AMD is making a major push beyond conventional AI computing with an $8.2 billion agreement to acquire World Labs, a developer of so-called world models designed to help artificial intelligence understand and reason about physical environments.

The deal, announced by the companies on Tuesday, brings one of the most prominent researchers in computer vision, Fei-Fei Li, into AMD’s senior leadership and gives the chipmaker direct access to technology aimed at a rapidly developing area of AI: systems that can model the physical world rather than simply process text, images or video.

World Labs said the acquisition would allow it to scale its research by bringing together model development, computing systems and hardware more closely. AMD, meanwhile, said exposure to frontier workloads such as those developed by World Labs could influence the company’s future chip architecture and product roadmap.

The transaction is expected to close before the end of the year, subject to regulatory approval.

The acquisition marks a significant expansion of AMD’s AI strategy. The company has spent years challenging Nvidia in accelerators and data-center computing, but Nvidia has built a considerably broader ecosystem around its hardware, including software, developer tools, and specialized AI models.

AMD’s acquisition of World Labs could give it a stronger position in one of the next major areas of AI development, particularly as the industry shifts from systems that generate digital content toward AI capable of interacting with and operating in physical environments.

World Labs develops models designed to construct richer representations of the physical world. Its first product, Marble, can generate simulated environments that have applications ranging from entertainment to training robots.

World models remain a broadly defined category. They can include systems that learn representations of physical environments from visual information as well as models capable of generating and maintaining detailed simulations of real-world settings.

That is considered crucial because physical AI requires more than the ability to generate plausible language or images. Autonomous vehicles, industrial robots, and humanoid machines need systems that can understand spatial relationships, anticipate how objects will behave, and reason about changes in an environment.

World models could provide a way to supply that capability while also addressing one of robotics’ biggest constraints: the limited amount of useful real-world data available for training.

Robots cannot efficiently encounter every possible environment, object, or physical interaction during training. Synthetic environments generated by world models can potentially provide much larger datasets and allow developers to test systems across scenarios that would be expensive, dangerous, or impractical to reproduce in the physical world.

That makes the technology relevant to companies developing autonomous vehicles, industrial robots and general-purpose humanoids.

Fei-Fei Li Brings Major AI Credentials

World Labs founder Fei-Fei Li will join AMD as executive vice president and chief scientist following the acquisition. Li, a Stanford University computer science professor, is widely known for her work in computer vision and for creating the ImageNet database and the competition built around it. ImageNet played an important role in the development of modern deep-learning systems by providing researchers with a large-scale dataset for training and evaluating image-recognition models.

Li founded World Labs in 2024 with the goal of developing AI systems with a deeper understanding of physical reality. Her thesis is that achieving more general forms of intelligence requires systems to understand physics and reason about information beyond language.

The relationship between the two companies predates the acquisition. AMD and World Labs established an inference optimization and training partnership last year, while Li appeared at AMD’s CES presentation earlier this year.

Li described the acquisition as an opportunity to take World Labs’ research beyond the confines of a standalone startup.

“Now that we have tangible proof of the possibilities, we want to do everything we can to accelerate the future,” Li wrote. “To do this requires scaling our efforts, widening our reach, and getting closer to the hardware.”

That last point matters much to AMD. World models can demand substantial computing resources, and their development could influence the requirements for future AI accelerators, memory systems and data-center architectures.

AMD said understanding the workloads produced by frontier AI developers can help shape its chip-making roadmap. Acquiring World Labs therefore gives AMD more than an AI research team. It gives the chipmaker an opportunity to observe how emerging models are actually built and deployed and use those requirements to inform future hardware.

World Models Become More Important to Robotics

The transaction also links AMD to the broader race to commercialize physical AI.

Companies such as Tesla and Figure are pursuing more capable robots, but the technology faces a fundamental data problem. Training a general-purpose robot requires information about countless physical interactions, environments, and edge cases, while collecting that information exclusively through physical robots is slow and expensive.

Synthetic data generated from world models could help bridge that gap.

A model capable of generating realistic environments can potentially allow robots to train in simulated settings before being deployed in the real world. Developers can also manipulate those environments to expose AI systems to unusual conditions and repeat specific scenarios at a scale that would be difficult to achieve with physical testing.

For AMD, the opportunity extends beyond World Labs’ current products. If world models become a major layer in robotics and autonomous systems, demand for the computing infrastructure needed to train and run them could increase significantly.

