The emergence of GPT-6 Astra has unexpectedly opened a breakthrough in solving the long-standing "brain" problem plaguing the embodied AI industry, bringing significant disruption.
In Astra’s real-world demo video, this general-purpose large model, not specifically trained for robotics, controls a $几百-dollar robotic arm holding a paintbrush to draw the Golden Gate Bridge on a white sheet of paper. The first attempt was uneven and wobbly, but by the fourth attempt, the outline and details of the Golden Gate Bridge were clearly visible.
After Thijs Simonian from OpenAI’s robotics team posted this real-world test video on X, the debate between China and the U.S. over whether general large models can serve as the brain for embodied intelligence has not ceased, with some even asserting that “OpenAI itself, not Physical Intelligence or Skild, will be the OpenAI of robotics.”
At a time when enthusiasm for investments in embodied intelligence in China’s primary market is temporarily cooling, discussing Astra provides a rare technical benchmark for narratives about embodied brains that were previously nearly impossible to verify.
Ranks can be manipulated, demos can be edited, and the scale of real-device data collection can serve as a valuation rationale—but anything that Astra can already do well may no longer command a premium. Players in the embodied AI industry are no longer competing solely within what appears to be an isolated sector.
01 Danger and Opportunity
Yushu Technology is one of China’s earliest leading embodied AI companies to publicly discuss Astra. Founder Wang Xingxing admitted that his biggest past mistake was “not believing in AI enough,” and predicted that when GPT-7 is released next year, it “might directly solve the embodied AI problem.”
Without any specialized training of the large model, Unitree Technology successfully enabled its robot to autonomously identify palettes, change paints, and clean brushes using GPT-6 Astra. More complex tasks, such as organizing a room according to instructions, may be achievable as early as next year.
After evaluating GPT-6 Astra, SenseTime also stated that general-purpose models can now handle robot manipulation and navigation tasks that were previously primarily managed by specialized models such as VLA and WAM, demonstrating capabilities such as zero-shot generalization, in-context learning, and failure recovery.
For leading hardware manufacturers like Unitree Robotics and AgiBot, this is a delicate moment filled with both risk and opportunity.
On one hand, original equipment manufacturers, whose past competitive advantages relied on hardware performance, motion control, and cost, need to find new sources of differentiation as the capabilities of the "brain" are expected to be rapidly leveled by general-purpose models.
On the other hand, as the supply of "brains" becomes diversified or even excessive, the "bodies" capable of hosting intelligence, collecting data, and delivering results may transform from previously low-barrier components into indispensable strategic assets.
Peng Zhihui, co-founder of Zhì Yuán, made an interesting statement, arguing that regardless of whether a general foundational model ultimately dominates embodied capabilities, intelligence must be grounded in a reliable physical body.
If the base model covers everything, the competition becomes who can embed intelligence into the real world at lower cost and higher consistency; if the base model cannot generalize, then domain-specific models still rely on data from their deployed ontologies—but ultimately, scalable, high-quality ontologies are becoming increasingly important.
At the "2026 Zhiyuan Innovation Day" on October 29, Peng Zhihui further stated, "Zhiyuan can also be an AI company." To support 30 million embodied tasks, the company's training cluster this year has approached a scale of ten thousand GPUs, placing its computing power consumption among the highest in the industry.
It is foreseeable that embodied intelligence is competing with large models for computing power, and Zhiyuan is striving to establish new legitimacy for its body-plus-full-stack approach. If the core elements of industry competition shift toward models and computing power, then leading embodied intelligence companies are already operating according to the logic of AI companies.
Moreover, Astra is not all-powerful; for example, it inevitably suffers from high latency in cloud-based inference, and its success rate for high-dynamic, fine-grained operations involving haptic feedback is virtually zero. Therefore, there is still a window of opportunity for leading hardware manufacturers.
However, for sci-tech companies in the embodied brain camp, the situation is much more awkward. The core narrative—that an independent embodied brain is a sustainable, defensible entrepreneurial direction—has been undermined.
In the first half of this year, companies whose core business was centered on the “brain” accounted for over 50% of domestic embodied intelligence funding, while companies focused on full-body hardware accounted for only 12.8% during the same period. But now, the era of widespread, unfocused funding has come to an end.
02 Question on Valuation
Recently, embodied AI company Yuanli Infinite completed an A++ round of financing totaling nearly RMB 100 million, led by Zhejiang Venture Capital Group. Just over two months ago, Yuanli Infinite announced its A and A+ rounds, raising nearly RMB 1 billion.
The high-profile funding rounds of only a few embodied AI companies do not reflect the current true level of interest in this industry.
According to NE Times, in September, there were 46 funding rounds in China's embodied intelligence primary market, with an approximate total amount of RMB 11.624 billion—down 4.2% and 40.1% respectively from August. Many of these deals had actually completed investment decisions in May or June but were only announced as closed in September.
Large-scale funding for leading embodied AI companies occurred mostly before May; for example, Xinghai Tu, Qianxun, Galaxy General, Zhi Square, and Zibianliang had already secured their primary funding during Q1 to early Q2, with the combined financing of the top five companies accounting for approximately 70% of the entire embodied AI sector.
After May, companies that still secured financing of over 100 million yuan were primarily those with clear state-backed support (such as Yuanli Infinite and Ju Nao Pan Shi) or those anchored by genuine orders (such as Zhongke Fifth Era).
Moreover, among the new funding raised after May, the amount and round levels of companies positioning themselves solely as a "general-purpose brain" have clearly declined.
