Sign up for our newsletter about national security here.

In recent weeks, Silicon Valley has been fawning over an AI model released by a lab in China. The program, called GLM-5.2, has been called a “marvel,” “very good,” and a “step change.” The billionaire venture capitalist Marc Andreessen posted on X that “AI insiders are saying GLM-5.2 is the first Chinese AI model to match and often beat” the top public U.S. models. Guillermo Rauch, the CEO of the AI-coding platform Vercel, said that he is “genuinely impressed, almost shocked” by GLM-5.2’s coding abilities. Or as one AI founder at a San Francisco dinner party recently told me: “Praise GLM-5.2.”

In a sense, GLM-5.2 is China’s answer to Claude Code, Anthropic’s agentic assistant that has reshaped the AI boom. This has been the year of AI agents—tools that don’t just chat but promise to do stuff on your behalf, whether coding a website or booking a vacation. Chinese AI models have been steadily improving, but none previously proved capable or consistent enough to be used as agents. Now GLM-5.2, developed by the Chinese company Z.ai, rivals some of OpenAI’s and Anthropic’s top offerings and, by many measures, has leaped ahead of Google Gemini. And it is several times cheaper.

[Read: Move over, ChatGPT]

Despite Silicon Valley’s awe over GLM-5.2, an inexpensive competitor couldn’t have come at a worst time for America’s frontier AI labs. Having successfully persuaded corporate America to give their products a try, OpenAI, Anthropic, and Google are now struggling to prove that their tools are worth the money. These bots can be very expensive to use, running up bills into the thousands of dollars, per employee, per month. Uber reportedly spent its entire 2026 budget for Anthropic models in only a few months. Other Big Tech companies including Meta, Amazon, Tesla, and Adobe are also reportedly clamping down on employee AI usage. Citi at one point shut down employee access to OpenAI’s and Anthropic’s most expensive models, according to reporting from 404 Media (which Citi has contested).

Even though political leaders and tech executives have framed the U.S. as in the middle of a contentious technological race with China, Americans can still use Chinese models. While it’s too soon to know whether GLM-5.2 is really capable of replacing America’s top-tier AI agents, any firm or developer who is balking at the costs now might have an alternative. The arrival of GLM-5.2 poses a business dilemma for Silicon Valley—and possibly a national-security dilemma for the country as a whole.

In some ways, America’s AI industry has been here before. Many previous impressive Chinese AI models have not triggered a mass exodus of OpenAI and Anthropic customers, with one major exception: In January 2025, after DeepSeek launched a cheap AI model on par with America’s best, adoption of Chinese AI models leapt up. Within two months, the share of global web traffic to Chinese AI models jumped from roughly 3 to 13 percent, according to research from RAND. OpenAI, Anthropic, and Google quickly responded to DeepSeek with cheaper models of their own, and then the rise of AI agents early this year seemed to solidly put American labs back on top.

[Read: China’s DeepSeek surprise]

Even before Z.ai launched its new model, there have been quiet signs of a shift to cheaper, Chinese models amid mounting concerns about AI bills. Over the first five months of 2026, DeepSeek’s adoption among the 70,000 U.S. firms that use Ramp, a financial-operations platform, increased from 0.1 to 0.3 percent, Ara Kharazian, Ramp’s lead economist, told me. Kharazian added that 6 percent of Ramp customers that spend money on AI use third-party platforms that provide access to many different AI products. On OpenRouter, one such platform, the six most popular AI models are Chinese; in under a month, GLM-5.2 already ranks fifth. These data don’t capture all the software developers and companies directly downloading Chinese AI models, which are typically open source, and configuring them on their own computers—that is, largely scrappy start-ups and academics who don’t have the budget to use a fintech service like Ramp. Chinese models accounted for nearly half of all open-source AI downloads from February 2025 to 2026, according to data from the popular AI platform Hugging Face.

