On July 29, 2026 (Beijing time), artificial intelligence company Recursive Superintelligence announced a multi-year cooperation agreement with Amazon Web Services (AWS). The total value of the announcement is $410 million. Recursive will use AWS as its cloud service provider for its automated AI research system, and both parties plan to jointly develop infrastructure for large-scale automated research. TechCrunch reported on July 28 at 13:19 UTC that the agreement amount was also $410 million, which is within the day of the company's announcement.
Scope of Agreement and Automated Research Loop
The current scope of the agreement confirms multi-year cooperation, a total value of $410 million, AWS cloud services, and joint development of infrastructure. The announcement does not provide specific years, minimum consumption amounts, payment schedules, termination conditions, chip series, processor or accelerator quantities, reserved capacity, cloud regions, delivery dates, or energy requirements. The $410 million is the announced value of the agreement and should not be described as an amount Recursive has already paid to AWS; there are also no disclosures on division of labor, ownership, or access terms for jointly developed facilities.
In its official technical article, Recursive describes the system as a set of automated research loops: it proposes a solution for a target, implements it, runs an experiment, validates the result, and uses the information to select subsequent experiments. The company says the system can maintain multiple long-running research threads in parallel, retain useful context from earlier experiments, combine promising branches, and check variance and reward hacking before treating a performance change as progress. This is Recursive's first-party description of its own system; the public materials provide no third-party operational audit.
Two Small Model Training Benchmarks
In the NanoChat Autoresearch test, Recursive reported a validation result of 0.9109 bits per byte, with a reference value of 0.9372; lower numbers indicate better performance in this task. The company states that the system first searches for solutions on NVIDIA H100 GPUs and then switches to B200 hardware, evaluating ten random seeds against public references. Recursive acknowledges that the underlying model may already know the public techniques, so results cannot prove independent rediscovery by the system; there is no independent audit of this data set.
NanoGPT Speedrun requires a small language model to achieve a validation loss of 3.28 on the FineWeb dataset and minimize training time. Recursive claims that its system reduced the required time from 79.7 seconds to 77.5 seconds, while meeting leaderboard statistical requirements. The official article attributes changes to combined adjustments in attention calculation precision, optimizer behavior, embedding updates, training scheduling, and fusion of GPU kernels rather than a single algorithm change. This result is self-reported by the company and cannot be used to infer production model costs or quality.
Kernel Testing and Team Background
In SOL-ExecBench testing, Recursive says its system jointly optimized 235 GPU kernels for NVIDIA B200 hardware, raising the average score from the previous leaderboard result of 0.699 to 0.754. The benchmark sets 1.0 as an analytically estimated performance optimum, and the company describes the change as an 18% reduction in the gap to that estimate. Recursive also notes that reward hacking was particularly challenging in this test; while the team checked some high-scoring kernels, it may have missed errors in areas outside its expertise. There is no independent recheck of these results in the public materials.
Recursive's official website states that Richard Socher serves as CEO and co-founder, that the team has more than twenty-five people, and that it has offices in San Francisco and London. The team backgrounds listed on the site include OpenAI, Google DeepMind, Google Brain, Meta, Salesforce, and Uber. The company defines its current focus as automated AI research and released its first benchmark materials and some run artifacts on June 11. The public pages do not disclose revenue, customer count, profitability, or cash levels; the team's background cannot be used to infer business performance.
Amount Differences and Subsequent Verification
There is one textual difference that needs to be retained between sources: TechCrunch's canonical link path includes '400-compute-deal', but TechCrunch's published headline and the company announcement both clearly state $410 million, so the article uses $410 million while retaining the canonical link as-is. The three benchmarks cover fixed-budget small-model training, training speed, and GPU kernel optimization; that scope is not equivalent to general scientific discovery and does not prove that Recursive has achieved a generally self-improving system. The company's public run artifacts are available for inspection, but that does not mean the results have been independently certified.
This agreement translates Recursive's compute needs for parallel automated research into a multi-year cloud-services commitment with a disclosed value and extends the cooperation to joint infrastructure development. This is editorial analysis based on the published transaction structure, not validation of the system's capabilities or business model. The next points to verify are the agreement term, actual spending, chips and capacity, deployment regions, delivery arrangements, facility ownership, and access terms, as well as whether independent teams can reproduce the public benchmarks. Until that information is available, the size of the transaction, the company's self-reported performance, and general self-improvement capabilities should be treated separately.