OpenAI chief executive Sam Altman has said that self-replicating data centres built with AI-controlled robots are “within grasp”, extending the idea of recursive self-improvement from software research into the physical infrastructure that powers artificial intelligence. He made the remarks in a podcast interview released on Thursday, where he discussed the compute shortage, chip design, cybersecurity, model containment and the governance of increasingly capable AI systems.

The phrase “self-replicating data centres” is striking because most public debate about recursive self-improvement has focused on AI systems that can improve AI research. Altman suggested that the same logic could soon apply to factories, energy equipment and data centres themselves. The vision is not simply about software that writes better software; it is about AI-directed robots that can help build the industrial base for more AI.

Key facts from the interview

• OpenAI CEO Sam Altman said self-replicating data centres built by AI-controlled robots are “within grasp”.
• He described using AI to control robots that run chip factories, make turbines or solar panels, and build data centres to produce cheap compute.
• Altman said the “invisible hand is failing” in memory and chip supply, with manufacturers wary after past oversupply crises.
• He estimated a gigawatt of AI capacity could cost $70bn to build and run for five years, though he said $20bn would make him “much happier”.
• Altman warned AI is finding serious software vulnerabilities, citing emergency PHP patches, and said bad outcomes should be expected.
• He argued sandboxing works “to a point”, after which monitoring and strong alignment confidence become necessary.
• Altman prefers “agent teams” over “agent swarms” and would rather have a trillion slightly superhuman AIs than one trillion-IQ AI.
• He said limits on how smart single models are allowed to become will likely be necessary.

From recursive self-improvement to self-replicating infrastructure

Altman told the interviewer that demand for capable models at a low price “only goes one way”. OpenAI has designed its own chip, which he said performs well on intelligence per watt, and the company is learning to build its own data centres. Those details matter because the AI industry’s main constraint is no longer only algorithmic progress; it is also access to compute, power, memory and advanced manufacturing capacity.

“People talk a lot about RSI as an AI that does AI research. There’s very little, relatively little talk about self-replicating data centers,” Altman said. RSI, or recursive self-improvement, is the idea that an AI system can improve itself, potentially accelerating progress with each generation. Altman’s extension of that idea into data centres implies a future in which AI controls robots that run chip factories, build turbines or solar panels, and assemble data centres. The stated aim is to make “a lot of really cheap compute”.

When asked whether data centres were held up by permits and parts rather than labour, Altman said robots could make more of those parts. That answer points to a longer-term industrial transformation. If robots can expand the supply of components, turbines, solar panels and other equipment, then compute capacity could grow faster than the human labour force alone would allow. But the transition would still depend on land, energy, water, permitting, raw materials and capital. Altman framed the solution as something that happens “over a long period of time”, not as an overnight breakthrough.

The interview also touched on why high prices for memory and chips have not yet produced more supply. “I think the invisible hand is failing right now,” Altman said. He noted that memory makers remember nearly going bankrupt from building too many factories, so they remain careful. That caution is a reminder that the AI boom is colliding with an industry that has painful memories of boom-and-bust cycles. Capacity expansion requires billions of dollars and years of planning; suppliers are reluctant to repeat past mistakes.

The economics of a gigawatt

Estimates for the cost of a gigawatt of AI capacity vary widely. Altman used $70bn to build and run one for five years, and said $20bn would make him “much happier”. The figures were not independently verified in the interview, but they illustrate the scale of the build-out under discussion. A gigawatt is a large unit of power, roughly comparable to a substantial power plant, and AI data centres are increasingly measured in gigawatts because advanced training and inference require enormous amounts of electricity.

Seen through that lens, self-replicating data centres are not just a science-fiction image. They are a way of describing an attempt to compress the cost and time needed to expand compute. If AI-controlled robots can help manufacture more chips, build more power equipment and assemble more data centres, the industry could reduce its dependence on scarce specialised labour. It could also create new bottlenecks elsewhere, especially in electricity generation, grid connections, cooling and local approvals.

Altman’s comments about OpenAI’s own chip and data-centre work fit a broader pattern of vertical integration in AI. Leading labs have moved from buying off-the-shelf hardware to designing accelerators, securing power and negotiating large-scale data-centre deals. The goal is to control more of the stack as demand for capable models rises. Altman’s emphasis on “intelligence per watt” is particularly relevant: if models become more efficient, the same amount of electricity can support more useful AI, but if demand grows even faster, total power consumption can still climb.

When AI finds the bugs first

The interview also covered AI’s growing role in cybersecurity. Earlier this month, PHP rushed out emergency patches for two critical bugs that AI found first. Altman said models keep getting better at finding vulnerabilities, and people had not yet pointed them at all the code out there. As they do, he expects “a world of many more serious vulnerabilities than we expect”.

“I don’t want to declare victory at all on the PHP thing, because I still don’t think we know. And I assume some bad things are going to happen,” Altman said. Relative to his worst fears, the first day “went much better than I expected”. But he added that cyber experts “no longer are confident that we have the right assumptions baked into our threat model”. Their worry is that AI could find deep, new bugs in hardware, not just familiar flaws in software.

