Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI

by Karen Hao

Generating one thousand images used on average as much energy as 242 full smartphone charges; in other words, every AI-generated image could consume enough energy to charge a smartphone by roughly 25 percent. Luccioni’s and Strubell’s papers are among the few that still provide quantifiable measures of the carbon behind generative AI models.

The land and energy required to support these megacampuses are but two inputs in the global supply chain of data center expansion. So, too, is the extraordinary volume of minerals including copper and lithium needed to build the hardware—computers, cables, power lines, batteries, backup generators—and the extraordinary volume of potable—yes, potable—water often needed to cool the servers. (The water must be clean enough to avoid clogging pipes and bacterial growth; potable water meets that standard.) According to an estimate from researchers at the University of California, Riverside, surging AI demand could consume 1.1 trillion to 1.7 trillion gallons of fresh water globally a year by 2027, or half the water annually consumed in the UK.

As more and more communities have watched data centers affect their lives, a growing number have pushed back vehemently against their unfettered development. In response, data center developers have grown more sophisticated with tactics to maintain business as usual: They’ve entered communities in secrecy under shell companies; they’ve donated to community programs to dampen resistance; they’ve made promises to cities about the sustainability of their facilities before walking them back one by one after projects have broken ground and are more difficult to reverse.

No one within Microsoft or OpenAI even knew whether Phase 5 was technically possible. In Microsoft and OpenAI’s design plans, The Information later reported, the $100 billion facility could need as much as 5,000 megawatts, nearly matching the average power demands of all of New York City. While the plan didn’t seem the most financially sound as a business investment, money wasn’t the main bottleneck. It was energy. “We’re running out of land and power,” an OpenAI employee says. Within both companies, it was understood that Phase 5 would only become possible with some amount of innovation.

As OpenAI trained GPT-4 in Iowa, the state was two years into a drought. The Associated Press later reported that during a single month of the model’s training, Microsoft’s data centers had consumed around 11.5 million gallons, or 6 percent, of the district’s water. GPT-4 had trained there for three months. (A Microsoft spokesperson said the company is working to increase its water efficiency by 40 percent above its 2022 baseline and to replenish more water than it consumes across its global operations by 2030, with a focus on the water-stressed regions where it works.) Arizona, too, faces a severe water crisis. In 2022, as Microsoft laid the groundwork for Phase 3, a study in Nature Climate Change found that the Southwestern US had been facing the worst drought it had seen in over a thousand years. That drought, combined with severe mismanagement, has drained the Colorado River, which Arizona and six other states rely on for fresh water, to dangerously low levels. Without drastic action, the river could cease to flow. The shortage compounds a power crisis, as climate change has slammed the region with relentless record-breaking temperatures and families have cranked up their air-conditioning. The region relies in part on hydropower from the Hoover Dam and water-cooled nuclear power plants. In other words, it needs water to produce more energy. In 2023, the Phoenix metro area hit multiple new heat records as well as the worst year for heat-related fatalities, which surged ...

The mining industry is a system, and that system, left to its own devices, will seek profit at any cost. “The worker doesn’t exist,” she says. None of the victims received any ceremony or commemoration; none of their families received compensation. “In that place, there is no humanity.” Chile is the world’s largest producer of copper, accounting for a quarter of the global supply.

Today Chile produces roughly a third of the world’s lithium, second only to Australia.

As companies and the Chilean government have been forced to invite them to negotiations, central to Indigenous demands are the need for the government to conduct research into the health of the Atacama Desert’s ecosystems and to quantify the water loss and any irreparable damage.

In the middle of the plot, a purple sign announces in Spanish, “Welcome to the Quilicura Urban Forest,” a project, it explains, that Google began in 2019 to give back to the community for hosting its data center. The sign includes a diagram to elaborate on the “forest’s” benefits: on the left side is an illustration of Quilicura as an industrial zone, packed with factories producing greenhouse gases and air pollution; on the right is an illustration of the forest flourishing under the generous rain pouring out from a big cloud labeled “SMOG.” Google boasts about this forest on its website and in its PR releases. When I ask the company’s country spokesperson for an interview about Google’s development in Chile, she sends me instead some polished briefing materials, later adding that the company creates a community impact program for each new data center to support local projects such as in education, sustainability, internet access, and health; for Quilicura, Google has invested over $1.2 million. In her materials, the part about the forest talks about residents using the green space. There are no residents. The place is too far from any bus line, and there are no homes in the surrounding area to speak of. Outside the modest plot, too small to fit Google’s data center itself, a dozen stray dogs meander around, barking and rummaging through the trash. The spokesperson said the forest is being updated to “evolve the experience for the community.”

