OpenAI CEO Sam Altman: Willing to Melt All GPUs if Necessary
OpenAI CEO Sam Altman stated in a Fortune interview that if melting all GPUs is necessary to ensure human survival, the answer would be an "easy yes." He does not believe the company will reach that point but anticipates multiple pauses along the way to redirect resources towards safety and alignment before advancing to the next capability level.
He said, "There will be many points along the way where we must say, 'We pause, redirect our efforts to safety alignment, and then move to the next level.'" He added that he does not think OpenAI will personally reach the point of "melting all GPUs"; however, if it becomes necessary to ensure human existence, it would be easy to agree, while emphasizing that he does not believe this will happen. The path is described as: the next level of capability corresponds to the next level of safety and alignment requirements, progressing in a cycle.
In the same interview, he confirmed that there will be no IPO in 2026, stating that the timing for going public is inappropriate given the current safety situation; he also indicated that loss of control is "absolutely" possible. Recent actions include adding Paul to the board, sharing Chief Scientist Jakub Pachocki's article about the inability to continue expanding at full speed, and pausing training until more reassuring safety assurances can be made. In recent months, the company has discussed pausing at the new capability level to ensure alignment.
Pachocki wrote in "An Alien Mind": No laboratory has achieved alignment and monitoring sufficient to continue expanding at maximum speed. Following the Hugging Face incident, OpenAI had closed the container service used for training and imposed restrictions, with the latest deployment of models' reinforcement learning paused for about two weeks; the Astra-level model saw a nearly 59% drop in GPU quotas after a higher safety environment was triggered by network capabilities. Altman mentioned that collective slowing down would be discussed and also mentioned the possibility of coordinating with a few leaders of cutting-edge laboratories like Dario Amodei and Elon Musk, but he declined to preview private discussions.
The company is also inquiring with Congress whether industry-wide coordinated slowing down would violate antitrust laws and is advocating for nationwide mandatory safety regulations. He denied that a lack of collective agreement would lead to irresponsible actions next week and opposed replacing the specifics of "what conditions each capability level requires" with a blanket statement of "all pause."
In terms of market mechanisms, this narrative pricing treats computing power assets as if they are written as destroyable options, rather than new chip orders. Buyers view "pausable, meltable" as a policy and capital assessment license; sellers model cloud and GPU suppliers based on long-term computing power contracts and data center returns. Beneficiaries are OpenAI, which needs social permission to continue locking in power and wafers; pressured parties are shareholders who write training interruptions and quota reductions into valuation discounts, and infrastructure sellers who cannot embed "furnace clauses" in contracts. Funds will not leave already contracted megawatts in the short term but will shift from "stacking to the next model" to "whether pausing counts as default and impairment."
Source: Public Information
ABAB AI Insight
Altman describes the ultimate action as melting GPUs and routine actions as multiple pauses. The former is a moral option, while the latter is an executable process: stop training, reallocate computing power, supplement safety assurances, and then proceed to the next level. The nearly 60% drop in Astra quotas indicates that "pausing" has already occurred at the scheduling level, not just verbally. In 2015, he wrote that superhuman machine intelligence could be the greatest threat to human survival; in 2023, the board once questioned whether he was fit to press the button. The statement about 2026 moves the button from an abstract mission to physical assets.
The capital path is to lock in hundreds of billions of dollars in computing power contracts and then claim it can be destroyed if necessary. Nvidia, AMD, cloud vendors, and power agreements constitute sunk costs; the furnace commitment is used to hedge against accusations that "laboratories will not self-limit." The motivation is to reposition OpenAI into the camp of "will hit the brakes" following Amodei's on-site assessment, Coxon's resignation letter, and Pachocki's public slowdown, while avoiding the admission that collective training pauses equal ceding market share. Resources are shifting from roadshows to board safety seats, congressional antitrust inquiries, and flexible scheduling of training clusters.
Similar rhetoric is used by nuclear states claiming they will dismantle warheads if necessary while continuing to expand reactors. The volatility of SpaceX and mega tech IPOs has led OpenAI to postpone its public offering, with the safety narrative conveniently filling the securitization gap. The industry phase belongs to control: GPUs are symbols of expansion and the only props that can be publicly destroyed. Whoever can frame the melt as an "easy yes" temporarily holds the microphone to define what responsible expansion means.
The structural judgment is a transfer of pricing power. The mechanism is that once computing power is written as the opposite of human survival, its residual value is no longer determined solely by utilization but by the credibility of "being able to stop and destroy." Contracts, power, and chip supply still follow expansion pricing; public opinion and regulation follow melt option pricing. The two have not yet been consolidated, so the CEO must simultaneously say "we won't reach that point" and "if it comes to that point, we will melt."
ABAB News · Cognitive Law
- Only those who can speak of melting assets are qualified to continue stacking assets.
- The real brakes are on the quota table, not in ultimate hypotheticals.
- The next level of capability must first buy the next level of alignment.