NVIDIA and CoreWave enter the Fera Robin generation with an agent execution leap that redesigns development environments

Core Wave (CoreWave) announced at its annual conference in San Francisco that the NVIDIA Fera Robin NVL72 systems and Spectrum-X networks are available at speeds up to 102.4 terabits per second, thereby moving accelerated computing infrastructure into the production-grade operation stage for AI agents. Cognition Labs, the developer of independent software engineer Devin, became the first customer to shift production workloads to the Fera Robin platform, alongside the unveiling of the NVIDIA Fera central processor, the first CPU designed specifically for independent agents.

The initial benchmark tests conducted by Cognition Labs to compare the performance of Fera Robin NVL72 with the previous generation GB 200 NVL72, using a sample of real programming tasks sourced from the Frontier Code platform, reportedAchieving a jump in code token throughput of 4.8 times in software model inference workloads.According to the lab’s founding team, code generation in agent environments requires long contexts and intensive synchronization, where the cost of a single token determines the ability to launch a software product, making multi-step processing acceleration a decisive factor in actual response time.

Agent systems place a dual strain on infrastructure: inference execution demands a sharp reduction in response time, while post-training and reinforcement-learning stages require running thousands of isolated environments concurrently. To solve this, the central Fera processor provides 128 processing units and 11,264 cores in a single cabinet, enabling more than 11,000 simultaneous work environments at roughly one core per environment, supported by Spectrum-X switches and Bluefield-4 data processing units. Experiments showed more than a threefold acceleration in the boot time of isolated test environments, as well as a 1.7-times improvement in terminal-bench tests across successfully completed tasks.

Simultaneously with the hardware stack, Core Wave launched the Forge platform to integrate training and evaluation tools with OpenAI's expertise and the open-source programming notebook project Marimo. The platform includes the Aria tool to suggest experiments, modify code, and save it on GitHub, and the Egent Lens service, which raises fault detection rates by 20 percent and halves error-handling costs by tracking millions of agent execution traces. The open-source Dynamo inference technology also allows test points to be updated during live operation without redeployment, a path that companies such as Canva, Capital One, and the Inoval Care health-care foundation have begun using to run clinical document automation and support medical decision-making.

This hardware shift directly impacts development teams, banks, and government agencies in the Gulf and Egypt, which are currently piloting internal process automation and software engineering. Multiplying code-generation throughput by roughly five times and halving the cost of managing isolated runtime environments alters the technical decision-making of technology leaders in the region: instead of restricting agents to simple retrieval tasks for fear of inflating inference bills and long contexts, moving toward full software automation becomes an economically feasible option. The availability of monitoring solutions and reduced fault-repair costs also compels local personnel to shift from merely writing command-line scripts to roles in operations engineering and continuous evaluation of agent safety in live environments.