food and providing jobs to the local community. Google’s data center—like most data centers—did not provide many jobs beyond its initial construction. A job posting from 2024 for a mechanical technician, one of the few long-term positions available, was advertised on Google’s job board only in English; the posting stated that Google wouldn’t consider résumés submitted in any other language. The data center—as activists point out—did not provide much other benefit to the local community either. Nearby, public schools still lack good internet or devices for students to access

In 2020, the government went a step further. It announced a project to build a new underwater cable, akin to a data highway, for connecting the Asia Pacific straight to the Americas through Chile’s central coast, not far from Santiago. Chile would become a global hub for digital infrastructure. Google backed the partnership. But in July 2019, as Google began the paperwork for its second data center in Chile, a group of residents was watching. The company had chosen Cerrillos for its new location, another working-class municipality bordering Santiago. Like Quilicura, Cerrillos has a long history of being overlooked and abandoned. From the 1930s to the 1990s, a cement factory that belonged to a Belgian company contaminated the community with lethal levels of asbestos, leading to what one Chilean historian called “the largest industrial genocide” in the country.

MOSACAT, a water activist group, began combing through all 347 pages of the filing. Buried in its depths, Google said that its data center planned to use an estimated 169 liters of fresh drinking water per second to cool its servers. In other words, the data center could use more than one thousand times the amount of water consumed by the entire population of Cerrillos, roughly eighty-eight thousand residents, over the course of a year.

MOSACAT was founded in 2019 after activists from several different movements fighting for women’s, housing, workers’, and environmental rights joined in solidarity to form a unified collective. Many had met while protesting an illegal mining project. MOSACAT’s activism successfully chased out the miners, shut down the project, and designated the land a protected nature reserve, the group says. It was shortly thereafter that a friend of the group, who now serves as a member of Chile’s national congress, tipped them off about Google’s data center project and urged them to look at its projected water consumption.

In most cases, projects in Chile that require water take a long time to receive approval. In Google’s case, the approval came quickly, even in the midst of a series of drought-related water emergencies. At first, MOSACAT sought to contest the project through Google’s local partner, a Chilean investment and services firm named Dataluna. The initial meeting went badly, MOSACAT says: The Dataluna representatives seemed to have little understanding of the project and denied that it would use fresh water. From there, MOSACAT went to the local government. The mayor and city council had themselves been meeting with Dataluna, the group remembers, and held the false impression that the data center needed only wastewater for its cooling. After MOSACAT briefed them, the government, alarmed, demanded an explanation from Dataluna. The matter escalated from Dataluna to Google’s Chile division all the way to Google’s headquarters in Mountain View, California.

They come offering us trees while drying out our earth.” Residents of Cerrillos didn’t need trees. They didn’t need Google to build them a park—as if the Santiago metropolitan area were such a backward place that it didn’t already have parks. They needed Google to stop treating their land as a place to plunder precious water and other resources; they needed the company to stop dismissing the community as bystanders instead of participants in the development of its local projects. “We know that we feed the world, that we provide raw materials like copper and lithium,” Salinas says. “Nobody is saying our treasure is ours alone and we won’t share it. Yes, we can help each other. But they are not going to come and use the water, which is vital for life, and leave us with nothing.”

the risk of pushing back as a Global South country is always that a Silicon Valley company will pick up and take its money somewhere else. As the project continued to stall and Google’s desire for more compute intensified, the company announced that it would shift its next planned data center in Latin America from Chile to another country.

science park, Parque de las Ciencias, which operates as a zona franca, a free-trade zone. Some cheekily call it zona America, with a hint of bitterness, for housing mostly American companies that do not pay taxes to the government.

Google purchased twenty-nine hectares of land here, fifty-eight times the size of Antel’s total data center footprint, to establish a different home for its second data center in Latin America. But at the time, Uruguay did not in fact have an abundance of water. Like Chile, like Iowa, like Arizona, it was also experiencing a devastating drought. The water shortage was so severe that farmers were losing their entire harvests, costing the country over $1 billion in agricultural losses; by the summer of 2023, the Montevideo government would start mixing contaminated salt water into the city’s drinking supply. Families opening their taps saw a putrid brownish fluid pouring out that smelled intensely of chemicals. Those who could afford it purchased bottled water for drinking and bathed with their windows open to avoid breathing in too many carcinogens. Those who couldn’t drank the tap anyway, leading many to suffer stomach pains, skin rashes, an aggravation of existing health conditions, and the agony of a growing rate of miscarriages.

