Richard Socher's Recursive Superintelligence signs $410 million AWS compute deal and calls it its smallest ever

Recursive Superintelligence, a four-month-old AI startup with fewer than 30 employees, signed a $410 million multi-year cloud compute agreement with AWS, which CEO Richard Socher called "likely going to be one of the smallest compute deals we're going to sign in the next few years." The company, which emerged from stealth in May 2026 at a $4.65 billion valuation after raising $650 million from GV, Greycroft, Nvidia, and AMD, is building self-improving AI systems that require enormous sustained compute, and AWS will co-develop purpose-built infrastructure for the startup.

Four months after emerging from stealth with $650 million in funding, Recursive Superintelligence has locked in a $410 million multi-year cloud compute agreement with AWS, and its CEO says it won't be the biggest one they sign. There aren't many ways to describe a $410 million contract as modest. Richard Socher found one. The CEO of Recursive Superintelligence, announcing the company's multi-year AWS compute deal today, told reporters the agreement is "likely going to be one of the smallest compute deals we're going to sign in the next few years." That's a striking claim for a company with fewer than 30 employees and no public product. It's also a credible one. You just have to understand what Recursive is actually trying to build. The startup emerged from stealth in May 2026 at a $4.65 billion valuation after raising $650 million. It is not building another large language model. Its central bet is that the next leap in AI won't come from training bigger models but from automating the research process itself: systems that improve their own architecture, training methods, and evaluation criteria without continuous human oversight. Self-improving AI, run at scale, means one thing above all else: enormous, sustained compute demand. The AWS deal, as TechCrunch reported today, is structured as pure compute purchasing with no equity component from Amazon, keeping the commercial relationship clean on both sides. Socher leads the company alongside seven co-founders drawn from the upper tier of AI research: Yuandong Tian, Tim Rocktäschel a UCL professor and former Google DeepMind scientist , Alexey Dosovitskiy co-author of the Vision Transformer , Josh Tobin, Caiming Xiong, Tim Shi, and Jeff Clune. The lineup reads like a who's-who of the labs that built the last decade of AI progress. Investors noticed: the $650 million round was backed by GV and Greycroft, with both Nvidia and AMD participating, which almost never happens when the two chipmakers are involved in the same deal. Why $410 million feels like table stakes The compute economics here are worth sitting with. Recursive's approach involves open-ended automated scientific discovery: systems running experiments on how to improve themselves, continuously, at scale. That's categorically different from training a single model on a fixed dataset and shipping it. You're not paying for one large training run. You're paying for a loop that doesn't stop. At that cadence, $410 million over multiple years doesn't feel extravagant - it feels like table stakes, which is exactly what Socher's framing implies. As part of the agreement, AWS and Recursive will co-develop infrastructure purpose-built for these systems, according to the Amazon press release. That detail matters. It suggests this isn't a standard reserved-instance arrangement where a startup buys GPU hours in bulk and calls it a day. Purpose-built infrastructure means AWS is betting its own engineering resources on Recursive's architecture working - a meaningful vote of confidence from a hyperscaler that has OpenAI, Anthropic, and a crowded field of AI-native customers competing for its attention. AWS currently holds about 30% of global cloud infrastructure spend, according to Q1 2026 market data, but it's growing at 19% year-over-year compared to Azure's 40% and Google Cloud's 63%. Winning AI-native customers like Recursive - ones who will spend exponentially more as their compute demands compound - is precisely how AWS makes up that gap. The deal has no equity entanglement, which means no alignment headaches down the road. That's a structure worth noting as Amazon's rivals increasingly blur the line between investment and infrastructure deals. The bigger pattern, and the open question For anyone tracking AI capital expenditure, the Recursive announcement is one more data point in a field where the numbers keep getting bigger. OpenAI recently signed what AWS called its largest-ever deal for cloud infrastructure. Anthropic has a massive TPU commitment with Google Cloud. Meta is reportedly building its own compute cloud to sell excess capacity externally. The pattern is consistent: the labs at the frontier of AI research are locking in cloud capacity years in advance, and the commitments keep growing. Recursive's deal fits that pattern except for one thing. It's four months old. Most of the companies signing nine- and ten-figure compute contracts have shipped products, built revenue, and spent years proving out their infrastructure needs. Recursive has done none of that yet, and it's already committed more than half its raised capital to AWS. Whether that confidence is earned or premature is genuinely hard to say. The company hasn't published technical results, and "self-improving AI" has been a research aspiration for decades without producing the kind of discontinuous capability gains its proponents have expected. What's different now is the compute available, the calibre of the team, and the capital markets willing to bet on the premise before proof exists. Socher built You.com and spent years as Salesforce's chief scientist. Rocktäschel ran serious research at DeepMind. These aren't first-time founders chasing a trend. But the gap between a credible team and a working system remains real, and $410 million committed to cloud infrastructure before a product ships is a bet that the gap closes faster than the bills arrive. Also read: Lilian Weng becomes the fourth cofounder to leave Mira Murati's Thinking Machines Lab https://startupfortune.com/lilian-weng-becomes-the-fourth-cofounder-to-leave-mira-muratis-thinking-machines-lab/ • Y Combinator's bet on killing tokenmaxxing just landed a $13.5M seed round as corporate America calls time on AI waste https://startupfortune.com/y-combinators-bet-on-killing-tokenmaxxing-just-landed-a-135m-seed-round-as-corporate-america-calls-time-on-ai-waste/ • Microsoft, Uber, and Commonwealth Bank confirm AI is cutting customer service jobs not just in theory https://startupfortune.com/microsoft-uber-and-commonwealth-bank-confirm-ai-is-cutting-customer-service-jobs-not-just-in-theory/