Capital Is Withdrawing from “World Models”
Summer Never Lasts Forever
Hengyu from Aofeisi
QbitAI | WeChat Official Account QbitAI
The sensational startup saga of Tian Keyu may have been no more than a footnote to this year’s “world model” boom taking a sharp downturn.
Amid the public debate, venture capitalists’ demands regarding founders’ professional ethics became a topic of ongoing discussion and reflection both inside and outside the industry.
But beneath the surface, the “world model” track on which Tian Keyu was betting has shifted from a torrent of speculative capital in the first half of the year to a shortage of Series A deals in the second half. Capital is indeed raising the bar for world model investment memos, while speculative money is dwindling and so, too, is the number of world model founders who can secure funding. Had Tian Keyu not been the “Best Paper Award winner at NeurIPS,” the outcome might have been different.
Because starting this summer, “We’re done investing in world models.”
This shift was especially evident in QbitAI’s recent conversations with several frontline investors and heads of investment and financing at world model companies.The number of world model projects that leading funds are now willing to pursue can be counted on one hand.
We asked, “Can any companies still raise funding relatively smoothly?”
Some can, folks. Some can.
An investor who had made several moves this year thought for a moment and then told usthe names of four or five founders.
After further thought, they added a well-funded embodied AI startup—because it was founded by the children of a major tech company and focused solely on “the brain,” it was also placed in the world model camp.
“Anyone else?” I asked.
“That’s it. Just remember these few.”
Remarkably, the names offered by other interviewees did not exceed this lean and exclusive shortlist either.
The impressions shared by several heads of investment and financing at world model companies were strikingly consistent.
Funding is becoming increasingly difficult to secure.
No one has written off the technological prospects of world models. Some investors are still actively searching for the next company worth backing.
But the market’s conditions have definitely changed markedly since the first half of the year.The valuation premium that the four words “world model” once brought across the board is rapidly shrinking—and shrinking at breakneck speed.
The window for raising capital on the basis of concepts and imagination is steadily narrowing. Teams that can present technological evidence, research credentials, or even tangible application progress are still being eagerly sought out by investors.
As for everyone else…
The market’s patience is running thin. Everyone is bound to feel a little anxious.
Rapid Cooling
In fact,the winds began to shift from the height of summer.
Until this year’s WAIC in mid-to-late July, world models were undoubtedly the hottest and most eye-catching topic in tech.
It felt as though everyone had been saving up their big reveal for half a year, waiting for the chance to show off at WAIC all at once.
A large number of startups unveiled new models, technological advances, and demos, with some systems allowing users to explore freely in AI-created environments as if they were playing a game.
For ordinary audiences, this was certainly novel enough.
Butfor industry insiders who had already seen dozens of projects, that novelty was rapidly fading.
The more they saw, the more obvious the problems—even common problems—became. As a result, after WAIC and WRC drew to a close, the previously lively project talks and partnership discussions also began to cool down.
That said, honestly speaking, this cooling-off cannot be blamed entirely on world-model companies or companies associated with the concept of world models.
Capital investment itself was a little overheated in the first half of this year.
Many investment firms were worried about missing out on the next star company and had already made their investments while the projects were still at an early stage.
To bet on AI, a leading RMB-denominated fund made more investments in the first half of the year than it had in any comparable period throughout its more than decade-long history.
A leading US-dollar-denominated fund had already exhausted its entire “budget” for AI investment in 2026 during the first half of the year.
The spectacle of consecutive investments in quick succession, existing shareholders repeatedly increasing their stakes, and companies raising three rounds of funding at once—with each round valued higher than the last—further inflated market expectations for the sector as a whole.
In addition,some US-dollar-denominated funds have entered a new fundraising cycle(with Sequoia, Hillhouse, Boyu, Qiming, and Monolith raising funds, while Starlink, Linear, and Chengyi, among others, announced the first close of their new US-dollar funds in the second half of the year). This will inevitably force them to reassess their capital reserves and investment pace.
Just then,the secondary market gave everyone a lesson as well.
On August 19, Unitree Robotics debuted on the STAR Market.
On its first trading day, Unitree’s share price briefly surged to RMB 1,100, taking its market capitalization to approximately RMB 444.9 billion. It then steadily adjusted downward, falling to RMB 404.03 at one point by October 15.
Although the stock rebounded sharply in late October, this roller-coaster ride served as a wake-up call for many investors in the embodied intelligence sector.
Unitree is, of course, not a world-model company.
But there is no denying thatfor institutions pricing embodied intelligence and physical-AI-related projects, Unitree provided an unavoidable public-market benchmark.
If even a leading robotics company with established products and a foundation for commercialization could face such dramatic swings in valuation, what were model companies that had yet to prove the value of scaled applications worth?
Investors began calculating this equation much more carefully.
Some people may wonder why, upon reading financing news from the second half of the year, the world-model sector still seems to be buzzing and money still appears to be flowing in.
