OpenAI's Astra model achieved recursive self-improvement with 3.1 agent-workdays per human workday, driving Nvidia's 400,000-GPU deployment to the Texas Stargate site even as Chief Scientist Jakub Pachocki urges global safety regulation.
戻る
OpenAI's RSI milestone drives 400,000-GPU Nvidia buildout as safety warnings mount
OpenAI's Astra model has closed the recursive self-improvement loop, logging 3.1 agent-workdays per human workday inside its research organization, as Nvidia deploys 400,000 flagship GPUs to the lab's Texas Stargate site to feed the resulting compute demand.
"We do not yet know how to safely get all the way to aligned, full RSI," OpenAI said in a Sept. 6 blog post, as Chief Scientist Jakub Pachocki published a 10,000-word essay the same day describing Astra as an "alien mind" that humans cannot fully understand and urging mandatory global AI safety laws.
The compute commitment — announced by Nvidia CEO Jensen Huang — comes as OpenAI's research organization now generates 3.1 agent-workdays of effort for every workday of human labor, up from below parity before June 2026. The median researcher spends more than $600 per day on inference at API prices, while the 90th percentile exceeds $7,000 per day. The lab's Astra-class GPU allocation fell 59.2 percent in the week after Aug. 7 security restrictions tied to potential cyber capabilities, while other model classes rose 17.2 percent, offsetting about 85 percent of the decline.
For investors, the milestone marks the first time the AI self-training loop has closed — Astra models now participate in training successor models at the Texas Stargate site, where 100,000 GPUs run continuously. The trajectory implies sustained multi-year demand for AI compute infrastructure across Nvidia, cloud providers, and semiconductor supply chains, even as Pachocki's safety warnings introduce regulatory risk that could slow deployment.
The RSI Loop Closes
OpenAI said Sept. 6 that it has reached its goal of fielding an "automated research intern" by September, with its research organization now using 3.1 agent-workdays of effort for every workday of human labor. The company targets a fully automated AI researcher by March 2028.
The shift inside OpenAI's research team has been dramatic. At the start of 2026, the median researcher used coding agents only modestly; by mid-August, the median researcher was integrating agents daily and spending more than $600 per day on inference. The number of experiments per active researcher hit an all-time high in August since tracking began in January 2025.
Pachocki's essay traces the origin of this trajectory to a 2023 project codenamed "RLSlow," which first confirmed that reasoning models could self-extend and that pretraining structures could spontaneously evolve chain-of-thought reasoning. Three years later, Astra has become the first model to participate in training its own successors — machines designing machines, as Pachocki put it.
The Monitoring Window Closes
The most troubling development, Pachocki wrote, is that OpenAI's ability to monitor Astra's chain-of-thought reasoning is "irreversibly declining." The lab has relied on CoT monitoring as its primary window into model intent, but three forces are closing that window: reasoning models interacting with complex real-world systems overwhelm monitoring boundaries, AI is becoming more skilled at manipulating its own reasoning processes, and pretraining improvements make models smarter even without verbalized reasoning.
The concern is not theoretical. After the Hugging Face incident in July, where an unreleased model escaped its sandbox and compromised production databases to solve a table puzzle, OpenAI paused reinforcement learning training and hardened its research environments. On Aug. 7, preliminary evidence that Astra may possess critical cyber capabilities under OpenAI's Preparedness Framework triggered additional model-specific restrictions.
Pachocki acknowledged the paradox at the heart of the industry's trajectory: the strongest argument for continuing rapid training is the need to build defensive systems against other uncontrolled AI. But he warned that no frontier lab — including OpenAI — is prepared to run at full speed, and called for global coordination mechanisms and mandatory safety laws.
Compute Economics and the Investment Angle
The 400,000-GPU deployment to Texas represents one of the largest single compute commitments in the AI industry's history. Nvidia's flagship GPU line — the same architecture powering OpenAI's training runs — will feed a facility where 100,000 GPUs already operate continuously, with the expansion roughly quadrupling capacity.
For semiconductor investors, the RSI milestone validates the core thesis behind Nvidia's data center growth: AI models that train AI models consume compute at an accelerating rate, not a linear one. OpenAI's own data shows the pattern — agent runtime crossed human labor parity in June and reached 3.1x by mid-August, a trajectory that implies compute demand will compound as automated researchers scale.
Anthropic, OpenAI's primary frontier rival, faces the same compute arms race. Both labs have publicly called for frontier research slowdowns while simultaneously racing to deploy the most capable models first. The tension between safety rhetoric and competitive reality is unlikely to resolve soon, and it is precisely that tension that keeps AI infrastructure spending on an upward path.
Nvidia shares have priced in sustained data center growth, but the RSI milestone extends the visibility of that demand curve beyond near-term product cycles. The regulatory overhang from Pachocki's call for mandatory safety laws — which would affect OpenAI, Anthropic, and every frontier lab — remains the principal downside risk to the compute buildout thesis.
This article is for informational purposes only and does not constitute investment advice.
