AI Isn't Coming for Hollywood. It Has Already Arrived

WIRED

a day ago

Aug 20, 2025 6:00 AM An early winner in the generative AI wars was near collapse—then bet everything on a star-studded comeback. Can Stability AI beat the competition? Photo-Illustration: Mark Harris; Getty Images
Lady Gaga probably wasn't thinking that a coup would unfold in her greenhouse. Then again, she was cohosting a party there with Sean Parker, the billionaire founder of Napster and first president of Facebook.
It was February 2024, and the singer had invited guests to her $22.5 million oceanside estate in Malibu to mark the launch of a skin-care nonprofit. One of the organization's trustees was her boyfriend, whose day job was running the Parker Foundation. In the candlelit space, beside floor-to-ceiling windows that looked out over the Pacific, Parker's people mingled with Gaga's, nibbling focaccia and branzino alla brace to music from a string quartet (Grammy-winning, of course).
Prem Akkaraju, one of Parker's close friends and business partners, arrived in a tailored suit, his thick hair coifed to perfection. The two men had known each other since Parker was at Facebook and Akkaraju was in the music industry. Over the years, they'd tried unsuccessfully to launch a movie streaming platform together and—much more successfully—had taken over a renowned visual effects company. Lately they had been talking about starting an AI venture.
That evening at Gaga's, Akkaraju found himself sitting next to an investor in Stability AI, the company that launched the wildly popular text-to-image generator Stable Diffusion in 2022. Despite its early success, Stability was 'circling the drain,' the investor recalls. It was 'within days of not having options.' He told Akkaraju: 'You should take Stability and make it into the Hollywood-friendly AI model.'
Hollywood did seem to be in need of a friend. Since 2022, the number of films and TV shows made in the United States had dropped by about 40 percent, thanks to ballooning production costs at home, competition from overseas, and long-running labor disputes everywhere. AI promised to bring the numbers back up by speeding production and slashing costs: Let computers automate the grunt work of translating dialog, adding visual effects frame by painstaking frame, and editing boom microphones out of a zillion shots. Maybe one day they could even write scripts and act! Two of the industry's biggest unions had gone on strike in part to obtain assurances that generative AI wouldn't replace union jobs in the near term. But every major studio and streaming service was racing to figure out its AI strategy, and a host of startups—Luma, Runway, Asteria—was working on tools to pitch them.
Akkaraju saw the opportunity in front of him. Stability AI had the technology. It just needed that Hollywood finish. As far as he could tell, there was only one problem. Didn't the company already have a CEO?
When Emad Mostaque, a former hedge fund manager, founded Stability in 2020, the company's mission was to 'build systems that make a real difference' in solving society's toughest problems. By 2022, the system Mostaque felt he needed to build was a cloud supercomputer powerful enough to run a generative AI model. OpenAI was gaining traction with its closed-source models, and Mostaque wanted to make an open source alternative—'like Linux to Windows,' he says. He offered up the supercomputer to a group of academic researchers working on an open source system where you could type words to generate an image. The researchers weren't going to say no. In August of that year, they launched Stable Diffusion in partnership with Mostaque's company.
The text-to-image generator was a breakout hit, garnering 10 million users in two months. 'It was fairly close to state-of-the-art,' says Maneesh Agrawala, a computer science professor at Stanford University. Openness was core to the model's success. 'It allowed researchers to essentially extend the model, fine-tune it, and it spurred a whole community into action in terms of creating enhancements and add-ons,' Agrawala says. By October 2022, Stability AI had only 77 employees, but with thousands of times that many people in the wider Stable Diffusion community, it could compete with its bigger rivals. Mostaque raised $101 million in a seed round from venture capital firms and hedge funds including Coatue and Lightspeed (the final million, he tells me, was for good luck). The company was a unicorn. Photo-Illustration: Mark Harris; Getty Images
Employees from this period describe Mostaque as a visionary. He spoke eloquently about the need to democratize access to artificial intelligence. In the not-too-distant future, Mostaque told employees, the company would solve complex biomedical problems and generate season eight of Game of Thrones . 'It was an incredibly fun and chaotic startup that was throwing a lot of spaghetti at the wall, and some of it stuck really hard,' a former high-ranking employee tells me. (Like others I spoke with, the employee requested anonymity to speak freely about Mostaque and the company.)
