Dear Aventine Readers, 

Last week the White House ordered Anthropic to suspend access to its most advanced AI models to all foreign nationals. Earlier in the spring, Anthropic had released its most powerful model, Mythos, but to only 50 customers. As large language models become ever more sophisticated, access is becoming stratified and will likely become more so. This week we look at what that could mean going forward and who the gatekeepers of the most advanced AI models might be.

Also in the issue:

Thanks for reading! 

Danielle Mattoon 
Executive Director, Aventine

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

The past seven days have made one thing clear about access to advanced AI: Most of us won’t have a say in who gets to use it. 

Last Friday, the White House ordered Anthropic to suspend access to its Fable 5 and Mythos 5 AI models to all foreign nationals. Citing national security concerns, the directive forced Anthropic to shut down the models entirely, as there was no way to distinguish between foreign nationals and citizens using the models within the United States.

The government’s restrictions are notable for being both sudden and extreme, but they follow a pattern established by AI companies themselves. In February 2019, OpenAI chose to release only a modified model of GPT-2 out of concerns that the full model could be used maliciously (it released the full model later that year). This year, worried about releasing a model that it claimed could “surpass all but the most skilled humans at finding and exploiting software vulnerabilities,” Anthropic initially gave access to Mythos to only 11 companies and organizations (and limited access to 40 more). Two months later, on June 2, it expanded access to 150 institutions across 15 countries. Its release of Fable 5, a version of Mythos with guardrails intended to make it safe for public use, went live for a few hours last week before it was shut down. 

Not surprisingly, both Anthropic’s early restrictions around Mythos (which outraged British banks and annoyed US utilities) and the government’s order to limit access to it have not been popular with those outside the tent. European politicians said that limiting access to US nationals was a wake-up call about the geopolitical realities of AI access. 

Yet gatekeeping is likely to become more entrenched, not less, according to experts who spoke with Aventine. As models grow more sophisticated, chances increase that they could be used for nefarious purposes like cyberattacks and bioterrorism, making the case for more restrictive access. At the same time, as AI companies struggle to find enough computing power, they might need to constrain access for reasons of capacity as much as trust. 

This points to a potential near future in which the world's most capable models are rationed. By whom, and according to what criteria, are newly pressing questions. 

Buying time

The idea of giving a model to a limited group of early users isn't new, said Lennart Heim, a researcher on AI policy who until recently worked at RAND. AI companies routinely share new models with groups such as the US Center for AI Standards and Innovation and the UK AI Safety Institute for purposes of testing a few weeks ahead of launch. (UK AISI was the only non-US organization to get early access to Mythos.) 

Nor was Anthropic’s approach to the Mythos release a new one, at least to AI insiders, said Janet Egan, a senior fellow at the Center for a New American Security. In the AI research community, she said, the idea of limiting access to models as their capabilities became more advanced and potentially dangerous had been “percolating for a while.” 

The logic of differentiated access is to give defenders early, controlled access to a cutting-edge tool so they can build resilience before it circulates freely. While many companies may have been affronted by being excluded from early access, there are open questions around which organizations have the capability to make use of such a tool. “A hospital isn’t going to know what to do with it,” said Raffi Krikorian, chief technology officer of Mozilla. Heim made a similar point: An organization needs "the capability to actually do something with it, to evaluate it, to deploy defenses." For now, it may not matter much that a given bank or government department lacks direct access to Mythos, provided it runs systems built by a vendor that does. 

But Heim said Anthropic’s longer window of restricted access, combined with the model's alarming security implications, made the decision feel like something new. And because the tiered access and reasons for it were headline news, for many outside the AI industry it was the “first very clear warning” of what’s coming down the pike, said Egan. 

New gatekeepers

Hanging over this is an obvious question: Who will be the gatekeepers of this new technology? “At the moment, the issue is that AI companies have a lot of power to decide what they think is dangerous,” said Egan. “That's kind of a weird role for a private company to play.” 

Anton Leicht, a fellow with the Carnegie Endowment's Technology and International Affairs program, would like those determining access to define a set of criteria any company or public institution must meet to qualify for access to frontier models. The appeal of this approach is that it formalizes and clarifies the requirements for access, removing the potential for unfairness. "It's then not a question of which partners the lab picks," he said. 

A second force may sharpen the problem: There may not be enough AI computing power to go around. If demand for AI outstrips supply, providers will have to choose which workloads to take on and when — a different kind of preferential access.

To an extent, this is already happening. AI companies offer different membership plans to users, from free to $20 a month to enterprise contracts. But so far, pricing hasn't been enough to manage demand and AI companies are looking for creative ways to manage the imbalance and guarantee their customers access. 

