Recursive Superintelligence signed a $410 million, multi-year compute deal with AWS this week — and it has nothing to do with building another AI assistant. The company, founded by former Salesforce chief scientist Richard Socher and backed by Google Ventures, Nvidia, and AMD, is building AI that runs its own scientific experiments. It proposes a research idea, implements it, runs the experiment, validates the result against a statistical significance threshold, then uses what it learned to design the next experiment. No human in the loop.
That last part is not marketing copy. They have benchmark results to back it up.
The Research Loop, Explained #
Recursive’s core system is an automated research loop: generate hypothesis, implement, test, validate, iterate. The system runs multiple threads in parallel, preserving findings across iterations and merging successful branches. Each result is checked against a p<0.01 significance threshold and screened for reward hacking — a critical safeguard, because an AI searching hard for improvements will eventually find ways to game the metric rather than genuinely improve performance.
This is meaningfully different from what GitHub Copilot or Claude Code do. Those tools generate code on request. Recursive’s system decides what to try next, designs the experiment, and runs it autonomously across extended timescales. AWS VP Jason Bennett described the infrastructure’s role directly: “turn months of sequential research into days.”
What the Benchmarks Actually Show #
The company published results on three benchmarks. Their technical write-up goes into full detail, but the headline numbers:
NanoGPT Speedrun: The AI community had spent two years optimizing the time to reach a fixed validation loss on a single H100 node, going from roughly 45 minutes in 2024 down to 79.7 seconds. Recursive’s system got it to 77.5 seconds — a modest delta, but comparable to recent human contributions on a benchmark that had been stable for months.
NanoChat: Training a small language model to minimum validation loss within a 5-minute GPU budget. Recursive improved bits-per-byte from 0.9372 to 0.9109, a 1.3x speedup equivalent. The system found hashed n-gram embeddings injected via learned attention gates — not a configuration humans had specifically tried before.
SOL-ExecBench: GPU kernel optimization across 235 real workloads, scored against theoretical hardware maximum. Recursive moved the mean SOL score from 0.699 to 0.754, closing 18% of the gap to hardware limits. It discovered methods like pre-scaled weight matrices that absorb normalization layers directly into GEMM operations — the kind of optimization that typically takes months of human kernel engineering.
These are real benchmarks with public leaderboards. The NanoGPT Speedrun has a well-documented community history. This is not a synthetic demo.
What the Deal Structure Tells You #
AWS holds no equity in Recursive. The $410M deal is pure compute-for-service. That matters because it reveals Recursive’s theory of leverage: the constraint on AI research velocity is compute throughput, not headcount. The deal accounts for most of their $650M raise, and Socher was direct about his expectations: “This may be one of the smallest deals the company signs in the coming years.”
The two companies are also co-developing infrastructure specifically designed for RSI workloads — parallel, long-running experimental loops that differ architecturally from standard training or inference jobs.
What to Watch For #
Socher has committed to “tangible, useful things you’ll be able to play around with” by October 2026. No specific product names yet. Given Nvidia and AMD Ventures among their backers, the most likely initial products are in AI inference and GPU optimization — territory where Recursive has already demonstrated measurable gains.
The broader pattern is worth noting. As TechCrunch reported, OpenAI is targeting intern-level AI research agents by September 2026. Claude Code already generates 80% of Anthropic’s production code. Recursive is the most direct current attempt to automate research design itself — the layer above code generation. The difference is the loop: not “write this code” but “figure out what to try next.”
The actual bottleneck in AI research has always been human bandwidth. Labs like OpenAI and Google run hundreds of major experiments per year. A system that runs thousands in parallel, validates rigorously, and compounds its improvements doesn’t need to be perfect to matter. It just needs to outpace human research velocity. IEEE Spectrum put it clearly: if that loop closes, every other AI milestone timeline accelerates with it.