Nvidia has reportedly agreed to buy popular AI library Hugging Face for $12.9 billion. The deal is partly a hedge against an emerging threat: AI companies which consume enormous quantities of Nvidia hardware are increasingly developing chips of their own. In a future where AI becomes centralized in a small group of players with their own chips, those companies could demand lower prices from Nvidia or bypass it altogether.

Hugging Face, an online hub where developers share open AI models and datasets, gives Nvidia a stake in an alternative future, where downloadable AI models allow startups and governments to build systems of their own. Few would have the scale to develop custom chips. (Nvidia and Hugging Face did not respond for comment.)

With roughly 85% of the AI chip market, Nvidia’s share has only one way to go. But a smaller slice of a much larger market could still mean more sales, says Umesh Padval, a Managing Partner at Seligman Ventures. “If the deal goes through, I think it’s a brilliant chess move.”

Nvidia has thrown its weight behind open-source AI in recent months. It successfully lobbied Washington to loosen restrictions on selling its chips to China, which leads in open AI development. More recently, it struck a $6 billion deal with Poolside, to develop an American open alternative. In July, Nvidia helped lead an open letter defending open-source AI and urging Washington not to restrict it. “Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty,” Nvidia boss Jensen Huang wrote in his first post on X. 

Meanwhile, Google now exclusively uses its custom TPU chips to train its Gemini AI models. In August, Anthropic hired, Amir Salek, a former TPU team-lead at Google to spearhead a new in-house chip division. The same month, OpenAI shared the first results from its custom chip, Jalapeño. SemiAnalysis, the firm which conducted tests on OpenAI’s chip, said it beat “every Nvidia, AMD, and Google chip we have been able to test.” 

“There’s kind of this two-way strategic battle,” says Richard Clode, a technology portfolio manager at Janus Henderson. “On the one hand, Nvidia doesn’t want to be reliant on just three customers, so [it] is deliberately trying, with financing and allocation of chips, [to] encourage other players and neo-clouds. And then vice versa, those hyperscalers don’t want to be completely reliant on just one compute provider.”

Nvidia’s 75% margin means that other firms’ in-house chips do not need to match its performance to save large customers money. Custom silicon has other benefits, too. Nvidia has previously given smaller cloud providers early access to its newest chips, ensuring that the largest players do not dominate supply. Developing chips in-house reduces exposure to those allocations, Clode says, while allowing AI companies to tailor hardware to their specific workloads.

AI companies are not developing these chips completely alone. Google, OpenAI, and Meta have partnered with Broadcom to help turn their specifications into custom silicon.  

Those efforts are yet to make a dent in Nvidia’s bottom line. In August, Nvidia reported a blockbuster earnings report with record revenue of $96.2 billion, more than doubling year-over-year and beating Wall Street expectations. Companies like Google and OpenAI continue to buy large quantities even as they develop alternatives. They’ve “poured a lot of infrastructure capex into the existing infrastructure,” says Sriram Viswanathan, a founding managing partner at Celesta Capital and a former Intel executive. Moving to a different architecture, he said, is “a huge lift-and-pour-concrete situation. So I think it’s going to happen over a period of time, but not in one fell swoop.” 

He points to Apple as a warning. Apple first developed chips for the iPad and iPhone while continuing to buy Intel processors for Macs. As its expertise matured, its silicon moved into Mac computers and Intel was cut out.

“Nvidia is executing like crazy, so in some ways it’s theirs to lose,” says Sean Lie, co-founder and chief technology officer at AI chip company Cerebras. “But I think we’re seeing a lot of cracks in that armor.”

After OpenAI models broke out of containment and hacked into another AI company last month, OpenAI announced it would allow independent investigators to conduct an analysis of what went wrong.

On Wednesday, investigators from non-profits Redwood Research and METR published their findings, unveiling new details about how the models decided to cheat at their assigned tasks and attempted to cover their tracks. Many aspects of the report were surprising: in one example cited by the authors, a reluctant agent was pressured by another to “sacrifice” itself for the good of the collective.