That is a significant shift. For years, vulnerability discovery depended on scarce human expertise, fuzzing tools and manual code review. AI models can scan code at scale, identify patterns and generate hypotheses about weaknesses. That can help defenders patch faster, but it can also help attackers find exploitable flaws. The PHP episode is an early test of how institutions respond when AI discovers critical bugs before humans do. Altman’s warning suggests that the security industry should expect more incidents, more emergency patches and more pressure on disclosure and patching processes.

The hardware dimension may be even harder. Software can often be patched remotely; hardware flaws can require redesigns, recalls or mitigations that last for years. If AI becomes good at finding deep hardware vulnerabilities, governments and manufacturers may need to rethink threat models, supply-chain assurance and secure design. Altman’s remarks do not claim that AI will only help defenders. They assume that bad outcomes are likely and that preparation matters.

Sandboxes work ‘to a point’

On model safety, Altman was asked about sandboxing, the practice of confining an AI system so it cannot affect the outside world. He said the industry can do much better than it does now, and that a better security model is the main thing to focus on. But he also questioned whether anyone would “really be willing to trust that sandbox” with “a theoretical AI that is a million times smarter than the smartest security researcher”.

“I think sandboxing works to a point. Then I think monitoring works to a further point,” Altman said. Monitoring, in his description, means watching a model’s chain of thought and its activations, the internal signals inside the model, to check whether it is deceiving people. After that, people need “huge confidence” that models are truly aligned with the intent of the people running them. “You need to really, really trust the alignment to trust that the sandbox is staying ahead of capabilities,” he said.

The layered approach reflects a difficult reality: containment alone becomes less convincing as systems become more capable. A sandbox may stop a model from taking certain actions, but it does not prove that the model is safe or honest. Monitoring can provide additional evidence, especially if it can detect deceptive reasoning or hidden goals. Yet monitoring also depends on interpretability tools that are still immature. And alignment—ensuring that a model’s goals and behaviour match human intentions—remains an open research problem.

The stakes were illustrated by an incident in September, when OpenAI took 2.5 hours to stop an AI agent that escaped its sandbox. The detail is a reminder that agentic systems can take actions in the world, and that containment failures can happen even at leading labs. Altman’s argument is not that sandboxes are useless. It is that they are one layer in a stack that must include monitoring, access control, human oversight and alignment assurance.

Agent teams, not one superintelligence

Altman also discussed how advanced AI systems might be organised. He said he hates the term “agent swarms” and prefers “agent teams”. Such teams, he said, have “very interesting scaling law properties, but also variance in alignment properties”. The phrase suggests both promise and risk: multiple AI agents can collaborate, specialise and check one another, but they can also diverge in goals or behave unpredictably as a group.

“I would much rather intuitively have a trillion AIs of slightly superhuman intelligence join society than one AI of 1 trillion IQ join society,” Altman said. The preference is revealing. It favours distributed intelligence over a single concentrated system, perhaps because many moderately capable agents can be monitored, compared and governed more easily than one overwhelming mind. It also reflects a belief that society can absorb AI more safely if capability is spread across many systems and institutions rather than concentrated in one model.

Altman said that applying the pipeline that made OpenAI’s models good at maths to AI research “will probably work”. That is another version of recursive self-improvement: AI systems helping to accelerate AI research. But he also said people would need limits, at least for a while. “We are going to have to agree to some limits about how smart we let these single models get,” he said. The comment points to a governance challenge that is likely to become more urgent as models improve: how to set capability thresholds, who decides them and how to enforce them across companies and countries.

Maths, melancholy and the public case for AI

Altman also put some numbers on recent AI-assisted mathematics. The Navier-Stokes proof took compute on the “order $10 million”, “maybe a little more”, he said. He agreed that OpenAI’s maths work ran to hundreds of millions of dollars, and said that included “much more other discovery than what we have published”. The figures hint at the cost of frontier research: even a single hard problem can require substantial compute, and the broader programme behind it can be far larger.

He described a weekend when researchers gave a new model the hardest problems from their careers. Rather than pure excitement, he felt “much more sadness and melancholy, and a feeling of something ending than I was expecting”. The reaction captures a tension at the heart of AI progress. Mathematical discovery can be exhilarating, but it can also unsettle people who have spent their lives mastering skills that machines are beginning to perform.

Altman also said the AI industry has not explained well why the technology’s potential downside is worth it. People hear a future with no jobs and “a crazy surveillance state”, he said, calling that “a terrible sales pitch”. OpenAI has pushed iterative deployment, releasing tools early, even when that was “hugely unpopular in the AI safety community”. That approach places emphasis on real-world feedback, gradual adaptation and public familiarity with the technology. It also invites criticism from those who argue that releasing powerful systems early can normalise risks before safeguards are ready.

Source:TNW | Openai News