The $200 million angel round by Wujie Dynamics, the hundreds of millions of yuan seed round by YuanCe Future, and the hundreds of millions of yuan Pre-A round by Moushen Intelligence may seem substantial on their own, but when compared to the $2 billion to $5 billion funding rounds secured by Xinghai Map and Qianxun in Q1, they likely belong to an entirely different valuation paradigm.
Previously, capital heavily bet on the embodied intelligence赛道, assuming it was a technological frontier requiring dedicated companies to conquer.
The CFO of StarSea Map cited "the scarcity of authentic brain targets" as one of three core reasons for the valuation doubling, exemplifying this narrative framework. Within this framework, brain-focused companies are positioned as platforms, gateways, and potential candidates for the "operating system" of the embodied intelligence era.
Capital has even given rise to another belief: that the data required to train embodied brains must be collected, annotated, and commercially delivered by specialized companies.
Astra has shattered the very notion of valuing a company as an "independent OpenAI." If general-purpose large model companies are inherently the strongest providers of embodied intelligence, then the standalone value of pure brain-focused startups must be re-evaluated—not by questioning whether their technology is good, but by asking why it must be you who does it.
More challenging is that large model companies have significant advantages over pure brain startups in terms of scenario data and customer relationships.
For example, Alibaba launched the Qwen-Robot suite of embodied intelligence large models in June—including VLN navigation models, VLA action models, and world models—aiming to build a comprehensive, general-purpose foundation for robotics. Alibaba’s existing offline fulfillment scenarios, such as Cainiao warehousing and instant delivery, serve as ideal natural testing grounds for these embodied intelligence models.
In contrast, pure brain-based startups typically receive only remote operation videos and lab workstation data provided by the hardware manufacturers. In the future, they are also likely to fall behind large model companies in terms of customer acquisition costs.
However, it should be noted that this does not mean embodied brain startups form a unified, equally vulnerable group.
03 Test the limit
Stress tests never distribute harm evenly; instead, they are more likely to reveal differences in resilience among various business structures when faced with the same shock.
The survival space of Embodied Brain depends on the business structure in which it is embedded—whether it is embedded in a fundraising narrative geared toward an IPO, in an established industrial cash flow, or in a closed-loop scenario operation.
The Core Brain Camp (StarSea Map, Qianxun Intelligence, Origin Infinity) is betting on the long-term future where the brain itself becomes a platform, offering the largest narrative space but the shallowest defense against the spillover capabilities of general-purpose large models. The future will depend on who first develops a commercial anchor that Astra cannot replace.
Taking Qianxun Intelligence as an example, it follows the integrated approach of "VLA + world model." Previously, its consistent leadership in benchmark tests allowed Spirit v1.6 to reach the top of RoboArena, surpassing NVIDIA and PI. However, with Astra’s impressive performance, the market’s focus will inevitably shift to questions about commercialization.
Prior to this, Qianxun Intelligence adopted a hybrid business model combining full-system sales, algorithm licensing, and capability subscriptions, introducing clear tiered subscription pricing (Standard Edition: 300,000 RMB/year, Enterprise Edition: 880,000 RMB/year, Flagship Edition: 3,000,000 RMB/year). This may be viewed as the starting point of a business model shift from “selling the brain” to “selling tokens.”
The scenario anchored by Force Unlimited is more tangible: its FORCE charging robot is priced at 368,000 RMB per unit. By the end of 2025, 300 units had been delivered, serving over 20 cities and more than 90 leading clients, with a cumulative total of 200,000 orders. This year, the company has set a revenue target exceeding 100 million RMB.
If it becomes a company rooted in vertical scenarios with an "embodied brain + specialized ontology," Origin Infinity’s floor at least remains secure.
In the brain-body fusion camp, Galaxy General and Cloud Deep represent two opposing risk structures.
Galaxy General integrates intelligence to orchestrate its operations, with its core assets being the Galaxy Brain and an end-to-end scenario operation loop. On the retail side, it operates smart pharmacies with over 100 locations across more than 30 cities, having processed over 1 million orders cumulatively. On the industrial side, it collaborates with CATL on production lines, with partnership models encompassing full-machine sales, Robotics-as-a-Service (RaaS), and data services.
Faced with pressure from Astra, Galaxia General’s differentiation for its self-developed Star Brain must increasingly rely on proprietary data loops generated by scenario-based operations. While becoming a scenario-based operation platform centered on embodied intelligence may lower the upper limit of valuation, it is more likely to ensure business sustainability.
The greater risk for Yunshen lies in the erosion of its core business; although 1.169 billion yuan from its IPO funding was allocated to embodied algorithms and models, only four humanoid robots have been sold in two years. In contrast, its investment in embodied intelligence appears more defensive than strategic.
This is similar to the situation faced by companies like Mech-Mind that focus on components: Astra will not replace industrial 3D vision guidance, nor will it eliminate the engineering demands for cycle time and yield on production lines, but the company’s desired “brain” business has never generated scalable revenue, which clearly undermines the valuation narrative built around its “eye-brain-hand” concept.
It is certain that the market space for embodied AI as standalone products is shrinking; the first to be pushed out will not be the one with the weakest technology, but the one lacking both scenario-driven cash flow and a verifiable data loop, relying solely on model narratives to justify its valuation.
Twitter: https://twitter.com/BitpushNewsCN
BitPush Telegram community: https://t.me/BitPushCommunity
BitPush TG subscription: https://t.me/bitpush