Meanwhile, OpenAI and Anthropic keep releasing more, rather than less, expensive products. And soon, they will likely have even more competition. The Claude Code and broader AI agent frenzy started seven months ago, which is also approximately how far behind Chinese AI firms have been lagging in model development. China’s other labs, including DeepSeek and Moonshot AI, will almost certainly release similarly capable and inexpensive AI agents before long. Coinbase, a popular crypto company, claims to have nearly cut its AI spending in half by defaulting to cheaper models including GLM-5.2 and Kimi, another popular Chinese bot. “The scenario to worry about is China has good-enough models at a quarter of the price,” Kyle Siler-Evans, an AI researcher at RAND, told me. “I think that is likely the future we’re headed toward.”

Kharazian cautioned against overstating this narrative. Whereas some tech-forward firms are pulling back on their AI spending, most American companies already spend very little on the technology: The median Ramp customer spends just $11 per employee on AI. “Not to say that the rest of the market won’t go in that direction,” Kharazian said, but Anthropic and OpenAI will have “ample time to respond” with competitive pricing—as they did with DeepSeek.

Perhaps the biggest obstacle to the ascendance of GLM-5.2, at least in the U.S., will have little to do with model capabilities or pricing. There are serious concerns about Chinese firms essentially using their AI to hoover up sensitive data and steal corporate secrets. The perceived risks of using Chinese technology could create a profound chilling effect among the customers of any U.S. firm—not to mention the uncertainty created by the possibility of federal regulation. Consider that Chinese electric vehicles are by all accounts better and cheaper than their Western counterparts, but they can’t be bought in the U.S.; one possibility is a near future in which Americans are functionally or legally barred from accessing the most cost-effective AI models.

The biggest implications of GLM-5.2 and the coming onslaught of cheap Chinese AI agents may also be geopolitical rather than economic. Whatever soft power the U.S. gains from its technology may wither as software developers in countries with less tense relationships with China turn to DeepSeek and GLM-5.2 in greater numbers (consider how the EU, U.K., and Canada are importing Chinese EVs). Meanwhile, despite Big Tech spending ever greater sums on advanced AI chips to train their models—chips that are banned from export to China—the gap between U.S. and Chinese models has not widened; if anything, it may be shrinking. That means any military, economic, cyberoffensive, or other advantage that AI can grant the nation could be vanishing. America’s lead in the AI race, for the first time since DeepSeek, is at real risk of slipping.

On a Friday in April, I hopped into an Uber to a fish market in San Francisco with a couple of tech founders on a mission to buy lobsters. Not for dinner, but for science: The duo dreamed of one day altering human consciousness, but they would start by toying around with some crustaceans. They intended to perform neurosurgery on the lobsters in the hopes of controlling them with an AI bot.

Leading the way was Elliot Roth, a bearded 32-year-old wearing a black T-shirt with Longevity printed across the chest and a silver chain with a double-helix pendant. To push the boundaries of the five senses, Roth has implanted a magnet in his left ring finger. He told me his nerves have grown around the magnet, giving him some sort of magnetoperception—he can feel when a microwave turns on in another room and sense when a radio tower is nearby. In the car, Roth took off his watch and allowed it to dangle, magnetized, from his finger.

Roth did not have any experience with lobster surgery, so he had enlisted a co-conspirator: William Joy, a lanky, 19-year-old redhead who also had never operated on a lobster but seemed confident in his fine motor skills. Joy would modify an off-the-shelf kit that can be used to remote-control a cockroach and implant the controlling device in the lobsters. If successful, Roth and Joy would then be able to send targeted electrical signals to direct the lobsters’ movements and, hopefully, the pinching of their claws. The final step: connecting the lobsters to the popular AI agent OpenClaw, which uses a lobster for its logo (get it?), and allowing the bot itself to decide what the lobsters should do. Perhaps the lobsters could even be made to control OpenClaw, Joy excitedly told me. “I’m pretty sure it’s going to be the first real instance of a complex AI agent interfacing with a biological organism,” he said during the ride.

With two lobsters secured in a pink plastic bag, we returned to an office building in downtown San Francisco where Roth and Joy would conduct the experiment. They began setting up an aquarium tank, fiddling with the water’s salinity and acidity. The duo claimed to be very concerned about animal welfare. “We are going to give a lot of thought to, How can we ensure that they don’t suffer?” Joy said, adding that he would look into giving them some kind of anaesthesia. These particular animals were “already destined for the dinner table,” Roth said, and they planned to later eat the lobsters out of respect.