The reason is that there is a time lag involved.
The announcement of a financing round often comes after the actual transaction has taken place.It is common practice in the industry to wait three months before making a public announcement.
The pace at which capital has been deployed has therefore already slowed, but transactions completed in the first half of the year are still gradually turning into financing news in the second half.
Although some companies are indeed still raising capital against the trend—for example,Force Infinite.
In August, this embodied-AI company announced the completion of nearly RMB 1 billion in Series A and Series A+ financing.In late October, it was revealed that the company had secured nearly 100 million yuan in an A++ round led by Zhejiang Venture Capital Group.
QbitAI has learned that the companyhas joined the ranks of unicornsand has swiftly launched a new funding round.
That said, companies able to raise capital at such a nonstop pace can hardly be taken as representative of the temperature of the entire market.
According to data from Cailian Press’s Venture Capital Database, there were five primary-market financing events related to physical AI and world models in October, based on incomplete statistics—a marked decrease from before.
Even when physical AI companies are included, financing events across the market have clearly become sparse.
As a result, changes in the world-model investment and financing market in the second half of the year have gradually become clear.
Some companies are still raising money—and raising it in large sums. But most projects no longer enjoy the kind of待遇 of being chased by investors that they did in the first half of the year.
That said, claiming that world models have completely gone cold would perhaps be unfair to them.
Rapid Divergence
While many projects were worrying about how to raise their next round, a small number of companies continued to secure funding with considerable ease.
And when they did, the amounts reached hundreds of millions of yuan.
Which companies are still receiving such treatment?
Looking back at the list provided by the investor mentioned at the beginning, some patterns do emerge.
First, the money.
In July this year, Natural Will, founded by Tsinghua University assistant professor Ding Ning, brought in new investors including Sequoia China, Ant Group, and CATL after completing a funding round in April.
In September, LiberAI, founded by post-00s entrepreneur Liu Songming, announced the completion of several hundred million yuan in Pre-A++ financing.
Wuyuan Technology, founded by Wu Haixu, had already completed multiple funding rounds just a few months after its establishment, reaching a valuation of hundreds of millions of US dollars.
Inverse Matrix, founded by Peking University graduates Chen Boyuan and Ji Jiaming, moved even earlier, officially announcing more than 100 million US dollars in seed++ financing in June. QbitAI has learned that Inverse Matrix’s latest funding round has also closed, although it has not yet been made public.
At first glance, these financing timelines seem somewhat at odds with the cautious attitudes of the investors mentioned earlier. But Tsinghua, Peking University, MIT... their backgrounds suggest a possible explanation.
Chen Boyuan founded Inverse Matrix at age 22 and had previously served as head of the Innovation Center for Behavioral World Models at the Beijing Academy of Artificial Intelligence. Ding Ning is an assistant professor and doctoral supervisor at Tsinghua University. During his undergraduate studies, Wu Haixu received Tsinghua University’s Special Scholarship for Academic Excellence before pursuing postdoctoral research at MIT. Liu Songming is also a recipient of Tsinghua University’s Special Scholarship for Academic Excellence and a post-00s entrepreneur.
Elite universities, youth, and impressive research credentials.
When world models took off in the first half of the year, candidates with exactly these attributes were the most sought after. Now, in the second half of the year, as investors begin to pick and choose, they are still selecting the same kinds of founders.
The answer is that only a few outliers can secure funding smoothly.
Moreover, beyond their founders’ credentials, each of these companies has something else of its own to show.
For example,Inverse Matrixhas released the Physis series of models, aiming to enable AI to learn state changes in the physical world and further predict and simulate them.
Wuyuan Technologyfounder Wu Haixu previously published GeoPT, which explored how to learn the laws of object motion from static geometric information and made a considerable impact within the field.
Both teams are researching how AI can understand the physical world, just from different technical perspectives.
The other two companies have taken somewhat different approaches.
Natural Will released Natura-1 in October, attempting to use more than a million hours of human first-person data to teach robots human-like dexterity.
LiberAI, meanwhile, unveiled the world action model Liber-0 this week. It is also betting on large-scale human experience, hoping to enable robots to transfer the manipulation abilities they have learned to more tasks.
There is one more especially distinctive name on the list.
Ant Lingbo.
It does not have the “child prodigy from Tsinghua or Peking University” startup story, but it is the offspring of a major tech company. It has already launched several LingBot series models, and its computing power, engineering capabilities, and application resources also differ from those of startups starting from scratch.
The overall investment and financing logic is starting to take shape.
The conceptual dividend from world models is fading, but teams that have already produced results in relevant research directions, possess unique technical accumulations, or can mobilize substantial R&D resources still have a chance to secure funding.
Much of this funding is, however, still a bet on the future—the technological path of world models has not converged, and both products and business models remain under exploration.
That is why companies known in the first half of the year for “having no model, no product, and merely talking about concepts” can still occasionally secure funding.