Mostaque was thrilled by the success. But he was also in over his head. 'I was brand-new to this,' he says. 'With my Aspergers and ADHD, I was like, 'What's going on?'' Mostaque talks fast, his tone matter-of-fact: 'On the research side, we did really good things. The other side I was not so good at, which was the management side.' Two former employees told me that they felt Mostaque didn't think deeply about building a marketable product. 'He just wanted to build models,' one said.
The company's success brought heightened scrutiny—particularly around how the models were built. Like many text-to-image models, Stable Diffusion 1.5 was trained on LAION-5B, an open source dataset linked to 5.8 billion images scraped from the web, including child sexual exploitation material and copyrighted work. In January 2023, Getty Images sued Stability AI in London's High Court for allegedly training its models on 12 million proprietary photographs. The company filed a similar suit in the US weeks later. In the stateside complaint, Getty accused the AI firm of 'brazen theft and freeriding.'
Then, in June 2023, Forbes published a blockbuster story alleging that Mostaque had inflated his credentials and misrepresented the business in pitch decks to his investors. The article also claimed that Mostaque had received only a bachelor's degree from Oxford, not a master's. (Mostaque says that he earned both, but a clerical error on his part was responsible for the mix-up.) What's more, Stability reportedly owed millions of dollars to Amazon Web Services, which provided the computing power for its model. Though Mostaque had spoken of a partnership, Stability's spokesperson acknowledged to Forbes that it was in fact a run-of-the-mill cloud services agreement with a standard discount.
Mostaque had answers for all of this, but investors lost confidence anyway. Four months after the article came out, VCs from both Coatue and Lightspeed left the board of directors, signaling they no longer had faith in the business. By the end of the year, the company's head of research, chief operating officer, general counsel, and head of human resources had left as well. Many of Stability's prominent researchers would follow. Under pressure from investors, Mostaque finally left the company on March 22, 2024—just a few weeks after Lady Gaga's greenhouse soiree.
Akkaraju and Parker wasted no time in taking over Stability, installing Akkaraju as CEO and Parker as chairman of the board. They never spoke to Mostaque, although the former CEO says he reached out to offer his support.
The pair set about trying to remake Stability AI for the moment. Not long after they took over, the competition got fiercer. That September, another startup, Runway, signed the AI industry's first big deal with a movie studio. Runway would get access to Lionsgate's proprietary catalog of movies as training data and develop tools for the studio. 'The time it takes to go from idea to execution is just shrinking—like a lot,' says Cristóbal Valenzuela, CEO of Runway. 'You can do things in just a couple of minutes that used to take two weeks.' In the coming years, he predicts, 'you will have teams of two, three, four people making the work that used to require armies and hundreds of millions of dollars.'
The deal with Lionsgate pushed the AI-fication of Hollywood into overdrive. 'I can tell you, last year when I came to Los Angeles versus today, it's night and day,' says Amit Jain, CEO of Luma, another Stability competitor. 'Last year it was 'Let's prototype, let's proof-of-concept'—they were deferring the inevitable. This year it's a whole different tone.'
Moonvalley, an AI company founded by former Google DeepMind researchers (and the parent company of Asteria, an AI film studio cofounded by the actor Natasha Lyonne), recently told Time magazine that more than a dozen major Hollywood studios are testing its latest model—signaling openness to the technology, if not yet a full embrace.
'It was really about me and Sean coming in and providing that direction, that leadership, and really taking advantage of what we call the three T's: timing, team, and technology,' Akkaraju says.
I'm sitting not at his TED Talk but in his $20 million mansion near Beverly Hills, on an immaculate overstuffed white couch overlooking a manicured garden. Akkaraju is fit, with a gleaming white smile and a button-up that shows off his biceps. His eye contact and handshake are equally strong.
Early on in his tenure, Akkaraju says, he decided that Stability would no longer compete with OpenAI and Google on building frontier models. Instead, it would create apps that sat on top of those models, freeing the company from enormous computing costs. Akkaraju negotiated a new deal with Stability AI's cloud computing vendors, wiping away the company's massive debt. Asked for specifics on how this came about, Akkaraju, through a spokesperson, demurred. Investors, however, came flocking back.