A natural endpoint, Leicht said, is access by willingness to pay. "Whoever is able to build around these models and derive the greatest surplus value is able to bid the highest price on AI tokens," he said, "and therefore they get access to these AI systems." But in that world, what about the university researchers who can't afford to run frontier models to hunt for new drugs or advance mathematics? Society might prefer that certain users — critical infrastructure operators, scientific researchers and so on — not have to outbid banks for access, Krikorian pointed out. But then someone has to decide who qualifies for a discount. Who will that be? 

America First

Private companies don’t necessarily want to be the ultimate arbiters of access. In conversations with people on the AI security and safety teams of various labs, Egan said, many have told her they would welcome clearer government guidance on what an acceptable level of risk actually looks like.

Obviously, we are seeing some of that guidance now, with the government’s decision to limit noncitizens’ access to Mythos. And a recently signed executive order will create a voluntary predeployment testing scheme for new models: When a lab has a new system it shares it with the government, and if the model clears a certain threshold for cyber capability, the government gets exclusive access for up to 30 days. The administration's recent actions are indicative of “a step change in how government perceives AI capabilities,” said Egan. 

Given these recent moves, experts who spoke with Aventine think that it’s almost inevitable that the US will impose America-first requirements that supersede everything else. 

“The US will be making calls and influencing these companies, whether [on a] voluntary [basis] or not, to adhere to US national interests first,” said Egan. Leicht tends to agree: “It depends a little bit on what administration is in power,” he said, “but I think it seems likely to me that in something like the current political environment, you'd see American-buyers-first policies being instantiated over time.”

What such a regime would look like is for now hypothetical, but the shape is not hard to imagine. Domestic access might itself be ranked: American government first, American companies next, American users after that. Leicht imagines a percentage of AI capacity reserved for domestic use, with the remainder available for export. Inside importing nations, tokens for frontier models might be scarce enough that those countries have to run tiered regimes of their own. 

The decisions about who participates in the defining technology of our age now look like they could become concentrated faster, and more aggressively, than many people previously realized. “The scale and scope of this [situation] is pretty monumental,” said Egan. The situation around Mythos, she said, was a very clear demonstration of how the discussions around access to AI are changing. “We shouldn't expect this to go away.”

Underwater data center before it was submerged. Getty

China is using the world's first wind-powered underwater data center. More than six miles off the coast of Shanghai, 33 feet beneath the waves, is a giant yellow cylinder stuffed with computer servers. The facility, developed by HiCloud Technology in partnership with Chinese authorities, is powered by electricity from nearby offshore wind turbines. According to The Guardian, it provides 24 megawatts of computing capacity — modest compared to some of the 100-megawatt-plus AI data centers now built on land, but still large enough to be commercially useful. Its biggest advantage is cooling: Seawater dissipates heat more efficiently than air. HiCloud launched a commercial underwater data center off the coast of Hainan in 2023, but the Shanghai project is the first to pair the concept with offshore wind power, creating a facility that is both cooled and powered by its surroundings. In 2018, Microsoft submerged a small experimental data center with one-hundredth of the capacity of the Shanghai facility off the coast of Scotland, though it has since closed down the project to focus on less exotic approaches. Others are exploring similar oceanic data center ideas. Panthalassa, a renewable energy company based in Portland, Oregon, is developing spherical offshore data centers powered by wave energy. Nautilus Data Technologies, based in Pleasanton, California, is building floating data centers for harbors that tap into seawater cooling, offshore wind generation and submarine internet cables. As the public sours on the buildout of data centers on land, taking the systems to sea now seems far less audacious than it might have a decade ago.

Evidence is mounting that smartphones affect birth rates. When we saw one paper published on this topic, we were interested but skeptical. Now there’s a second, suggesting that there may be something to the theory. The first paper, published in May on the Social Science Research Network by researchers at the University of Cincinnati, examined smartphone adoption and teenage fertility across 128 countries using World Bank data. The team found fertility declines tended to accelerate once smartphones reached mass adoption across countries with vastly different healthcare systems, abortion laws, cultures and religious traditions. A second study, published in June by researchers at Middlebury College through the National Bureau of Economic Research, exploits the fact that the iPhone was exclusive to AT&T between 2007 and 2011. By comparing fertility trends in US counties with strong AT&T coverage against those with little or none, the researchers estimate that the iPhone may account for as much as half of the fertility decline observed during that period. There are plausible explanations as to why this might happen: Smartphones may reduce in-person socializing and therefore sex; they also make information about contraception and abortion more accessible. But this debate is far from settled. Neither paper has yet been peer reviewed, and because no one can randomly assign entire populations to use or not use smartphones, proving causality will always be difficult. Still, the fact that two different research approaches result in similar conclusions suggests that there may be something to the theory.