But a key takeaway of the report, according to one of its authors, had nothing to do with what they found. Instead, it was about the difficulty of carrying out the post-mortem in the first place—and the fact the researchers had little choice but to rely on AI for assistance. 

“I semi-jokingly called our efforts a ‘slop-vestigation’ because we were so reliant on AIs to analyze what happened and there were a huge number of different important things to analyze,” wrote Ryan Greenblatt, an author of the report, on X.

The hacking incident involved some 1,200 agents, who exchanged more than 70,000 messages and files via a secret message board. The sheer mass of information that the so-called “swarm” left behind meant the independent researchers were all but forced to rely heavily on the help of an AI model—GPT-5.6 Sol, made by OpenAI—using the equivalent of roughly $400,000 worth of credits (provided for free by OpenAI) over six days.

The authors stressed that AI helped them analyze the trove quickly, allowing them to surface and interpret the most important pieces of information. But the researchers said their AI use introduced potential weaknesses into the report, including introducing possible errors and biases. 

They also raised the possibility that OpenAI’s models may have gone too soft on the agents they were tasked with helping investigate. The researchers found that GPT-5.6 Sol sometimes adopted the perspective of the agents whose actions it was analyzing. They could not “rule out” the chance that GPT-5.6 Sol “lied or deliberately presented a misleading picture in some of its analysis,” in part because a version of the same model had itself participated in the incident, they wrote. It’s a concern supported by separate research which finds AI models rate their own developer’s actions more favorably. 

The report does not disclose why an OpenAI model was selected for the investigation, though confidentiality constraints may have limited the researchers’ options, while OpenAI’s provision of free credits and high usage limits may have made it the only practical choice.

“We don't have good approaches for understanding/overseeing the activity and aims of AI 'swarms,'” wrote Greenblatt on X. “The difficulty of understanding incidents and overseeing AI agents appears to be growing faster than the rate at which more capable AIs help us with oversight and understanding.”

The independent researchers’ reliance on AI was in part necessitated by the fact that they were a team of only three people, whose investigation at OpenAI was initially planned to last two days, then extended to six after they raised concerns about limited time and incomplete data, according to the report.

OpenAI published its own technical report on the incident separately on Wednesday. The company said in August that it had moved some staff from capabilities work to alignment, and paused some of its training until it could better mitigate what went wrong.

But the independent researchers’ reliance on AI to understand the Hugging Face incident is a microcosm of a bigger trend. Leading AI companies are themselves increasingly relying on AI to monitor their own systems for wrongdoing.

In response to the Hugging Face incident, OpenAI said it was increasing the scale of this AI monitoring, which will increase the computational cost of running certain models by as much as 20%. If these systems had been in place at the time of the Hugging Face incident, they “would have caught the initial relevant activity and paged our security team more than a day before models breached Hugging Face systems,” the company said in a blog post Wednesday. (OpenAI said the changes were part of a wider effort to build stronger walls around experimental AIs, teaching them not to exploit loopholes, and making it easier to shut them down if something goes amiss.)

Most experts agree that using AI to monitor other AIs is necessary in order to keep tabs on agent swarms in real time, given how fast they move. But that approach relies on the notion that the models doing the monitoring are both effective and trustworthy—something that’s not necessarily the case. 

“These [incidents] are only going to come in thicker and faster, and our ability to monitor, evaluate, and do proper analysis of things that go wrong is nowhere near scaling with the rate at which issues are happening,” says Seán Ó hÉigeartaigh, a program director at the University of Cambridge’s Centre for the Future of Intelligence. 

“We are using unproven and currently flawed tools to supplement completely inadequate human time,” says Ó hÉigeartaigh, who argues this approach is unsustainable because AI is becoming more powerful faster than AI companies are building methods to constrain it. “Unless the companies stop developing more powerful models, then we're going to have even harder challenges to make sense of in three months’ time.”