Where Mostaque painted a picture of AI solving the world's most difficult problems, what Akkaraju is building, in brutally unsexy terms, is a software-as-a-service company for Hollywood. The goal is not to generate films, he says, but to use AI to augment the tools that filmmakers already use. 'I really do think that our differentiation is having the creator in the center,' Akkaraju says. 'I don't see any other AI company that has James Cameron on its board.'
Yes, the irony writes itself: The guy who once had a fever dream about murderous machines while 'sick and broke' in Rome and proceeded to turn it into The Terminator —the creator of Skynet!—is on the board of an AI company. What's doubly surprising, though, is that Cameron is on the board of an AI company run by Parker and Akkaraju. A decade ago, Cameron was helping lead Hollywood's charge against them. He didn't appreciate the premise of their streaming platform, the Screening Room, which let people watch new releases at home for $50 on the same day they came out in theaters. Cameron reportedly told a crowd at CinemaCon that he was 'committed to the theater experience.' In the years that followed, none of the major studios publicly announced deals with the Screening Room, and in 2020 the company rebranded as SR Labs. Photo-Illustration: Mark Harris; Getty Images
That same year, Akkaraju and Parker took over Weta Digital, the visual effects studio behind blockbusters such as Lord of the Rings , Game of Thrones , and Cameron's Avatar movies. Weta developed virtual cameras that let Cameron see a real-time rendering of the artificial environment through a viewfinder, as if he were filming on location in the fictional world of Pandora.
One night, Cameron, Akkaraju, and Parker met for dinner to discuss how technology was changing the film industry. 'The tequila was flowing,' Cameron recalls. 'A friendship formed.' Any tension that had existed over the Screening Room melted away. ('I never really talked with him about it,' Akkaraju says. 'He knew, and I knew. It was very funny.')
So Cameron is on the board, but is the 'creator in the center,' as Akkaraju said? When I spoke with Parker, he emphasized the importance of using open source models and spoke of 'respect for creators and respect for IP.' He added: 'That sounds potentially kind of rich, coming from me, given my past association with Napster and early social media. But it is a lesson learned.'
In June, the company scored a major win when Getty dropped its copyright infringement claims from a broader lawsuit as the trial neared a close in the UK. The US trial is ongoing. Akkaraju said the company 'sources data from publicly available and licensed datasets for training and fine-tuning,' and that when 'creating solutions for a client' it 'fine-tunes using the dataset provided by the client.' When I asked Akkaraju if the company trained exclusively on licensed data, he responded: 'Well, that's the majority of what we're using, for sure.'
Even those who are bullish on AI admit that, for the most part, the technology isn't ready for the big screen. Text-to-image generators might work for marketing agencies, but they often lack the quality required for a feature film. 'I worked on one film for Netflix and tried to use a single shot,' says a filmmaker who asked to remain anonymous, not wanting to discuss their use of AI publicly. The AI-generated footage got 'bounced back' from quality control because it wasn't 4K resolution, the filmmaker says.
Then there's the problem of consistency. Filmmakers need to be able to tweak a scene in minute ways, but that's not possible with most of the image and video generators on the market. Enter the same prompt into a chatbot 10 times and you will likely get 10 different responses. 'That doesn't work at all in a VFX workflow,' Cameron says. 'We need higher resolution, we need higher repeatability. We need controllability at levels that aren't quite there yet.'
That hasn't stopped filmmakers from experimenting. Almost every person I spoke with for this story said that AI is already a core part of the 'previz' process, where scenes are mapped out before a shoot. The process can create new inefficiencies. 'The inefficiency in the old system was really the information gap between what I see and what I imagine I want moving forward,' says Luisa Huang, cofounder of Toonstar, a tech-forward animation company. 'With AI, the inefficiency becomes 'Here's a version, here's another version, here's another version.''
One of the first people in Hollywood to admit to using generative AI in the final frame is Jon Irwin, the director and producer of Amazon's biblical epic House of David . He became interested in the technology while shooting the first season of the show in Greece. 'I noticed that my production designer was able to visualize ideas almost in real time,' he says. 'I was like, 'Tell me exactly how you're doing what you're doing. What are you using, magician?'' he recalls.