In Uruguay, more than 80 percent of the country’s fresh water goes to industry instead of human consumption—most notably, cash crop agriculture. These include industrial farms for soybeans and rice, and for trees that feed into paper production.

Pena had read about hyperscalers using potable water, even during major droughts, and the activism of communities like MOSACAT that had resisted the projects. But when he downloaded the details of the project, the water numbers were marked as confidential. After submitting a public information request, which he had successfully done around twenty times, the ministry continued to withhold the numbers, saying they were proprietary information.

In March 2023, four months later, Pena won the case in a surprising victory. The environmental ministry revealed that Google’s data center planned to use two million gallons of water a day directly from the drinking water supply, equivalent to the daily water consumption of fifty-five thousand people. With much of Montevideo receiving salt water in their taps not long after, the revelations were explosive. Thousands of Uruguayans took to the streets to protest Google and all of the other industries that had led the government to squander the country’s precious freshwater resources. The slogans of resistance are still scrawled across the city’s walls and roadside barriers during my visit in June 2024. “This is not drought,” reads one. “It’s pillage.” Pena sees data centers in the same way that Ramos sees mining. It’s not the infrastructure itself that’s the problem, but the scale at which Silicon Valley is trying to build it.

Near the end of 2023, Google silently updated its proposed data center in Uruguay to use a waterless cooling system and said it would reduce the facility to a third of its size. Pena says the fight is still not over: The government is now withholding the projected energy consumption of the latest proposal as a commercial secret. Pena also sent a petition to the ministry, with over four hundred signatories, demanding a more extensive environmental and social impact study of the full supply chain of the data center: where the minerals for producing its hardware are being extracted, how the labor involved is being treated, how much carbon will be emitted, how the generated e-waste will be disposed of in a way that doesn’t leach chemicals into someone’s community.

In 2024, Chile’s environmental court ruled that Google cannot build a water-using data center in Santiago.

Microsoft projected that it would need a significantly lower amount of water than Google had in Cerrillos. Even still, Vallejos worried. The drought in Chile had only gotten worse and was expected to last until 2040. Quilicura’s wetlands were suffering acutely, on top of industrial encroachment, from accelerating desertification. On its website, Microsoft boasted about new cutting-edge innovations in data center cooling systems that would mean its facilities didn’t need to use water. If Microsoft had the capability to build waterless data centers, why wasn’t it doing so in Quilicura? “It is deeply striking that a company with as much reputation as Microsoft publicly presents an environmentally friendly discourse, but in reality does not comply with global innovation standards in a third world country like Chile,” Vallejos later wrote in an article. Microsoft would subsequently explain that the innovations it had advertised were still under development and only being piloted in a place in the US. “Then why do they promise these things” on their website? Vallejos asks.

Otero knew she had one thing Vallejos did not: affiliations with prestigious universities like Harvard and Columbia that would command Microsoft’s and the government’s attention. She began to mount a multipronged campaign, growing so deeply involved that she quit her job: She spoke at high-profile conferences about the environmental impacts of data centers and Resistencia’s fight against them; she forged connections with the Chilean Ministry of Science, Technology, Knowledge and Innovation and with representatives at Microsoft and Google; she connected Vallejos and Arancibia to other international researchers to elevate their profile. With Dambrosio and Díaz, Otero also developed a more speculative project. All three had architectural backgrounds and had been studying the infrastructure of modern digital technologies through the lens of the built environment. They began to wonder: What if they treated data centers as architecture structures and fundamentally reimagined their aesthetic, their role in local communities, and their relationship with the surrounding environment?

the three researchers applied for funding and put together a fourteen-day workshop, inviting architecture students from all around Santiago to reimagine what a data center could look like for Quilicura. The students designed stunning mock-ups. One group imagined making the data center’s water use more visible by storing it in large pools that residents could also enjoy as a public space. Another group proposed tossing out the brutalist designs of the typical data center in favor of a “fluid data territory” where data infrastructure coexists with wetland, mitigating its damaging impacts. The structures of the data center would double as suspended walkways, inviting Quilicura residents to walk through the wetland and admire the ecosystem. Plant nurseries and animal nesting stations would be interspersed among more traditional server rooms to rehabilitate the wetland’s biodiversity. The data center would draw polluted water from the wetland and purify it for use before returning it. The computers themselves would collect and process data about the health of the wetlands to accelerate the local environment’s restoration.

