自改进AI（RSI）或驱动创新但将给数据中心带来压力

Opposing silver and black robotic arms gripping a golden AI microchip between precision claws

RSI posits models that design their successors, raising new operational questions.Getty Images

Recursive self-improvement (RSI) describes AI systems that can design, code, and deploy their successors with minimal human input. If achieved, improvements could compound rapidly as smarter models beget even smarter generations – potentially accelerating beyond human comprehension and, worryingly, control. For data center operators, RSI would reverberate across power, cooling, orchestration, governance, and certification.

What RSI Is and Where It’s Heading

Although RSI is loosely defined, much like artificial general intelligence (AGI), there are signs it’s edging closer to reality. Achieving RSI depends on fast advances in AI-assisted coding, which labs say are already underway. For example, frontier AI model creator Anthropic, in its report titled “When AI Builds Itself,” says its Claude model now authors a substantial proportion of its own codebase: “As of May 2026, more than 80% of the code we merge into Anthropic’s codebase was authored by Claude. Before Claude Code launched in research preview in February 2025, this number was in the low single digits.” If sustained, this kind of contribution is a step toward systems that can improve themselves.

On the hardware side, breakthroughs such as quantum computing could catalyze step changes in RSI or AGI. “In the recursive AI world, it's basically an optimization problem. What is the best AI? What is the best AI model neural network going forward? Quantum can support this,” said Jay Guilmart, lead product manager at Q-CTRL. “There's an algorithm called [Grover’s algorithm], which is really good at searching large parameter spaces. It can support the recursive operation to help [the system] perform better and pick the best option faster.”

If AI can repeatedly build better versions of itself, the implications for physical infrastructure are immediate. What does that mean for IT hardware lifecycles, energy use, workload orchestration, and the way we design and operate data centers?

“The basic premise of ‘self-improving systems’ is not new in and of itself,” said John O’Brien, senior analyst at Uptime Institute. “What’s different now is that … advances in AI models and infrastructure make it likely we are at the point where some of this blue-sky theorizing is becoming technically feasible – or soon will be.”

A practical lens on RSI separates systems that truly author and deploy new code (RSI) from those that self-optimize continuously without rewriting themselves. The latter is already emerging in data centers, often underpinned by reinforcement learning (RL). “Successful AI applications in the data center today often make use of reinforcement learning (RL), a well-proven machine learning technique that uses penalties and rewards based on feedback from the environment,” O’Brien said. “This can work well in closed-loop systems where parameters are defined, such as data center cooling, IT, and power.”

Systems that can self-improve in situ are already emerging. “Startups Emerald AI and Phaidra are making breakthroughs in early pilots and demonstrators that show the potential of dynamic, self-improving systems,” O’Brien added.

Autonomous improvements, especially in energy use, are a goal of a new wave of data center management software platforms, agreed 451 Research senior analyst Zoe Roth. “Players like Phaidra, etalytics, and Vigilent are already proving that closed-loop AI can continuously optimize physical facility variables on the fly,” Roth said. “Current AI engines like Phaidra act as autonomous control layers over existing cooling (Building Management System) and power (Electric Power Management Systems) systems. They don't rewrite their own code, but they do retrain and update neural networks on real-time sensor telemetry.”

Operational Impacts for Data Centers

For operators, self-optimization promises tangible benefits in monitoring and maintenance. “Condition-based maintenance today needs a substantial historical dataset before it can identify fault risk with any confidence,” said Alex Cordovil, research director at Dell’Oro. “A system that improves on its own operating experience could get there much faster and much more effectively, with real gains in uptime and performance.”

But that raises a thorny question: how do you certify software that keeps changing? “If a software suite, or an AI agent, were judged capable of performing data center management duties, I'm not sure how you certify something that keeps changing underneath you,” Cordovil said. “You can't certify a moving target.” In practice, this could require ongoing re-approval of models or continuous monitoring and rollback controls that keep humans in the loop, reducing some of the efficiency gains.

“You can't certify a moving target.” – Alex Cordovil, research director, Dell’Oro

Self-improving AI that shifts between training and inference could also exacerbate the wide power swings already challenging AI data centers. “Sustained synchronous training behaves nothing like inference: large fleets swing between near-idle and near-peak in lockstep, tens to hundreds of megawatts at fractions of a hertz,” Cordovil said. “A fleet that trains and serves at the same time would make that a permanent design requirement rather than a training-cluster problem.”

Strategic Outlook and Industry Effects

Beyond near-term design choices, RSI could alter industrial dynamics. A system that can code and design its successor would have huge ramifications for those designing, building, and manufacturing data center infrastructure,” O’Brien said. “Would they want to automate half of their workforce? What would that do to their business?”

For now, RSI remains a nascent concept. But if it’s ever achieved, the effects – good and bad – will extend well beyond the models themselves, reshaping the entire AI stack and rippling through society. One plausible scenario features AI not only coding more powerful successors but also designing and even prefabricating the data centers that will house them. Even then, self-optimizing, self-replicating AI data centers would face hurdles in regulation, certification, safety assurance, and public acceptance. Expansion already feels too fast for many communities; RSI could accelerate the pace further.

About the Author

Reporter

Andrew Donoghue specializes in journalism, analysis, and thought leadership content for sectors including data centers, critical infrastructure, and sustainable technology.

He has held senior content, strategy, and market intelligence roles at organizations including Vertiv and Fujitsu. He is the author of several influential reports covering topics including renewable energy and the data center, IT power management, and data center cooling.

Andrew has worked for analyst companies such as 451 Research and has also held senior editorial roles at business publishing companies, including CBS Interactive and Incisive Media. He has also been involved in several European Commission-funded IT research projects on data center energy efficiency and sustainability.