Take another highly controversial young entrepreneur,Tian Keyu, as an example.
…
Looking back, in the first half of the year, every company was busy proving that it was working on world models. Even companies that did not originally build their reputation around world models were trying to prove that they could develop them.
By the second half of the year, investors began scrutinizing their choices more carefully.
And after much careful selection, they returned to top universities, star researchers, and young academic talents.
It is almost funny to read this: in the end, the story seems to have come full circle, with everyone placing their bets on child prodigies again…
But does this not align with investment logic?
Inflation in the “child prodigy” label does not mean that truly exceptional young talents will be overlooked.
Finally, it is worth emphasizing that this listis merely one market snapshot obtained through QubitBit’s latest interviews. It is far from an authoritative ranking, and it does not mean that these companies have already won.
Nevertheless, it should offer a glimpse of the investment market’s mainstream attitude.
Concept Inflation and the Standard-Bearers “Coming Ashore”
Looking back, there were certainly good reasons why world models became the hottest trend this year.
Language models and embodied intelligence are advancing rapidly, but AI’s ability to understand, predict, and interact with the real physical world still faces numerous challenges.
If models can develop an internal understanding of how the world operates and simulate in advance the consequences of different actions, AI’s ability to enter the physical world could reach a new level.
This is precisely why so much hope has been placed in world models.
Moreover, the field is being led by influential AI researchers such as Li Fei-Fei and Yann LeCun. Tech giants including Google and Nvidia are also investing in related technologies.
The capital boom in embodied intelligence has, in turn, provided world models with vast potential for applications.
Research breakthroughs, visions for the field, star entrepreneurs, investments by tech giants, and demand from the robotics industry—all converging quickly heated up the market.
Funnily enough, even though this technological direction has allowed countless companies to raise enormous amounts of capital, the industry has still not reached a unified understanding of what it actually is.
Generating photorealistic videos, simulating three-dimensional physical states, and predicting robot actions can all fall within the scope of world model research, but they address very different problems and are evaluated by vastly different standards.
During the period when the concept of world models was so overwhelmingly popular that it left people dizzy, these distinctions were easily overshadowed by one sweeping technological concept.
Meanwhile, even though the conceptual boundaries remain unclear, the modifiers can keep growing longer and longer.
Physical world models, spatial world models, causal world models, action world models...
If one word isn’t enough, add another.
After all, if one adjective isn’t enough to highlight uniqueness and cutting-edge nature, just add another. This isn’t done to subdivide the field, but merely to “highlight uniqueness and cutting-edge nature.”
Thus, the joke that “everything can be a world model” began to circulate.
In June, Li Fei-Fei and her startup World Labs published an article titled “A Functional Taxonomy of World Models,” attempting to establish a clearer classification for this widely used concept by dividing it into three categories: renderers (Renderer), simulators (Simulator), and planners (Planner).
But it doesn’t seem to have helped much.
The scramble to acquire “world models” nevertheless showed no signs of subsiding.
On October 11, The Wall Street Journal reported that Amazon was in talks to acquire the world model startup Decart for approximately $7 billion.
Decart had previously developed Lucy, a real-time video editing model, and Oasis, a system capable of generating simulated environments.
In Amazon’s calculations, the company’s appeal lay not only in its generative models, but also in its AI infrastructure and its technical capabilities for running across different chips.
One practitioner’s lament left a particularly deep impression on us:
It really feels like watching the sunset over the sea.
The sun is about to set, but hasn’t quite done so... There’s a sense of grandeur, a touch of solemnity, and every now and then, a hint of warmth.
One More Thing
At the end of September, AMD announced that it planned to acquire World Labs, the world model company founded by Li Fei-Fei, for $8.2 billion.
For a startup founded just over two years earlier, an $8.2 billion acquisition offer is considerable.
But not everyone in the world model community was pleased.
Li Fei-Fei enjoys globally top-tier research prestige, while World Labs has a model development team, technical expertise, and the ability to raise capital. Yet its fate is to become part of a chip giant’s AI infrastructure strategy.
For AMD, the deal also brings cutting-edge model research capabilities, helping the company understand future AI workloads and optimize its chips, software, and systems.
If even Li Fei-Fei and her team—one of the field’s two standard-bearers—ultimately chose to join AMD for $8.2 billion, weren’t people’s earlier imaginings about the valuation of world models a little too extravagant?
Some even began to question Li Fei-Fei’s original intentions in founding the company and agreeing to be acquired...
Regardless, this event became a landmark for world models.
It may mark the beginning of a reshuffling, or perhaps the turning point of an era.
World models have been on the rise for less than three years, but it is clear that they must—and inevitably will—undergo more robust validation in real-world applications, as well as demonstrate their value through a narrative that can withstand scrutiny.
Who could have imagined that the reshuffling would arrive so violently, with autumn seemingly descending overnight?