This experiment was eccentric but not outlandish—at least not to Roth and Joy’s housemates. They both are residents of Biopunk House, a repurposed college dormitory inhabited by more than 20 aspiring entrepreneurs with dreams of curing aging and rewiring the human brain. In the lingo of San Francisco, this kind of co-living situation is called a “hacker house.” Out of sheer convenience, if nothing else, young tech founders have long decided to band together to make rent in one of the most expensive parts of the country. Mark Zuckerberg and some other early Facebook executives famously lived together in the 2000s.

In recent years, hacker houses have become more deeply woven into the Bay Area’s culture than ever before. During the AI boom, crashing in a group house is something like a rite of passage for any young AI founder. Across San Francisco, hacker houses are filling up with college dropouts and, in at least one case, exuberant tinkerers spending their Friday buying lobsters for science. They want to change the world and become spectacularly wealthy, but first they’ll have to share a bathroom.

A couple of days before the lobster mission, I rolled my suitcase up to the front door of a slightly battered Victorian row house a few blocks from San Francisco’s famous Painted Ladies. This was the Accelr8 house, where I would spend the next week among half-a-dozen entrepreneurs. What better way to understand the hacker-house scene than to live in it?

Accelr8 has a more modest setup than the Biopunk House. There’s no theme other than trying to launch a tech company with little to no funding. After I’d settled into my room, which included a wobbly standing desk and a view of the adjacent building’s wall, I got a brief tour of the house. It has two floors, each with three bedrooms, two bathrooms, and a kitchen. On the top level is a shared workspace with desks and a small bookshelf featuring Atlas Shrugged; The Art of Power, by Nancy Pelosi; some Foucault; and three vintage issues of Playboy. Below, a living room is crowded with couches, succulents, and kitschy art. Nearly every room in the house had an Amazon Basics odor eliminator (lavender), and red placards implored residents not to leave dishes in the sink, although dirty dishes were consistently there. That first day, as I put some yogurt away in a communal fridge, a faint sour smell wafted out. All things considered, this was one of the cleaner hacker houses I’d been to.

I soon began to meet the other residents. Michael Adams, a former rocket engineer with long, flowy hair, barged in while I was sitting in the workspace. “We’re going to the moon!” he said, and showed me a livestream of Artemis II hurtling toward space. During the pandemic, Adams had become concerned about and then obsessed with the failures of San Francisco’s local government. He quit his job at a space start-up and eventually built CivLab, a platform that, among other things, offers an interactive map that overlays local news, police reports, and community events across the city. Humans can comment on the site, and to further provide the atmosphere of a public forum, Adams had created a group of AI agents (with names such as “Max Entropy” and “Pearl Plexity”) to post about various happenings.

Adams and most of my other housemates were in or around their 30s. Kai Song Eer and Jay Yen Lim, who had just moved here from another hacker house in San Francisco, were building an AI agent to automate sales calls. On my first day, they offered me some (delightful) roasted oolong brought from Malaysia. Alan VanToai, whose room was on my floor, had an awakening to the potential of AI about a year ago when taking a psychedelic. He had recently vibe coded an AI-guided-meditation app.

And then there was Daniel Morgan, a 30-year-old Texan who lived in the room below mine. Morgan co-founded the Accelr8 hacker house with his friend Patrick Santiago in 2024, and it’s a sort of start-up in its own right. Each room in the house has a rent of about $3,000 a month, which Morgan told me was enough to basically break even on the lease and other expenses. Accelr8 has also launched another location, a “hacker hotel” that offers lower prices and fewer amenities, and it is where Santiago lives.

On my second evening at the house, I lounged in the backyard with Morgan and Santiago, surrounded by lawn furniture and a rusting barbell. It was a typical San Francisco evening: sitting outdoors to sip ginger beer, munch on sour-cream-and-onion Pringles, and talk about how AI is warping the world.