Can’t find qualified workers? Train your own. While much of the conversation around AI is focused on job loss, companies building the infrastructure behind AI can't find enough people to do the work. That's why Meta has announced what it describes as the largest private-sector commitment to the trades in US history. The company is desperate for electricians, fiber technicians, welders, mechanics and other skilled workers needed to build the vast network of data centers underpinning its AI ambitions.The new initiative, called America's Workforce Academy, is backed by $115 million and will launch in Indiana, Louisiana, Ohio and Texas. Participants, who need no prior experience, complete a free five-week training program, earn money while they learn, and are guaranteed a job at the end. Graduates, according to Meta, will receive industry-recognized credentials, including certification from the National Center for Construction Education and Research. The company expects high demand: Earlier this year, it launched a fiber-installation training program and says it received 35,000 applications in its first week. According to a recent report by Lightcast, a labor market data provider, job postings for construction roles mentioning data centers rose 23 percent from September 2025 to February 2026 compared with the previous six months, and have roughly doubled over the past two years. Demand for technicians and engineers has grown at a similar pace. 

The “steroid olympics” were a circus—and a window into our culture, from MIT Technology Review
5,500 words, or about 22 minutes

The Enhanced Games made plenty of headlines when they took place in late May. Athletes were allowed to use performance-enhancing drugs, with the promise that shattered world records were a near certainty. That didn't happen. Held in a $50 million temporary arena built in a Las Vegas casino parking lot, the event featured swimmers, sprinters and weightlifters competing while following doctor-supervised, FDA-approved drug protocols. The results were underwhelming: The only standout achievement came from swimmer Kristian Gkolomeev, who broke the 50-meter freestyle world record while wearing a now-banned supersuit. But this story by Amit Katwala, who has followed the Enhanced Games story for years, argues that sport was never really the point. The company behind the Enhanced Games has evolved into a sort of telehealth longevity business, selling testosterone, peptides, supplements and other performance-enhancing products through subscriptions. Viewed through that lens, Katwala argues, the Games looked less like a sporting competition and more like a marketing campaign. And its most successful piece of marketing may be that one swimmer, Megan Romano, swam the 50-meter freestyle faster at age 35 than she had at 22 even though she failed to break any records. As Katwala puts it, “Making a 35-year-old feel 22 again is probably the perfect marketing message for the products Enhanced wants to sell.”

Hard rain, from Harper’s Magazine
8,900 words, or about 34 minutes

The problem with water isn't that the planet lacks it. It's that it often falls in the wrong places. This piece follows Rainmaker, a $31 million startup founded by Augustus Doricko, a Berkeley dropout and Thiel Fellow, who believes cloud seeding — releasing substances into the atmosphere to promote rainfall — can help combat drought. During the pandemic, Doricko returned home to Texas, found religion and came to view a passage in Genesis as a call to help humanity manage water more effectively. Now Rainmaker is trying to solve some of the problems that have plagued a technology that always sat on the fringes of science and public policy. It is building cheap autonomous drones (to make it cheaper than using planes) and developing sophisticated sensing equipment (to track whether its efforts are having the desired effect). That final part is still difficult to answer. While early results from the company's work in Utah's Bear River Basin appear promising, Rainmaker acknowledges that it can’t yet reliably distinguish seeded precipitation from rainfall that would have occurred naturally. What makes this story particularly compelling is that it discusses the technology in the context of some intense cultural anxieties surrounding the idea. Running through the piece is a parallel narrative about conspiracy theories claiming that governments already control the weather. Those claims are baseless, but they point to something important: Many people are deeply uncomfortable with the idea of deliberately manipulating the atmosphere, and Rainmaker will have an uphill battle ahead even if its technology does work.

The bold plan to save the world's coral, from New Scientist
2,600 words, or about 11 minutes

Between 2023 and 2025, 84 percent of the world's coral reefs experienced bleaching. Triggered by marine heat waves, bleaching occurs when the symbiotic relationship between coral and algae is damaged. Coral doesn’t necessarily die — it can recover if conditions improve — but repeated exposure to stress increases the likelihood that entire reef systems collapse. Researchers are looking for ways to intervene before that happens. Some are breeding corals in laboratories to increase genetic diversity and resilience in the wild. Others are cultivating algae that can better withstand high temperatures, then introducing them to reefs. Still others are experimenting with probiotics and nutritional supplements to strengthen coral health. And in a money-is-no-object Saudi-backed project in the Red Sea, researchers are experimenting across a square kilometer of reef, with a coral nursery reportedly 12 times larger than any other in the world. The entire site is mirrored by a detailed digital twin updated every six months using data collected by autonomous underwater vehicles. Some researchers worry that the field is becoming too enthusiastic about engineering its way out of a crisis. But the Red Sea trials may be a helpful proving ground, helping develop approaches that could be deployed far more widely.