AIs

Within a week, two of the three largest cloud computing companies in the world signed 20-year contracts for nuclear electricity. Amazon has taken 690 megawatts from Constellation’s Calvert Cliffs plant in Maryland, including 190 megawatts of additional generating capacity expected between 2030 and 2032. Google has signed an agreement that will bring 890 megawatts of additional nuclear capacity into the United States. Constellation is investing more than $4.3bn to upgrade 11 reactors across Illinois, Pennsylvania, and New Jersey. Neither agreement, however, involves building a new nuclear reactor. Rather, they are paying to generate more output from reactors that already exist – some of which are even older than the companies themselves.

A few years ago, such agreements would have seemed inconceivable. Nuclear power had fallen out of favour following accidents at Three Mile Island in 1979, Chornobyl in 1986 and Fukushima in 2011. Germany phased out nuclear power, while Japan shut down its reactors after Fukushima, although it has since begun restarting them. In many democratic countries, advocating for nuclear energy became politically difficult.

Today, an undamaged reactor at Three Mile Island is being brought back into service, with Microsoft contracted to buy its output for 20 years. The US Department of Energy is providing up to $4.4bn in loans to support the restart of three shuttered plants, with the aim of restoring about 2.25 gigawatts. Palisades in Michigan, which was closed in 2022, is contracted to be generating again by March 2027. If successful, it would become the first US commercial nuclear plant to restart after entering decommissioning.

What has changed?

The quiet reversal

In the last decade, the industry’s answer to AI’s energy problem was efficiency – things such as better chips, smaller models and smarter cooling seemed appealing and feasible. Industry conferences featured discussions about performance per watt, with modern artificial intelligence (AI) chips performing more computation per unit of electricity than their predecessors.

Yet those efficiency gains did not reduce overall electricity demand. Instead, each gain was quickly put towards something larger. Even when DeepSeek demonstrated that a competitive model could be trained at a fraction of the computing cost many had expected, the result was not a reduction in industry spending. Instead, laboratories continued planning more training runs, models that reason for longer, agents that work continuously in the background, and video generation.

This would not be the first time something like this has happened in history. In 1865, William Jevons observed that more efficient steam engines made Britain burn more coal, not less, because cheaper power created uses that had not existed before. The International Energy Agency estimates that global data centre electricity demand will more than double to around 945 terawatt-hours by 2030, even assuming continued improvements in efficiency.

Hence, the industry has changed strategy – unable to shrink demand, companies are now buying up supply.

The arithmetic of waiting

Money, however, cannot solve every problem quickly.

A new large reactor takes a decade or more from decision to electricity. Tech companies are investing in small modular reactors to power their data centres, although many of these projects are not expected to deliver electricity until the early 2030s. So far, very few of these have been built anywhere in the world. Thus, while reactor restarts and upgrades are the faster option, the “fast option” still means 2027, 2028, 2029, or beyond.

AI capital expenditure, however, does not operate on that same calendar. The chips being installed this year could be overtaken by newer technology long before the 20-year electricity contracts powering them expire. This discrepancy underpins all of these announcements, yet the industry has not addressed the implications if demand weakens while contractual obligations remain.

Who pays in the meantime

This is where the story stops being about technology.

Most of these deals sit inside PJM, the grid operator serving roughly 67 million people across 13 American states and Washington, DC. PJM uses capacity auctions to secure sufficient electricity-generating resources for future demand, with the associated costs ultimately passed on to consumers through their electricity bills. Capacity prices have reached the regulatory price cap in three consecutive auctions. According to the grid operator’s independent market monitor, data centre demand contributed approximately $29.4bn in capacity costs across the four most recent auctions. In the latest auction alone, around $6.3bn of these costs were attributed to data centre demand, including forecasts for facilities that have yet to be built.

In effect, households are already paying higher electricity capacity charges, partly because of projected demand from facilities that may never become operational. The Natural Resources Defense Council estimates that, by 2028, the typical household within PJM could face an additional $70 per month in electricity costs if current trends continue. These rising costs are also having political consequences. Recent US polling shows growing public dissatisfaction with data centres, with electricity bills among the main concerns. In response, Virginia has established a separate electricity rate class for data centres, Ohio has introduced minimum billing requirements, and Illinois is debating legislation.

The aim of the Google agreement is to tackle this objection. Its underlying principle is that major new electricity consumers should bring additional generation capacity rather than compete with existing users for an already constrained supply. This approach represents a constructive step towards mitigating the pressure that data centres place on electricity systems.