When it got cold, we went inside and settled on some couches in the living room. Morgan offered to cook what he described as “hacker-house slop bowls”—stir-fried noodles with eggs, mushrooms, and greens. While we waited, Santiago projected a YouTube show onto a large screen. It was a miniseries inspired by the residents of Crypto Castle, a hacker house from the 2010s. In one episode, a journalist visits and brings up Elizabeth Holmes, the Theranos founder who, at that point, had just been accused of running a massive scam. She may have made some mistakes, one resident of Crypto Castle said, but she “had a vision.” The hacker-house residents beside me laughed.

Hacker houses are popping up everywhere in San Francisco. “You could just walk a few blocks” in many neighborhoods, Amber Yang, an investor at Lightspeed Venture Partners, told me, and “there’ll probably be a few houses of start-ups.”

People who want to found an AI company have little choice but to head to Silicon Valley. Nearly two-thirds of the world’s AI-start-up investment, some $120 billion last year alone, is made in the Bay Area. The gold rush has made living in San Francisco even more difficult: Rents have risen roughly 15 percent over the past year and are likely to continue soaring as the public offerings of SpaceX, Anthropic, and OpenAI mint scores of AI millionaires. Adams put it bluntly: He and his partner could not afford their own one-bedroom apartment, so they moved into Accelr8. My stay at Accelr8 was itself motivated, in part, by a scramble to find a roof for a weeklong stretch between rentals in San Francisco. (I am not launching an AI company, but I did pay prorated rent.)

In the process, hacker houses themselves are changing. A number are becoming more professionalized and lavish. Down the block from Accelr8 is perhaps the most prestigious of them all, HF0, a mansion obscured by tall hedges that has raised a $100 million fund to invest in its residents. And the Biopunk House—where Roth and Joy, the two DIY lobster surgeons, live—is part of the Residency, a network of houses that counts Sam Altman as an adviser. The Residency collectively has an acceptance rate of 3 percent (and falling) and offers a three-month stay in homes where laundry, meal preparation, cleaning, social events, and basically any adult tasks are taken care of for you. “That’s 20 hours a week” back to dedicate to your company, Nick Linck, who founded the Residency, told me. We sat in the media room of one of his houses, a Noguchi table between us and fake ivy on the shelves behind.

The house where I stayed was hardly luxurious, but it was still a major step up from Accelr8’s original location, which had 15 bedrooms and a single kitchen for $1,500 a month. When I first met Morgan there last spring, I spotted a sign over the sink admonishing residents to Wash your pans or Sam Altman will get you. (Some things, I suppose, haven’t changed all that much; Adams would later dub me the “dish fairy,” on account of my putting away the dishes every morning on our floor.) Accelr8 just opened a second hacker hotel, which it markets as an all-inclusive residency—daily breakfast, full access to hotel amenities—and is trying to raise some $20 million to purchase another property to convert into a hacker house.

Too many companies, too many products, and too many venture capitalists are coursing through San Francisco. Hacker houses can help investors screen for genuine talent. Some houses and their investors take equity in residents’ start-ups. So-called angels—wealthy people who use their own money to fund companies—can drop by a hacker house, chat with someone in the living room, and cut them a $100,000 check on the spot. Still more venture capitalists visit for product demos.

For all of the pragmatic benefits of a hacker house, there is also a more soulful draw. Starting a company is lonely, time consuming, and risky. I met hacker-house residents who were spending down their life savings or had asked friends and family for loans. Most of them will fail. At the Accelr8 house, between sprints of coding and sales calls, the founders sat on couches and chatted about good date spots, or wandered to the bar next door to watch basketball. They were friends as much as they were housemates.

On my fourth day at Accelr8, Morgan barged into the kitchen carrying an armful of mugs that had been left out the night before. He said he sometimes felt like a “hacker-house dad.” After pouring himself coffee, Morgan took a gallon of 2 percent milk out of the fridge, smelled it, made a disgusted face, and put the jug right back where he found it. That afternoon, I ran into VanToai and Adams laughing in the kitchen. Adams was preparing coffee and had poured the remnants of a jug of 2 percent milk—that milk—into his glass without first smelling its contents. The milk, they joked, was “imbued” with the curdled whey. VanToai and I watched with incredulity as Adam tried to strain out the curdled bits. He gave up and reached for the instant coffee.