As part of the data center plan, the Ministry of Science, which oversees its creation, has for the first time formed a committee of activists to consult with regularly as part of the drafting process, and invited Vallejos, Arancibia, and members of MOSACAT to join them.

After centuries of extractivism, the country intimately understands what it means for its land to be hollowed out, dispossessed, and destroyed under a banner of progress. It could use those experiences as a wellspring from which to generate fundamentally new conceptions—decolonial conceptions—of AI. “In the planetary market of AI, we as a country are playing a specific role: giving materials to develop this technology,” he says. “Many companies are trying to extract a lot of material from us to create AI. “So we need to think from this position, this geopolitical position, this terrestrial position. We can think of another way to relate technological innovation with the earth.”

Altman was attending an exclusive dinner with sixty House members at the Capitol, feasting on an expertly prepared buffet with roast chicken. At the same time, the artists were hosting an interactive cocktail hour and trying to attract as many staffers with the best their budget could buy: wine and Chick-fil-A. It was a small but darkly comedic illustration of who commanded power and influence in the AI policy conversation and who didn’t.

industry. Amid the climate of frustration and fear in Washington, a policy white paper echoing Altman’s recommendations arrived two months after his hearing in July 2023. Written by a consortium of researchers, including from OpenAI’s Safety clan, Microsoft, and Google DeepMind as well as more than a dozen think tanks, many tied to the Doomer community, it pushed once again for a new licensing regime for AI models using compute thresholds, and the development of AI safety evaluations for dangerous capabilities including the ability to manipulate and persuade and the creation of novel biological weapon recipes.

If an AI developer produces a large language model that is able to create recipes for bioweapons, “it’s because they trained it on a dataset that included information on bioweapons,” says Sarah Myers West, the co–executive director of AI Now Institute and former senior adviser on AI to the FTC. As always, the neural network is surfacing patterns within its training data. Opening up that data would be the first step to establishing scientific clarity on what kinds of inputs could lead to dangerous outputs.

A Schumer spokesperson would later note in the press that the senator was personally consulting with Altman and other OpenAI executives as he moved closer to regulation. “That for me was a huge realization,” Raji says. “Wow, we need more people just debunking—just looking at what people are saying and being like, ‘Actually, reality is more complicated.’ ”

“Harry Potter and the Methods of Rationality.” The serial novel, which spans 122 chapters and over 660,000 words, reimagines Harry engaging in the wizarding world as a well-trained rationalist. It had served for many as a gateway into effective altruism and, in turn, to broader Doomer ideology. Yudkowksy had also cofounded the blog LessWrong, a central hub for AI safety researchers to foster community and propagate AI safety ideas, where he’d advocated with increasing alarmism to put a full pause on AI development as his p(doom) shot up to 95 percent.

For the independent directors, every instance added up to a single troubling picture: Bit by bit, Altman was trying to cloud their visibility and maneuver in ways that prevented the board from ever being able to check him. For years, Altman had advertised the board’s ability to counterbalance and even fire him as OpenAI’s most important governance mechanism. Indeed, he was now trumpeting this fact around the world to secure public and government trust.

A former Facebook data scientist and two tech and sex work experts say it’s possible that Annie’s very first SeekingArrangement account in December 2019 could have limited her online traction, based on the nature of how tech platforms track and shadow ban sex workers through automated systems by tagging their devices, emails, bank accounts, or other information that ties together their online presence, even for profiles completely unrelated to their sex work.

Over the years, she had been skeptical of its various configurations. Having three cofounders on what was meant to be an independent nonprofit board was far too many. And many of the independent directors had not had true independence from Altman; in one way or another, they’d had financial ties with him or had benefited from his networks. She had observed his investing in startups and donating to politicians to establish and entrench important relationships.

Q* had not factored into the board’s decision. But, as Pachocki alluded to in his talk, Research was indeed treating the new algorithm with intense importance. The algorithm had been a brainchild of Sutskever’s, rooted in research he had been developing since 2021 to advance OpenAI’s models without a need for more data. The idea was to get a deep learning model to make better use of its existing data by using more compute at inference time to deliver better results. It broke the logic of the original scaling laws, which tied together a model’s performance with three inputs used during training: data, parameters, and compute. With this new method, Sutskever hoped to improve a model’s performance by further scaling the one ingredient he had always believed to be the most important: compute—but inference compute rather than training compute; in other words, the compute used to generate the model’s responses.