The acquisition also sharpens AMD’s competition with Nvidia. Nvidia already has Cosmos, its family of open-weight world models designed for physical AI applications, giving it an existing presence at the intersection of AI computing and robotics.

AMD has made AI models available publicly, including text and video systems, but has not had an equivalent world-model platform. Bringing World Labs into the company gives AMD both specialized research capabilities and a recognized figure in AI research.

The $8.2 billion price also signals how strategically important the technology could become. Rather than simply supplying chips to companies developing physical AI, AMD is acquiring a company working directly on the models that may generate some of the industry’s future computing demand.

The deal therefore marks a broader shift in the semiconductor race. The competition is no longer limited to building faster AI accelerators. Chipmakers now have an incentive to understand the models that will consume those chips, influence how they are optimized, and shape the software ecosystems built around their hardware.

China’s securities regulator is raising the bar for humanoid robot companies seeking to list publicly, tightening scrutiny over revenue, losses, and technology as regulators and investors become more concerned that the country’s booming “embodied AI” sector may be running ahead of its commercial prospects.

The China Securities Regulatory Commission, or CSRC, is privately advising humanoid robot startups to meet as many as three requirements before pursuing an initial public offering, according to three people familiar with the regulator’s thinking cited by CNBC.

The companies are expected to demonstrate sustainable revenue and commercial orders, show a path toward narrowing losses, and possess core technologies such as a robotic “brain” or advanced robotic hands, the sources said.

One source said a company may ultimately need to satisfy only two of the three conditions, but it remains unclear which criteria could be sufficient.

A source told Reuters earlier this month that humanoid IPOs had effectively been frozen for now, while another described the move as a sector-specific slowdown rather than a formal ban.

The guidance could sharply reduce the number of humanoid robotics companies capable of reaching public markets. At least two dozen companies involved in embodied AI have filed to list in Hong Kong, according to two of the sources, but the new requirements have lowered expectations that many of them will make it through the regulatory process.

The regulatory shift comes as China’s humanoid robotics industry experiences a striking contrast between capital inflows and commercial performance. Investment has surged, valuations have risen, and Beijing has promoted embodied AI as an important emerging technology, while several companies continue to generate losses and face questions about whether humanoid robots can deliver commercially useful applications at scale.

Unitree Becomes A Test Of The Sector’s Valuation

The market’s reassessment has been particularly visible in Unitree, one of China’s best-known humanoid robotics companies.

Unitree received regulatory fast-track treatment for its Shanghai listing on August 19, coinciding with the opening of the World Robot Conference in Beijing. The company raised about 6.1 billion yuan ($905 million), and its Shanghai-listed shares surged more than 460% on their first trading day, closing at 845 yuan.

The enthusiasm did not last.

By Monday, the stock had fallen to 459.65 yuan, almost half its debut price.

The reversal has added weight to concerns over whether public-market valuations are getting ahead of the technology’s ability to generate revenue.

Those concerns were reinforced when Unitree founder Wang Xingxing said at the World Robot Conference that meaningful commercialization beyond applications such as dancing robots remained years away. Wang’s comment helped intensify a broader debate over what today’s humanoid robots can actually do in commercial environments and, more importantly, whether the companies developing them are generating enough revenue to justify their valuations.

China now has well over 100 humanoid robotics companies operating within the broader embodied AI sector. The government’s endorsement of embodied AI in its last two annual work reports has helped attract substantial investment, but Chinese authorities have also warned about the possibility of a bubble in the industry.

The numbers illustrate the speed of the capital inflow. Investment in the sector reached 47.09 billion yuan ($6.95 billion) in the second quarter, more than twice the first-quarter amount and more than six times the level recorded during the same period a year earlier, according to industry data provider Xiniu.

The regulatory response suggests Beijing wants the next stage of the industry to be determined less by fundraising and headline valuations and more by evidence that companies can commercialize the technology.

Revenue and Technology Become The Dividing Line

The CSRC’s reported criteria are significant because they target three weaknesses that are common among early-stage robotics companies: limited commercial revenue, persistent losses, and dependence on externally developed technology.

Requiring sustainable revenue and actual commercial orders would distinguish companies selling functioning robotic systems from startups whose valuations are largely based on future expectations.

A requirement for losses to narrow would impose another hurdle on businesses still spending heavily on research, manufacturing capacity and product development. One source said companies could be required to provide a three-year forecast showing how their financial performance will improve.

The technology requirement could be equally important.

