CoreWeave Details How It Runs AI Factories After Launch, Up to 96% Goodput

Why goodput is the number that matters

The post argues that once a system moves from validated to operational, the question changes from whether it works to whether it keeps working. A cluster that passes every benchmark can still lose real hours to interruptions, or to a checkpoint that silently did not save, according to CoreWeave.

CoreWeave defines goodput as the share of time a system spends doing useful work rather than recovering from interruptions. On long-running, distributed training workloads, the company says it has demonstrated up to 96% goodput on NVIDIA Hopper GPUs. Everything below that line is lost productivity, the company says.

The post breaks goodput into two variables, throughput and resiliency, and uses a 1,024-NVIDIA Hopper cluster benchmark to isolate both.

• Throughput: CoreWeave measured 20% higher model FLOPs utilization (MFU) on Hopper-based systems than other publicly reported benchmarks show. The company notes that MFU depends on model architecture, sequence length and parallelism, so the figure reflects its test context and is not a universal result.
• Resiliency: Mean time to failure improved roughly 10x, with an effective training time ratio of up to 98 percent.

CoreWeave attributes the gains to reliability, availability and serviceability (RAS) features built into the GPU, CPU and rack architecture, combined with its own software layer that detects, contains and recovers from issues. Failures still happen, the company says, but they cost less time.

The software layer: Mission Control and SUNK

CoreWeave Mission Control is the layer that decides what to do when hardware shows a problem. According to the post, it handles straggler detection, automated node draining, fleet and rack lifecycle management, and recovery workflows. It processes over 200 million metrics samples per second across all customer environments.

SUNK, CoreWeave's Slurm on Kubernetes offering, dynamically reallocates GPUs across training and inference workloads on the same cluster as demand shifts, so freed capacity does not sit idle between jobs. CoreWeave links these capabilities to its MLPerf Training v6.0 result on NVIDIA GB300 NVL72, where Mission Control held a consistent performance baseline and SUNK's NVLink-domain-aware placement kept communication local across the NVL72 domain.

The post includes a quote from Craig Falls, Head of Quantitative Research at Jane Street: "Our research depends on infrastructure that's both powerful and reliable, and CoreWeave has delivered on this as we've scaled across NVIDIA Hopper and Blackwell." He adds that the company's cluster observability and support team give Jane Street confidence to partner with CoreWeave on NVIDIA Vera Rubin.

Resiliency across NVIDIA generations

On the hardware side, the post says NVIDIA Blackwell introduced a dedicated RAS engine that applies AI-based predictive maintenance across thousands of hardware and software data points. NVIDIA Vera Rubin adds a second-generation RAS engine for proactive maintenance and real-time health checks without downtime. Vera CPUs add SOCAMM LPDDR5X serviceability and in-system core testing.

At the system level, the post says modular cable-free trays make Vera Rubin assembly up to 90x faster than Blackwell, while software-defined NVLink routing reroutes around faults to keep operation continuous with less maintenance overhead.

Preparing for Vera Rubin NVL72

CoreWeave says that in June 2026 it completed the industry-first bring-up and validation of NVIDIA Vera Rubin NVL72. The idea, according to the company, is that customers inherit a validated system instead of debugging a new architecture in production.

The post notes that NVIDIA normally redesigns one or two chips per generation, while the Vera Rubin platform redesigns every chip. Six purpose-built chips are unified into a single supercomputer: Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet Switch. Together they are meant to address communication, memory and compute bottlenecks and reduce the cost of training, post-training, reasoning and agentic inference.

Vera Rubin runs at 100% liquid cooling with no air-cooled option at this power density, so CoreWeave built three systems to manage it rack by rack. Valvey and Racky are described as patent-pending and available only on CoreWeave. Valvey is a per-rack liquid-cooling valve assembly that watches flow rate, temperature and pressure continuously and isolates a rack automatically if something goes wrong, without taking down neighboring racks. The Rack LifeCycle Controller orchestrates workflows across the system, and Racky is the per-rack control point that ties the pieces together. The source text is cut off before describing Racky in full.

What it means for customers

The post's central message is that the lifecycle stays the same with each hardware generation: design together, validate before deployment, and operate with full visibility once live. For customers running large training jobs, the pitch is that more of the compute they pay for becomes finished work. The figures, however, are CoreWeave's own, and the company stresses that 96% goodput is an upper bound, not a guarantee.