Unitree held a major conference this week to double down on Physical AI, but the larger the market opportunity, the harder it becomes to maintain its leading position. MiniMax has favorable conditions and is not lacking in compute power, yet it has fallen behind in foundation model competition. The Doubao smartphone faced rejection from 166 apps; even if AI can operate phones, it must still pass the gate of app developers.
Viewing these three events together, the advantages previously held by AI companies may not necessarily translate into leadership in the next round of competition.
01、The Larger the Physical AI Market, the Less Stable Unitree's Leading Position
This week, Unitree announced at its Physical AI Conference a plan combining self-developed models with external collaborations, focusing on advancing physical AI models adapted to its own hardware. The threat from foundation model vendors has become apparent, forcing robot manufacturers to bolster their model capabilities. However, the larger the potential space for the Physical AI market, the smaller Unitree's chance of remaining the industry leader.
Previously, Unitree's advantage lay mainly in robot bodies and motion control—often referred to as the "cerebellum"—enabling robots to stand steadily, run fast, and be produced at lower costs. Physical AI also requires a "brain" to understand the environment, judge tasks, and then decide how to act.
Mastering the cerebellum does not imply possessing the same advantage in building the brain. If Physical AI ultimately becomes a super-large market, foundation model vendors will inevitably invest heavily. At that point, Unitree would face competitors with stronger general-purpose models, larger compute investments, and top-tier research teams.
In the future, capability gaps among robots will be primarily determined by the "brain," giving foundation model vendors a greater opportunity to dominate this market. Unitree's leading position accumulated through motion control will be challenged. Conversely, if Physical AI evolves more like intelligent driving, forming commercial value mainly around specific hardware and limited scenarios, Unitree actually has a better chance.
Models must be adapted to robots, and customers must consider cost, reliability, and delivery. Foundation model vendors may not be willing to sustain heavy investment in these niche scenarios. Unitree, however, can rely on its existing hardware and mass-production experience to develop some models internally while introducing external capabilities to optimize complete products. For Unitree, a market that is profitable enough but not large enough to attract full-scale investment from foundation model giants may be more advantageous than a super-market leading toward AGI (Artificial General Intelligence).
02、MiniMax Drops Below IPO Price: Why Was It the First to Fall Behind?
This Friday, MiniMax's stock price briefly fell to HKD 155.2, dropping below its IPO offering price of HKD 165. MiniMax's conditions are not poor; its compute reserves lead most peers, meaning it does not need to worry about GPU shortages first.
In February of this year, Anthropic alleged that MiniMax distilled programming and tool-calling capabilities through over13 million interactions with Claude. However, by June, M3 had not widened the gap with competitors. In the subsequent months, no model emerged that was particularly impressive.
The M3.1 Flash update released this week showed some progress, but comprehensive evaluations still did not surpass DeepSeek's and Zhipu's latest models. With ample card resources, the failure to continuously keep up with models suggests the problem likely lies within the team. In the first half of this year, Liang Li, the former head of Agent business, left to start her own venture. Cai Jiaren, who participated in post-training and was responsible for open-source work, also joined a new company. By August, reports indicated that A Dao, the supervisor of the Agent Engineering Department, had also resigned. Personnel departures occurred in key areas such as evaluation, post-training, and engineering. Who will take over these roles, and whether original R&D coordination can continue, will impact the next model.
Consider Zhipu. Upon listing in January this year, Tang Jie announced in an internal letter a full return to foundational model research. Since then, the GLM 5 series models have performed well, approaching overseas top-tier closed-source models in long-horizon programming tasks. For MiniMax, the focus now shifts to Yan Junjie: can he lead the team to produce a model that regains recognition?
03、Doubao Asks Apps to State Their Positions, Receives 166 Rejections
Doubao wants AI to operate in-phone apps on behalf of users. It gave app developers 30 days to state their positions, allowing those unwilling to register a rejection. As of October 21, the rejection list already included 166 apps, with all financial applications restricted.
Why did Doubao proactively solicit public opinions, concentrating these "refusals"? This reflects somewhat the issues common in large corporations. The product team wanted to push the feature forward, while the compliance team needed to manage risks first. Eventually, they compromised: publish the notice first, allow apps to opt out, and gradually open up access. However, ByteDance can set rules for its own products, but other companies' apps are under no obligation to accept them.
Writing "gradual opening unless opposed" into the rules does not make cooperation partners see profit. As a result, Doubao collected a long list of rejections and publicly named those who refused. Cooperation that originally required case-by-case negotiation now begins with a public record of "non-cooperation."
When Meta advanced Muse, it also sought support from service providers. It introduced retail partners like Walmart and Best Buy for Muse and integrated payment services, enabling assistants to complete purchases through collaboration. At least, these service providers participated as partners.
Doubao also needs to answer what benefits apps gain after integration, how users and transactions are distributed, and who is liable when problems arise. If these matters cannot be agreed upon, even perfect rules allowing others to opt out will hardly secure genuine support.
There is a vast difference between "no rejection" and "willingness to cooperate."
If mainstream apps continue to reject, Doubao's ability to perform cross-app tasks will shrink significantly. When users ask AI to arrange travel, after checking information, booking tickets and hailing rides still require manual action, making the promise of "getting things done with one sentence" difficult to fulfill. If only chat, search, and partial system operations remain, why should users specifically switch to a new phone?