However, the arrangement remains voluntary. In October 2026, US federal regulators rejected most of PJM’s proposal to require large new electricity consumers to provide their own generation capacity. As a consequence, the industry is voluntarily offering a commitment that regulators have so far declined to impose through PJM’s proposed rules. This poses the question: how long will such voluntary commitments remain viable if the costs of fulfilling them begin to rise?

What the Gulf should take from this

For Gulf countries preparing their own AI infrastructure expansion, the US experience offers a warning about the costs of getting these arrangements wrong.

For countries with abundant energy resources, the expansion of AI infrastructure presents a significant opportunity.

Qatar’s sovereign fund has committed to a $20bn joint venture with Brookfield for AI infrastructure, beginning with a first phase of roughly 300 megawatts and potentially expanding to near 1.5 gigawatts. Abu Dhabi is pursuing plans for up to five gigawatts of AI data centre capacity. Saudi Arabia is targeting almost two gigawatts through HUMAIN. The region’s competitive advantage is substantial. As demand for AI computing grows, access to electricity is becoming an increasingly important factor in determining where computing infrastructure is built. The Gulf is well positioned to capitalise on this demand.

However, the scale of the proposed developments warrants scrutiny. If the proposed 1.5-gigawatt build-out is completed in Qatar, it would represent electricity demand equivalent to approximately 12 percent of the country’s installed generating capacity. Electricity allocated to AI infrastructure cannot simultaneously serve other purposes, meaning the opportunity costs of these projects deserve serious consideration.

In the United States, the debate over AI infrastructure is becoming increasingly centred on who ultimately bears the cost of its electricity demands, with many of the major commitments already in place. The Gulf is in a rare position of being able to settle that question first: what the industrial tariff is, who funds the new generation, what happens to the contracts when the hardware ages, and what long-term economic benefits the country would see beyond energy revenues and a limited number of jobs.

While hosting AI infrastructure can generate considerable economic value, the greater strategic opportunity lies in determining the conditions under which that infrastructure operates. For Gulf countries, the opportunity to establish these terms exists now, before large-scale commitments constrain their options.

The views expressed in this article are the author’s own and do not necessarily reflect Al Jazeera’s editorial stance.

For months before carrying out Canada’s deadliest school shooting in almost four decades, Jesse Van Rootselaar was consulting an unlikely adviser on how to perpetrate mass murder.

“In a hallway, indoors, or a crowded classroom, the 870 [Remington shotgun] is brutal … Close quarters is its playground,” OpenAI’s ChatGPT guided the 18-year-old, according to reports.

On February 10, 2026, Van Rootselaar shot his mother and half-brother at home before going to a school in the small British Columbia town of Tumbler Ridge, where he killed another six people – including five students – before taking his own life.

Earlier this week, on October 14, US authorities charged a Washington State man with allegedly conspiring with Van Rootselaar to plan the attack. But the mass shooting incident has raised questions that go beyond the traditional legal understanding of culpability.

The British Columbia government in September filed a civil lawsuit against OpenAI and its CEO Sam Altman, accusing them of failing to prevent the incident and claiming they were aware of the conversations that turned deadly. Families of the victims also filed separate lawsuits. OpenAI’s leadership issued a public apology for their failure to notify authorities.

The company has not responded to requests from Al Jazeera for its comments on the lawsuits.

But as the case against OpenAI progresses, it has spurred a familiar debate: Where should a line be drawn between reporting potentially dangerous behaviour and protecting an individual’s privacy?

“The question of whether companies should be required to turn such information over to law enforcement is extraordinarily difficult, particularly if they don’t already screen chats for indications of such types of threats,” United States-based civil liberties lawyer Jenin Younes told Al Jazeera.

“Forcing companies to implement a surveillance system implicates very serious civil liberties concerns, and that’s probably a bridge too far for me,” Younes added.

“On the other hand, AI gives individuals access to types and quantities of information they didn’t have previously that may facilitate the commission of mass casualty events, a concern it’s not easy to dismiss.”

AI’s ability to distinguish thoughts from threats

One of the biggest challenges facing AI companies is determining whether a user expressing violent thoughts actually intends to act on them.

In August, Anthropic, the company behind AI chatbot Claude, flagged a user to San Francisco police after discovering conversations allegedly threatening its CEO, Dario Amodei.

According to the San Francisco Standard, citing a police report, the user had claimed to possess an AR-15 gun and suggested he had Amodei “in his sights”.

However, when contacted by the newspaper, the man said he was joking and had no intention of carrying out any form of violence. The person was neither charged nor detained, the Standard said.

