Recursive Trades $410M for Custom AWS AI Research Clusters

On July 28, 2026, Recursive Superintelligence formalized a $410 million multi-year compute agreement with Amazon Web Services. The arrangement allocates 63% of the startup’s total capital directly into cloud infrastructure. Recursive intends to use this hardware to run an automated AI research system designed to autonomously write, implement, and test architectural changes to its own models.

Custom Infrastructure for Autonomous Research

This deployment goes beyond a standard cloud lease. The deal involves co-developing purpose-built hardware topologies optimized for high-parallelism autonomous research workloads. This setup allows Recursive Self-Improvement (RSI) systems to generate hypotheses, alter code, and validate benchmark performance continuously.

If you are building multi-agent systems, this scale of deployment illustrates the infrastructure required to shift from static inference to continuous execution. The cluster is engineered to support the rapid spinning up and tearing down of training environments necessary for automated experimentation.

The Scientific Discovery Loop

Recursive’s core platform automates the entire machine learning research cycle. The system operates in four distinct phases: proposing algorithmic modifications, implementing the new code, executing validation experiments, and selecting subsequent trials based on the data.

The system generates novel configurations by analyzing previous failure states and compiling successful proposals into executable models. This requires extensive context engineering within the agent’s memory to track which architectural paths have already proven suboptimal. By automating these steps, the architecture shifts the development bottleneck from human engineering time to raw compute availability.

Agent Count Over Headcount

Recursive operates with fewer than 30 employees distributed across San Francisco and London. This lean structure stems from a strict strategy to prioritize compute cycles over human researchers. The company currently holds a $4.65 billion valuation following its recent $650 million funding round.

Unlike recent agreements from Microsoft and OpenAI or Amazon and Anthropic, the AWS partnership involves no equity exchange. It remains a traditional customer-vendor contract, keeping Recursive fully independent while securing top-tier infrastructure access.

Timeline and Product Strategy

Led by Richard Socher and Tim Rocktäschel, Recursive plans to release its initial products by October 2026. The founders claim the internal system has already outperformed years of human optimization on specific benchmarks. Evaluating these upcoming models will require robust AI evaluation frameworks to verify the output of non-human scientific discovery.

Socher characterized the AWS agreement as a baseline, anticipating even larger compute acquisitions in the future. The company emerged from stealth earlier in May 2026 with backing from GV, Greycroft, Nvidia, and AMD Ventures.

If your organization plans to scale autonomous testing loops, assess your cloud infrastructure for high-parallelism bottlenecks early. Standard inference environments often lack the necessary throughput for continuous, programmatic code validation and retraining.