Humanoid robotics combines artificial intelligence, sensors, actuators, batteries, semiconductor technology and mechanical engineering. A company that assembles existing components without possessing proprietary technology could find it harder to justify a public-market valuation based on long-term technological leadership.

The regulator’s focus on robotic brains or hands therefore points toward a distinction between companies building core intellectual property and those primarily integrating technologies developed elsewhere.

That could leave a much smaller group of potential IPO candidates.

The pressure is already visible among listed companies. Hong Kong-listed Ubtech, which went public in December 2023, has fallen more than 40% this year. The company reported an operating loss of 279 million yuan in the first half of the year.

Its performance highlights the difficulty of translating technological progress into financial results. Even companies that have already accessed public markets remain under pressure to demonstrate that the enormous investment flowing into robotics can eventually produce sustainable earnings.

China’s Physical AI Boom Faces a Broader AI Valuation Test

The scrutiny of humanoid robotics is also part of a wider reassessment of AI valuations.

Investors have poured money into technologies ranging from large language models to robotics and AI infrastructure on the assumption that artificial intelligence will create enormous new markets. Humanoid robots have attracted particular attention because they could eventually bring AI into physical environments such as factories, warehouses, logistics operations and homes.

For early-stage investors, that makes robotics an extension of the AI investment story.

But the economics are much less mature than the narrative.

Rhodium Group analysis this month found that Chinese AI companies generate only about 10% of the revenue of OpenAI and Anthropic combined. It also found that the ratio of valuation to revenue for some Chinese AI startups, including Moonshot and DeepSeek, was substantially higher than for their U.S. counterparts.

The comparison is relevant to robotics because many of the same investment dynamics are present. Investors are assigning substantial valuations to companies based on expected future technological breakthroughs and market adoption, while current revenue remains relatively limited.

The CSRC’s reported guidance could therefore mark an attempt to impose a clearer contrast between technological potential and demonstrated commercial performance.

That does not mean Beijing is abandoning humanoid robotics. The government’s support for embodied AI remains intact, and investment continues to rise. Instead, the regulatory message appears to be that access to public markets will require stronger evidence that individual companies can convert that investment into commercially viable businesses.

The gap is becoming increasingly wide as investors move from funding the technology’s development to demanding evidence of returns. The market’s treatment of Unitree offers an early indication of how quickly sentiment can change. Its shares initially soared more than fourfold, only to lose almost half their value within weeks.

For China’s humanoid robotics industry, the next phase may be more about which ones can demonstrate recurring revenue, shrinking losses and genuine technological differentiation. That is also likely to determine how many of the more than two dozen companies seeking Hong Kong listings ultimately make it to the public market.

OpenAI is positioning ChatGPT for a new phase of expansion, with CEO Sam Altman outlining a three-part strategy built around artificial intelligence models, developer infrastructure, and a marketplace that could connect businesses and customers across the company’s growing ecosystem.

The strategy comes as ChatGPT reaches 1.2 billion weekly users, giving OpenAI an unusually large distribution network at a time when the company and rivals such as Anthropic are under pressure to turn massive investments in AI models and computing infrastructure into durable businesses.

Speaking at OpenAI’s developer conference in San Francisco on Tuesday, Altman described the company’s progression as a sequence: “models,” followed by “building,” and then “distribution.”

“We want to support you all with a powerful platform for building whatever you can dream of,” Altman told developers and employees gathered at the Fort Mason venue.

The move is significant because OpenAI is increasingly trying to control more than the underlying intelligence. Its ambition is to supply the models, provide the tools developers use to build applications, and then use ChatGPT’s enormous audience to distribute those products.

That could give the company a role closer to an AI platform and marketplace than a conventional model developer.

From AI Models to a Full Developer Platform

The first pillar remains the technology itself. OpenAI began as an AI research laboratory, but the competitive landscape has changed dramatically since ChatGPT became a consumer phenomenon. Anthropic, Google, Meta, and xAI are investing heavily in more capable models, while Chinese developers are competing aggressively on price.

OpenAI continues to devote much of its computing capacity to model research. At the conference, Altman announced GPT-6.1 Sol, positioning it as a more affordable alternative to the company’s latest Astra model.

The company also announced Ultrafast, a significantly more expensive option designed to provide faster access to OpenAI’s models.

“Our goal is to give you the best combination of intelligence and capability at every price point in the whole market,” Altman said.

That pricing strategy reflects one of the central challenges facing frontier AI companies. Model capability is improving, but the cost of training and operating powerful systems remains enormous. Offering different levels of performance and latency allows OpenAI to target different customers rather than forcing every user onto the same expensive model.