Fazl Barez, a senior researcher at the University of Oxford specialising in AI safety and governance, said current AI models cannot “reliably” determine whether an interaction with an AI chatbot “is purely hypothetical or is likely to lead to harmful behaviour”.

“Suppose you are a creative writer working on a fictional story and want to explore how a murder plot might unfold. From the context, it may be clear that this is simply a fictional narrative. If the threshold for flagging potentially harmful content is too low, such a writer could be wrongly flagged,” he told Al Jazeera.

“Similarly, consider a graduate student in biology or chemistry studying pathogens that, if misused, could have serious consequences. They might be trying to understand the risks associated with these pathogens and how to mitigate them. Such discussions may be necessary for legitimate research and education. Even when the subject matter is sensitive, they should not automatically be flagged without considering the broader context.”

The researcher said identifying a credible threat may require examining a user’s interactions across multiple conversations, rather than relying on isolated statements.

Barez said someone experiencing intrusive violent thoughts might turn to a chatbot for support, while another person planning an attack may reveal their intentions gradually across several conversations.

“Distinguishing between the two requires careful consideration of context, intent and patterns of behaviour, rather than relying on individual statements alone,” he added.

Chinmayi Sharma, a law professor at Fordham University, said the debate has parallels with the duty imposed on mental health professionals to warn identifiable third parties about credible threats.

However, she said there was a key difference – therapists have professional training, knowledge of their patients and enforceable standards governing their conduct. AI companies do not necessarily have comparable safeguards.

“My concern with AI companies is that right now there’s not a concerted effort to hire mental health professionals, sociologists, individuals that would have the unique expertise on what the indicators of threat are here,” Sharma told Al Jazeera.

“Secondly, licensed professionals are accountable to concrete, legally enforceable standards in their discipline. We don’t have something like that for AI companies.”

The privacy dilemma

The privacy implications extend beyond whether AI companies can accurately identify threats. Millions of people use chatbots to discuss relationships, emotional distress, mental health concerns and deeply personal experiences.

Civil liberties lawyer Younes said many users may not realise how vulnerable those conversations could be to disclosure.

“They often reflect a person’s innermost thoughts, questions, and desires. The majority of people assume that these dialogues will never be viewed by another person,” Younes, president and legal director of the American-Arab Anti-Discrimination Committee, said.

Younes warned that the consequences of greater monitoring could extend well beyond those planning violence.

“If people fear their chats can be turned over to law enforcement simply because they express disturbing thoughts/violent fantasies etc, many will be dissuaded from using them,” she told Al Jazeera.

Younes argued that forcing companies to establish systems for monitoring users and reporting conversations could raise serious constitutional concerns in the US, including protections against unreasonable searches.

Meanwhile, Fordham’s Sharma also raised an issue about who could be disproportionately affected by heavy monitoring by AI chatbots.

“The concern is always there about either intentional or unintentional targeting, because if you have models trained on what we consider dangerous speech, they’re almost always going to reflect biases,” she said. “Certain minorities’ speech is over‑flagged as a threat, while other comparable speech is not.

“Even in a world in which everyone is benevolent, we should still be concerned about disparate outcomes,” she said, warning that machines could reproduce such disparities at scale.

‘Not a licensed therapist’ – When do AI companies become liable to report?

While AI companies may face moral pressure to prevent violence, their legal obligations are less clear, legal experts say.

Bryan Choi, a professor of law at the University of Colorado Boulder, said US law recognises a duty to warn in certain circumstances, including when mental health professionals become aware of imminent threats or when a special relationship exists between parties.

But those principles do not necessarily apply to AI companies.

“Under present case law, it seems unlikely that a company such as OpenAI would be held to have such duties to warn,” Choi told Al Jazeera. “The chatbot is not a licensed therapist. There is no special relationship between the company and the victims.”

Choi said existing negligence and product liability laws could potentially be applied to AI companies, but proving responsibility would present significant challenges.

Finding a balance

Fordham’s Sharma suggested that independent institutions could help establish reporting standards rather than leaving decisions entirely to private companies or requiring immediate disclosure to police.

Such bodies she said could develop procedures for assessing threats and explore ways of sharing only relevant information while protecting users’ identities and other personal details.

Meanwhile, Younes argued that users should, at a minimum, be clearly informed about whether their conversations are monitored and what circumstances could lead to information being shared with law enforcement.

“This is almost certainly an issue that will be wrestled with over the coming years, and I’m not sure anyone knows the solution at this particular point in time,” she said.

“Events that have not yet occurred will likely influence how these policies are shaped.”