotics were evaluated for implementation, reproducibility, educational value, platform use, usefulness, and originality. First place went to Shmulik Avraham’s MediSimplifier, a LoRA-fine-tuned model that translates complex hospital discharge summaries into plain language and demonstrated rapid, low-cost parallel experimentation while revealing weaknesses in single-model LLM judging. Andrei Goldenberg placed second with an evaluation dashboard that compares hosted and self-hosted models on users’ own data for quality, latency, and cost, finding that self-hosting is economical mainly at high utilization and that reasoning outputs require preprocessing. Third-place winner Zaher Khateeb’s Balagan tested multi-agent systems under crashes and sabotage, showing that fault type can matter more than fault count and that majority voting improved resilience at a higher token cost. Seven special awards recognized projects in areas including production fine-tuning, legal-document recognition, robotics, agents, clinical AI robustness, and CPU-based planning, while a Community Choice award remains open for voting during the August 25 Serverless AI office hours.

Nebius Academy, the education and research division of AI infrastructure company Nebius, aims to help individuals and organizations turn AI access into practical business value through professional certifications, research grants, and enterprise training. Its role-based certification program combines targeted training, exams, digital badges, and Nebius cloud credits for positions including AI CloudOps Engineer, Agentic AI Builder, AI Cloud Leader, AI Infrastructure Engineer, and AI Solutions Architect. The Research Grants program offers eligible academic and nonprofit researchers free GPU cloud credits, model access, agentic search tools, and technical support for computational projects, with 2026–2027 applications open through an academic-year cycle. For businesses, Nebius Academy uses a four-stage model of assessment, role-specific training, internal AI Champion development, and ongoing skills measurement through its Evolve learning platform, emphasizing that sustained workforce capability rather than AI tools alone drives transformation.

Nebius argues that although modern AI relies heavily on open-source software, data, research, and public benchmarks, the most capable models have often remained proprietary, creating risks of concentration among a small number of companies. It contends that increasingly capable open-weight models can broaden competition, but their practical success depends on production infrastructure such as efficient inference, low latency, reliability, and affordability. Nebius says it supports customer choice between open and closed models while contributing to the open ecosystem through inference optimizations, research such as LK losses for speculative decoding, open-source tools including Kvax and Soperator, hardware designs, benchmarks, agent-development resources, and research funding. The company maintains that openness does not inherently ensure quality or safety, but enables more people to inspect, test, reproduce, improve, and address failures in AI systems. It frames open and proprietary AI as complementary rather than mutually exclusive, arguing that a durable open ecosystem is necessary to preserve choice, scientific scrutiny, and the broad circulation of useful ideas across the AI stack.

Training speculative-decoding draft heads for long-context language models faces two major bottlenecks: the memory required to materialize vocabulary logits for language-model losses and the inefficient sparse attention patterns used by EAGLE-3 draft heads. Streaming Cross Entropy addresses logits memory by calculating token-level cross-entropy loss in chunks rather than constructing logits for every batch token simultaneously, preserving the original objective when targets, masks, and head-specific shifts are aligned before chunking. This can reduce the loss-path memory requirement from hundreds of gigabytes to an amount determined by chunk size, although excessively small chunks reduce throughput through added iteration and matrix-multiplication overhead. Block-sparse FlashAttention addresses EAGLE-3’s banded diagonal attention by computing only block-level regions allowed by causal, document-packing, and draft-head constraints, avoiding dense quadratic attention processing and enabling efficient FlashAttention-style kernels. Benchmarks cited for GPT-OSS-20B-like configurations show substantially lower memory use and higher throughput than a JAX baseline, while combining both techniques reportedly reduced GPU memory utilization from 44% to 25% and raised throughput from 6,900 to 11,200 tokens per second. Together, the approaches support training workload-specific speculative draft models with larger vocabularies, longer contexts, and multiple prediction heads in Nebius Token Factory.