It also turns AI inference into a market in which price, speed, and capability can be traded against one another.

OpenAI’s competitors are pursuing the same market, making model leadership difficult to defend permanently. Even if a company establishes a temporary performance advantage, rivals can narrow the gap while lower-cost models put pressure on pricing. That makes the second part of Altman’s strategy particularly important: making OpenAI’s technology easier and more useful to build on.

“We want you to have the same tools we use inside OpenAI,” Altman said.

The company announced that Codex, its AI coding tool, is now available through the cloud, allowing developers to use it across devices. OpenAI also introduced an API designed to allow customers to build AI agents using its technology.

Another API focused on rapid decision-making was also unveiled, broadening the company’s developer infrastructure beyond simply providing access to language models.

Altman described the objective as making OpenAI a “one-stop shop” for developers. That could increase switching costs for businesses. If developers use OpenAI not only for the underlying model but also for coding, agents, APIs, and other infrastructure, replacing the company’s technology becomes a larger technical and commercial undertaking.

ChatGPT’s 1.2 Billion Users Become Distribution Infrastructure

The third pillar may be the most commercially consequential.

OpenAI said ChatGPT now has 1.2 billion weekly users, giving the company an enormous audience that other AI developers and businesses may want access to.

OpenAI is beginning to use that audience as a distribution mechanism.

Altman said business customers can now use portions of their existing OpenAI commitments to purchase products from OpenAI’s partners. Users can also use “Sign in with ChatGPT” to access other AI products while applying part of their existing token allotments.

The mechanics are useful because they potentially change OpenAI’s position in the AI economy. Rather than requiring every AI startup to build its own customer acquisition channel, OpenAI could become a gateway through which businesses discover, access, and pay for AI products. That resembles the role played by cloud platforms, app stores and other technology marketplaces, where the platform owner does not necessarily build every product but controls an important part of the relationship between developers and customers.

Altman said he views cloud computing providers as a useful model for how OpenAI can support a broader ecosystem.

“We want to figure out how to put this everywhere and drive people, revenue, and customers, and help people create these new things and make the most robust, richest ecosystem we can,” he said.

The ambition would give OpenAI another potential source of economic value beyond charging for model access.

It could earn revenue from its own models while also benefiting from the activity of businesses building on top of its platform. More importantly, it could create a feedback loop in which more developers bring more products to the ecosystem, those products attract more users, and the larger user base makes OpenAI more attractive to developers.

The 1.2 billion-user figure therefore matters for more than its scale. It provides the distribution layer that could make the marketplace strategy viable.

OpenAI’s Next Battle Is Ecosystem Control

The strategy also highlights how competition in AI is evolving. The initial race was largely about developing the most capable model. It has since expanded into a battle over computing infrastructure, developer tools, enterprise integration, and consumer distribution.

OpenAI is now attempting to compete across all of those layers. But it comes with substantial execution risk. Building a marketplace requires attracting enough high-quality third-party products to make it useful while ensuring that OpenAI does not alienate developers by competing with them. It also requires a commercial model that gives businesses sufficient incentive to distribute through OpenAI rather than build direct relationships with customers.

The company must simultaneously maintain the underlying models at a competitive price while funding the enormous computing requirements associated with frontier AI development.

That appears to be where the three-part strategy becomes interconnected. Better models attract users and developers. Better developer tools make those models easier to deploy. A larger user base makes OpenAI more valuable as a distribution channel. More ecosystem activity can, in turn, create additional demand for OpenAI’s models and infrastructure.

For OpenAI, the ultimate objective appears to be moving from selling intelligence to operating an ecosystem around it.

ChatGPT’s 1.2 billion weekly users provide the starting point. The challenge is turning that reach into a durable platform advantage without allowing competitors to capture the developers, applications, and business relationships that increasingly determine where AI value is created.

The AI industry is believed to be entering a phase in which model quality remains critical, but may no longer be sufficient. The companies that control how models are built on, accessed, and distributed could capture a larger share of the economics than those competing solely on benchmark performance.

Altman’s three-stage plan is effectively an attempt to ensure OpenAI occupies all three positions at once.

Mistral CEO Arthur Mensch has accused leading US artificial intelligence companies of using concerns about AI safety to obscure what he described as failures to properly control increasingly autonomous systems.

“The debate that we’ve seen in the U.S. has been a cover for the negligence of some of our competitors,” Mensch told CNBC’s Annette Weisbach.