The more OpenAI faced uphill challenges, the more Altman seemed to overcompensate with public declarations of its extraordinary success. The pattern was becoming so consistent it was turning into a signal: If Altman was being brazen and boastful, most likely something wasn’t going well.

In the same month, The New York Times filed its copyright infringement lawsuit, which added to a snowballing pile of other lawsuits from artists, writers, and coders over OpenAI’s reaping hundreds of millions, then billions, of dollars from models trained without credit, consent, or compensation on their work and that were now being used to automate away their jobs.

On top of the competition from Anthropic and Google, Microsoft had also begun to more aggressively diversify its AI portfolio in a response to the ouster.

Later in 2024, Microsoft would officially list OpenAI as a competitor in its SEC filing and not mention the startup even once during the fiscal year’s final quarterly earnings call.

“He told me that he felt that by my voicing the system, I could bridge the gap between tech companies and creatives and help consumers to feel comfortable with the seismic shift concerning humans and A.I.,” she wrote in her statement. “He said he felt that my voice would be comforting to people.”

The Blip and the Omnicrisis were one and the same: the convulsions that arise from the deep systemic instability that occurs when an empire concentrates so much power, through so much dispossession, leaving the majority grappling with a loss of agency and material wealth and a tiny few to vie fiercely for control.

revolution (what would become the country’s national motto)—“Liberté, egalité, fraternité”—and how it could be reinterpreted and wielded to consolidate his own power. It was ultimately under that banner that Napoleon did the opposite: He restricted freedom, dismissed fraternity—a philosophy based in unity and solidarity—and granted equality only to French men, not women, while reintroducing colonial slavery in an effort to reconstruct a French empire. “So he talked about how you build a system…where you can kind of control the people,” Altman reflected. “I was like, ‘Wow. I’m glad he does not run the United States ’cause that is a dude who understands something deep that I did not and clearly was able to use it for power.’ ” Six years after my initial skepticism about OpenAI’s altruism, I’ve come to firmly believe that OpenAI’s mission—to ensure AGI benefits all of humanity—may have begun as a sincere stroke of idealism, but it has since become a uniquely potent formula for consolidating resources and constructing an empire-esque power structure.

“When we originally set up the Microsoft deal, we came up with this thing called the sufficient AGI clause,” a clause that determined the moment when OpenAI would stop sharing its IP with Microsoft. “We all think differently now,” he added. There would no longer be a clean cutoff point for when OpenAI reached AGI. “We think it’s going to be a continual thing.” The two companies would continue to partner and release ever-advancing technologies—at a great price. It was a bizarre and incoherent strategy that only made sense under one reading: OpenAI would do whatever it needed, and interpret and reinterpret its mission accordingly, to entrench its dominance.

The following month, the publication confirmed more details. Altman was considering a few different scenarios: one could be transitioning OpenAI to a traditional for-profit; the other would be transitioning it to a for-profit public benefit corporation like Anthropic and xAI. Both scenarios would retain the existence of the nonprofit as a separate entity but dismantle its board’s control over the company’s business. Under this new structure, investors were also pressuring Altman to take equity in the company to align his incentives more directly with their own.

After the Omnicrisis, the competition facing OpenAI had only accelerated. Musk was expanding his computing capacity at an alarming pace to build xAI. Anthropic’s latest version of Claude was pulling customers away from ChatGPT. Sutskever had officially formed his new rival company, Safe Superintelligence, and had only just announced a starting $1 billion in funding. At the same time, after over a year of work, OpenAI was still struggling to attain the desired performance for Orion to justify its release. The company was beginning to stare down the barrel of an uncomfortable prospect: Its tried-and-true formula of scaling no longer seemed to be enough to work; to advance its AI systems further, it likely needed fundamentally new research ideas. This was far easier said than done in general, but even more so after OpenAI had spent years orienting its hiring and team organization around exploiting existing research rather than exploring uncharted science.

Anthropic began an ad campaign for Claude with a cheeky message on its billboards in San Francisco: “The one without all the drama.”

In 2021, I came across a story that felt different from any that I’d ever reported: the story of an Indigenous community in New Zealand that was using AI to revitalize te reo Māori, the language of the Māori people.

Large language models accelerate language loss. Even for models several generations earlier like GPT-2, there are only a few languages in the world that are spoken by enough people and documented online at sufficient scale to fulfill the data imperative of these models. Among the over seven thousand languages that still exist today, almost half are endangered according to UNESCO; about a third have some online presence; less than 2 percent are supported by Google Translate;