Every Sunday, the Accelr8 community members—both current and former residents—gather at the main house to demo their products at dinner. During my stay, it was hot-pot night. A couple dozen founders stood around two pots of simmering broth and ate thinly sliced beef, enoki mushrooms, and other fixings from paper bowls. With some, I discussed superintelligence and the coming automation of labor; with others, how superintelligence was mostly hype, AI is addictive, and sometimes agents “just make you feel like you’re getting somewhere” without accomplishing much. But everyone was enthusiastic to pitch their start-up.

Tigran Voskanyan, who was staying in Accelr8’s hotel, is developing an AI assistant that, he told me, centralizes your media, allowing you to access Spotify, Telegram, and so on in one place. Voskanyan asked me to pull out my iPhone and load his website so he could show me the program. This proved much slower than just opening my music app, and the bot had some voice-recognition difficulties. As I put my phone away, Voskanyan explained that his prototype is in its early stages, and he’s still working out some technical issues. Suddenly, he was interrupted by a robotic voice emanating from my jeans pocket. Voskanyan’s AI, despite being closed and my phone being asleep, was listening. I hurriedly revoked the bot’s microphone access.

Other products I heard about that evening included the crypto-based payments system Fumav, for “Fuck you, Mastercard, American Express, and Visa,” and a tool that was described to me, without irony, as “a new format” for the book. This turned out to be a text-based video-game app played on an e-reader, seemingly more akin to reinventing Oregon Trail. Matt Thompson, the founder who had made this tool, had hand-traced every illustration in his sample game to give it an artisan aesthetic. While we were cleaning up after dinner, Morgan told me that if he had the money, he would invest in Thompson’s start-up immediately.

The earnestness within these hacker houses is at times endearing and frequently confounding. Here is an ecosystem of people who generally fear the path that AI could lead us down and yet are eager to sell AI products, who see the first step in making the world better as conquering it with their technology, who earnestly believe that avarice is a prerequisite to altruism. In this city, for an idea to matter, it must exist as a product. To improve the world, you must turn people into users. Scale is a means to an end, until it isn’t. As Sourya Kakarla, a former Accelr8 resident I met at demo day, put it to me, he was “initially attracted to start-ups for the heart,” and a few years ago built an app to help elders in rural India use smartphones. But right now, he said, he “has to navigate short-term market realities to keep the heart alive.” He’s doing AI-consulting gigs to make money, like so many founders I met.

These tensions are baked into the hacker house itself, where young entrepreneurs scrap along together until they make enough money to leave the pack behind. Morgan was well aware of this dynamic, and expressed complex feelings about technology. One evening, walking back from dinner, Morgan lamented how in San Francisco, many people will stop mid-conversation and pull out their phone to request AI input. “How about what you, a human, have to say?” Morgan said, incensed. He would later tell me he’s worried about the “ChatGPT babies” growing up dependent on AI. But at the same time, he is spinning up his own AI-consulting firm.

On my last full day at Accelr8, I watched Morgan take a call with executives at a boutique construction company. Seated at the living-room table and wearing a dress shirt tucked into sweatpants that were outside his camera’s view, Morgan explained how he could build them a team of bespoke AI agents. “Sales people definitely need this,” one of the executives told him.

By then, Accelr8 seemed to be crumbling around me. After hot-pot night, the sink clogged and the upstairs dishwashers broke (the one downstairs wasn’t working when I arrived). I was holding my breath while opening the fridge. The morning I was set to leave, much of the crew gathered in the living room to see me off. I had graduated from Accelr8 and was added to an alumni group chat. Morgan handed me a souvenir, a glittering Accelr8 sticker for me to put anywhere. “Preferably a laptop,” he said.

I did still have some unfinished business. A couple of weeks after the trip to the fish market, I checked in with Joy to see how the lobster surgery had gone. We sat down in the same room where he and Roth had set up their experiment, but the tank beside us was conspicuously empty of any life or, for that matter, liquid. The lobsters had died before surgery. Perhaps they’d gotten the salinity of the water wrong, Joy told me. He had conducted preliminary dissections on their carcasses, and he said he was going through an ethical crisis about “whether I should even be doing this.” Nobody ate the lobsters.