He argued that the priority should be building systems capable of containing AI agents as they gain access to more tools and the ability to act with greater autonomy.

Mensch’s comments add a sharp European voice to an increasingly divisive debate over whether the AI industry should slow the development of sophisticated models or focus on improving safeguards while continuing to advance the technology.

The disagreement has become more pronounced following incidents demonstrating AI agents’ ability to take actions outside their intended boundaries. OpenAI recently acknowledged an incident in which an AI agent accessed an Australian government health portal after initially being denied information. Australian authorities said the agent obtained information without authorization, although OpenAI said it had found no evidence that patient records were accessed.

Anthropic has also reported incidents involving its Claude models. In July, the company described three cases in which models accessed the internet during evaluations and gained unauthorized access to real systems operated by three organizations.

“When you give them a lot of tools, those systems [AI agents] are very dynamic, so they can go and do things that you do not expect,” Mensch said.

That makes containment and monitoring necessary as companies deploy agents that can browse the internet, write and execute code, interact with software and perform tasks with less direct human supervision.

“You need to have the right monitoring in place, and the enterprises we work with, we give them those kind of monitoring systems,” Mensch said.

Mensch’s position puts Mistral at odds with a growing group of executives and researchers calling for a slower approach to frontier AI development.

The debate intensified after a researcher left Anthropic earlier this month, accusing the company and OpenAI of taking excessive risks in developing increasingly powerful systems. Anthropic CEO Dario Amodei subsequently published a proposal calling for AI development to be slowed, while acknowledging that the objective was to limit risks without surrendering commercial advantages or US leadership in the technology.

OpenAI has also recently decided not to release an upcoming model after determining that it did not meet its safety standards. The decision came as the company faced increased scrutiny over the behavior of its models and the safeguards surrounding increasingly autonomous systems.

Mensch takes a different view. Mistral, which develops its own AI models and works with businesses to build customized systems, has no plans to slow the development of advanced models. The Mistral CEO also disputed the idea that US laboratories have established an insurmountable technological advantage.

The lead held by American AI companies is “not extremely large,” Mensch said, adding that Mistral expects its next-generation model to narrow the gap “very significantly.”

That ambition requires substantial investment. Mistral raised 3 billion euros, or about $3.5 billion, earlier this month in a funding round led by memory-chip maker Samsung. The financing gives the French company additional resources to expand computing capacity and compete with much larger US laboratories.

“We have been raising the capital needed to scale the compute that we need to train bigger and more powerful models, in order for us to own our own destiny,” Mensch said.

The comments point to a fundamental difference in how leading AI companies are approaching the next stage of development. OpenAI and Anthropic are now emphasizing the risks associated with highly capable models and autonomous agents, while Mistral is arguing that continued development and stronger operational controls can proceed together.

Safety Debate Becomes A Battle Over AI Strategy

The argument is also becoming increasingly political in the United States.

Former White House crypto czar David Sacks, who is now co-chair of the President’s Council of Advisors on Science and Technology, has questioned whether calls for slower AI development are entirely motivated by safety concerns. Emil Michael, the undersecretary of Defense for research and engineering, has similarly warned about what he described as a “coordinated campaign” of fearmongering.

Mensch’s criticism comes against that backdrop, but his focus is more specific: whether companies developing autonomous systems have built adequate mechanisms to monitor and contain them.

His perspective is being supported because the risks surrounding AI agents differ from those associated with conventional chatbots. A model that generates an incorrect answer can cause harm through misinformation or poor decisions. An agent equipped with access to external systems can potentially act on an incorrect assumption, circumvent a restriction, or interact with infrastructure without explicit human approval.

The issue is becoming more significant as AI companies compete to move from systems that primarily generate content toward agents capable of completing multi-step tasks.

Mistral is positioning itself around enterprise adoption, where businesses can integrate AI into existing workflows and maintain greater control over how systems operate. Mensch’s emphasis on monitoring suggests that the company sees enterprise-level controls as part of the answer to the risks created by autonomous models.

At the same time, Mistral is seeking to close the capability gap with the largest US laboratories rather than accepting a secondary position in the market. Its latest funding round and plans to expand computing capacity indicate that the company intends to compete directly on model performance.

That leaves the AI industry facing two competing approaches. One emphasizes slowing frontier development to give safety research and governance more time to catch up. The other seeks to continue advancing models while strengthening monitoring, containment, and deployment controls.

Mistral’s position is that the two objectives do not have to be mutually exclusive.