Irwin started playing around with the tools himself. 'I felt directly tethered to my imagination,' he says. Eventually, he made a presentation for Amazon outlining how he wanted to use generative AI in his production. The company was supportive.
'We film everything we can for real—it still takes hundreds of people,' Irwin tells me. 'But we're able to do it at about a third of the budget of some of these bigger shows in our same genre, and we're able to do it twice as fast.' A burning-forest scene in House of David would have been too expensive to do with practical effects, he says, so AI created what audiences saw.
Irwin says he has spoken with the team at Stability but has 'not been able to use their tools successfully on a show at scale.' The comment reflects a theme I found in my reporting: While I was able to identify a number of filmmakers who admitted to toying around with Stability's text-to-image generators, none used the tools professionally—at least not yet.
Contains AI-generated imagery. Courtesy of Stability AI
The taboo on studios acknowledging their embrace of AI seems to be softening. In July, Netflix co-CEO Ted Sarandos told investors the company had allowed 'gen AI final footage' to appear in one of its original series for the first time. He said the decision sped up production tenfold and dramatically cut costs. 'We remain convinced that AI represents an incredible opportunity to help creators make films and series better, not just cheaper,' he said.
Hanno Basse, Stability's chief technology officer, is showing me an image of his backyard in Los Angeles: a grassy lawn surrounded by high hedges, rose bushes crowding a bay window, and a tree in the far left-hand corner. Suddenly, the 2D image unfurls into 3D. A generative AI model has filled in the gaps, estimating depth (how far away the hedge is from the rose bush, the tree from the window) and other missing elements to make the scene feel immersive. Basse can replicate camera moves by selecting from a drop-down menu: zoom in or out, pan up or pan down, spiral.
'Instead of spending hours or days or weeks building a virtual environment and rehearsing your shots, the idea here is actually that you can just take a single image and generate a concept,' Basse says.
Contains AI-generated imagery. Courtesy of Stability AI
Rob Legato, Stability's chief pipeline architect, seems pleased. A veteran visual effects specialist who worked on Wolf of Wall Street and Avatar , Legato joined the company in March. He was up until 2 am the night before shooting a film and has arrived at this meeting to act as both a company executive and a beta tester.
The only issue, Legato says, is the drop-down menu. 'You probably want to combine them and have a slider,' he says.
Contains AI-generated imagery. Courtesy of Stability AI
Stability AI's offerings are still in their early days. Even Legato admits the version of the virtual camera tool we are looking at has a ways to go before it could be used by a professional. 'Right off the bat my job is unfortunately to be critical,' he says.
The conversation drifts to rotoscoping. Legato explains that this process, where an artist sketches over a scene frame by frame, used to take hundreds of hours and was reserved for entry-level animators. Now AI can automatically isolate part of an image and add visual effects. 'You'd never want your child to work on roto,' he tells me.
The comment is meant to sound optimistic, but it gets to a looming fear about how AI will impact Hollywood. Namely, that the technology will lead to widespread job losses.
'I hear artists at VFX companies say, 'Hey, I don't want to get replaced.' Of course you don't want to get replaced!' says Cameron. 'If you guys are going to lose your jobs, you're going to lose your jobs over the work drying up versus getting bumped aside by these gen AI models.' The idea, echoed by Akkaraju and Parker, is that as movies become cheaper to produce, more films will get made and overall employment will rise.
When pressed on this point, Akkaraju reverts to an extended metaphor. 'Every major transition or technological invention is always met with apprehension at first, and then acceptance, and then it's obvious,' he says. 'When ATMs rolled out in the '80s, all the tellers were really up in arms. They were like, 'That's our job. We give withdrawals, we take deposits, and now you're having this machine do it.' What's happened since then is that there are more teller jobs than ever before, and their average pay is higher, even adjusted for inflation.'
Whether the coup that began in Lady Gaga's greenhouse ultimately saves Stability AI, the AI revolution is here and already transforming Hollywood. That collapsing building, that burning forest, that crowd of people you see when you stream a show or go to the movie theater? One person with a keyboard could've made them. The thing about that bank-teller anecdote is that it's often used by techno-optimists—including Stability AI investor Eric Schmidt. What they don't mention is that the number of bank tellers peaked around 2015. Since then, it's been on the decline.
Let us know what you think about this article. Submit a letter to the editor at [e-mail].