Start a trial

Operational notes Partnerships

Recursive and AWS: $410 million in cash still doesn’t buy independence from a supplier

6 min read

A black-and-white server module pulled halfway out of a rack, with other blurred modules lined up in the background
A module can be pulled out and swapped. Infrastructure built specifically for one customer cannot.

On 28 July 2026, Recursive Superintelligence — a start-up founded by Richard Socher, formerly Salesforce’s chief scientist — announced a multi-year agreement with Amazon Web Services worth $410 million in compute capacity. The detail that got specialist media talking wasn’t the figure: it was the structure. The deal is entirely cash, with no equity stake going to AWS. In an industry where large compute deals are often paid in equity or through circular arrangements — a supplier invests in the customer, who spends the money back on the supplier — that’s the unusual part of the story. But paying cash doesn’t remove dependence on a single supplier: it just moves it, and it’s worth understanding where.

What was announced

Recursive Superintelligence came out of stealth on 13 May 2026, with a $650 million round and a $4.65 billion valuation, led by GV (Alphabet’s fund), Greycroft, Nvidia and AMD’s venture arm. It has offices in San Francisco and London; the company’s own site describes a team of more than 25 people with prior backgrounds at OpenAI, DeepMind, Google Brain, Meta, Salesforce and Uber, alongside Socher. It has no public product yet: Socher told TechCrunch, “In October or so, you’ll see some actually tangible, useful things” — October 2026, a forecast, not an accomplished fact.

The AWS agreement, reported by TechCrunch and confirmed by Data Center Dynamics and other specialist outlets, is a multi-year compute commitment worth $410 million, with no equity exchanged. Jason Bennett, AWS’s vice president for start-ups and venture capital, said part of the agreement involves plans to “co-develop infrastructure purpose-built for these types of companies.” The exact type of accelerator involved — AWS’s own chips or Nvidia GPUs — has not been disclosed. Socher, on the size of the figure, called it “likely going to be one of the smallest compute deals we’re going to sign in the next few years”: a signal about future scale, not a binding commitment.

What a system that writes its own research actually does

In a technical post published in June, Recursive describes a system that automates a research loop: it proposes an idea, implements it, runs an experiment, validates the result, and uses what it learns to choose the next experiment — running many threads in parallel, retaining useful context between experiments, and applying an explicit check against “reward hacking” (a system that improves its score without genuinely improving its behaviour). These are figures the company reports itself, on benchmarks it chose, unverified by any third party: on fixed-budget language-model training (NanoChat Autoresearch) it claims a loss improvement from 0.9372 to 0.9109 bits per byte, a 1.3x speed-up; on a training-speed test (NanoGPT Speedrun) a cut from 79.7 to 77.5 seconds; on GPU kernel optimisation (SOL-ExecBench) a mean score of 0.754 against 0.699, an 18% reduction in the gap to optimal performance. The limitation is written by the company itself: the system works on problems with clear metrics and fast feedback loops, not on open-ended scientific research — that remains future work.

The debate over the name

“Recursive self-improvement”, RSI, is the term the company chose even for its own social handle. Not everyone agrees it describes what exists today. Helen Toner, of Georgetown’s Center for Security and Emerging Technology, draws a line between labs that are “just using AI for as much as they can” and the classical definition of RSI, where “there are no humans needed” — a gap that remains wide today. Even Google’s chief executive, Sundar Pichai, said publicly that on the pace of acceleration the term implies, “we aren’t quite there yet.” A panel of experts convened by Georgetown was split, some expecting a capability explosion and others a slower path followed by a plateau. We take the same cautious position: Recursive’s benchmarks are a real data point on a narrow set of tasks, not proof of autonomous scientific research.

What it means

One day earlier, on 27 July, Nvidia and Safe Superintelligence — Ilya Sutskever’s lab — announced a partnership of the opposite kind: an investment (reportedly, according to Reuters and other outlets, though not confirmed in the official announcement, around $5 billion) in exchange for access to the Vera Rubin platform, with a compute increase described as “an order of magnitude.” Two days, two opposite structures — equity for compute versus cash for compute — addressing the same problem: where a frontier AI lab finds the power to scale. For Recursive, paying cash avoids dilution and makes the relationship with AWS easier to unwind on paper. But the clause about “purpose-built co-developed infrastructure” describes a different lock-in, no less real: a facility built to measure for one customer is harder to walk away from than a generic GPU, whatever currency you paid it in. For anyone buying compute — not just billion-dollar labs — the lesson is the same: how you pay and how portable the infrastructure is are two separate questions, and the second one matters for longer than the first. We’ve written more generally about ownership of the stack, from the machine to the model: Hardware and open-weight models.

The operational lesson

  1. Separate the announcement from the contract: a “multi-year compute commitment” doesn’t yet say how many nodes, which hardware, which SLA — the technical details of this deal aren’t public, and until they are they remain a promise.
  2. Ask what “purpose-built” actually means: infrastructure customised for a single customer is often more efficient, but almost always less portable. A limit that isn’t written down and verifiable doesn’t exist — the same applies, in reverse, to exit clauses.
  3. Don’t confuse announced capacity with available capacity: the gap has already shown up elsewhere, and a contract signed today isn’t compute running today.
  4. Be wary of the acronym before the product: benchmarks on tasks chosen by the supplier, without independent verification, remain marketing figures until proven otherwise — that applies to “RSI” as much as to whichever acronym appears in the next press release you read.

Anyone assessing a compute or AI supplier should apply the same test to every partnership announcement: what is already proven, what is still a promise, and what happens if the supplier changes course. That’s why a dedicated AI system should be delivered in two ways, never one alone: on-premise in the customer’s own environment, or CSIDIA’s dedicated cloud — an environment reserved for a single customer, accessed via a dedicated VPN, with a data centre in Italy staffed directly by us. Questions about portability get written down before signing, not after. What counts is the method, not whichever supplier’s brand is fashionable this quarter.

Assessing a multi-year compute deal, or a supplier promising results “within a few months”? Let’s talk for thirty minutes.

Sources