On September 28, AMD announced that it had reached a definitive agreement to acquire World Labs, founded by Fei-Fei Li, in an all-stock transaction valued at approximately USD 8.2 billion. The transaction is subject to regulatory approval and is expected to close by the end of this year. World Labs focuses on spatial intelligence and world models, among other areas. Upon completion of the transaction, Fei-Fei Li will join AMD as Executive Vice President and Chief Scientist.
A semiconductor company renowned for its CPUs and GPUs is now integrating model research teams into its technological landscape.
Meanwhile, GlobalFoundries is also expanding beyond wafer manufacturing. In August 2025, it completed the acquisition of MIPS; in June this year, it acquired the ARC Processor IP business from Synopsys. RISC-V processor IP, software tools, custom design, and manufacturing are now integrated into a unified "software-to-silicon" ecosystem.
AMD is moving from chips to models, while GlobalFoundries is shifting from manufacturing to IP and software. Though heading in different directions, both point to a common shift: AI is moving from the digital world into the physical world.
Cars need to navigate roads autonomously, robots need to reach out and grab objects, and industrial equipment must adjust its actions in real time based on the environment. Once AI truly starts to "take action," the previously clear-cut division of labor in the semiconductor industry will hardly be able to continue in its original form.
A Single Large Chip Cannot Accommodate Physical AI
To understand this shift, one must first look at the new challenges Physical AI poses to chips.
Sameer Wasson, CEO of MIPS, describes Physical AI as a closed loop of "perception, reasoning, execution, and communication": sensors perceive the real world, processors handle understanding and reasoning, and control systems translate the results into actions for motors, steering wheels, or mechanical joints, with continuous communication among all nodes.
The most common misconception is equating AI upgrades with simply swapping in a larger GPU or NPU.
Wasson's judgment is exactly the opposite. Large centralized processors will continue to exist, but a massive amount of low-power intelligence will also be distributed across sensors, MCUs, actuators, and communication nodes, with centralized computing and distributed intelligence coexisting for the long term.
The automotive industry has already previewed this evolution.
Over the past few years, electronic and electrical architectures have continuously shifted from distributed ECUs to domain control, central computing, and zonal control. However, radars, cameras, high-speed communication, power management, and safety MCUs have not disappeared as a result.
Faisal Saleem, Senior Vice President of GlobalFoundries, revealed that over the past five years, the number of vehicles covered by GlobalFoundries chips has remained at approximately 9 million units, with little change, but the number of chips installed per vehicle continues to increase.
Robots embody this layering even more thoroughly.
A large model can understand "bring the cup over," but the motor closed-loop control, force control, balance, and collision detection for dozens of joints cannot wait for the large model to think slowly. One handles high-level tasks, and the other handles millisecond-level actions; they are inherently different workloads.
This is also why MIPS repeatedly emphasizes being "workload-native." Facing Physical AI, it has not tried to use a single processor to handle all tasks. Instead, it breaks down computing requirements into categories such as AI inference, real-time response, and safety-critical computing. On the other hand, the race for massive computing power has not stopped—its NPX6 AI accelerator boasts a nominal computing power of 3,500 TOPS.
Looking at these two things together is actually more interesting.
The computing power race is not over; it is just that in the era of Physical AI, computing power is increasingly becoming an entry ticket rather than the answer itself.
The changes have even reached power supply. When answering a question from Gasgoo about the 48V architecture, Faisal judged that as the power consumption of central CPUs, ADAS, and zonal controllers increases, the evolution of automobiles from 12V to 48V is "inevitable and imperative," although the transition will still take time.
When AI truly begins to control the real world, the greater challenge than building a more powerful chip is getting the entire system to run.
Systems Are Becoming Increasingly Complex, and Semiconductor Companies Are Expanding Outward
As system problems multiply, suppliers naturally refuse to settle for just selling a small piece of the pie.
At a recent MIPS media briefing, Sameer Wasson explicitly stated that the company aims to transition from an IP supplier to a full-stack technology solution provider, making custom chips a crucial part of its future business.
His subsequent remark is even more noteworthy: IP licensing remains the core business of MIPS, but a significant portion of future new markets and new revenues will come from chip manufacturing and post-tape-out customer delivery.
The business logic is not complicated.
Selling an IP earns money from just one segment of the industry chain; continuing into software, custom design, and even final chip delivery allows entry into the customer's longer development process.
According to MIPS itself, it is currently the world's largest RISC-V IP supplier and the world's second-largest processor IP supplier, second only to Arm. The annual shipment volume of chips based on its related technologies is approximately 3.5 billion units, of which over 200 million chips using the same microarchitecture have entered ADAS. It has more than 500 customers, including 9 of the top 12 global automakers and the top 5 global robotics and industrial automation manufacturers.
Therefore, what MIPS wants to change now is not just its product catalog, but how deeply it participates in a system.
GlobalFoundries' acquisition of MIPS and ARC follows the same logic. Originally closer to the manufacturing end, it has now incorporated processor IP, software tools, and custom chip capabilities into its system; AMD, from the other end, is pulling world model and spatial intelligence research into the semiconductor company.
It is not that semiconductor companies suddenly all want to go "full-stack," but rather that customer problems no longer stop at a single chip.
Technological complexity pushes companies upstream and downstream, while new revenue spaces make this endeavor even more worthwhile.
Automotive and Robotics Begin to "Borrow" Technologies from Each Other
While companies extend vertically, technologies are also flowing horizontally.
On September 8, XPeng Robotics' automated production line was officially launched, with IRON rolling off the line autonomously after automated final assembly. According to official information from XPeng, IRON has 76 degrees of freedom across its body, is equipped with 3 Turing AI chips, and delivers an effective computing power of 2,250 TOPS. The quality system and manufacturing capabilities accumulated from the automotive business have also been brought into the robotics production line.
This is a typical case of internal reuse. Automotive, AI chips, models, and robotics are all within one corporate system, and the technologies and manufacturing foundations already invested in automotive can continue to be amortized over robotics.
Independent robotics companies, however, are taking a different path.
UBTECH currently still relies on mature computing ecosystems, but in June this year, it co-founded XILECT with companies like MetaX, with a registered capital of CNY 100 million, to start laying out dedicated edge-side chips for embodied intelligence, planning to tape out in the second half of 2027 and mass-produce in 2028.
First using mature platforms to build products, and then participating in chip definition in reverse once workloads, costs, and scale become clear, is also a viable path.
On the other hand, automotive chip companies are also moving into robotics.
SemiDrive has formed a robotics chip portfolio comprising the R1 "brain," D9 "intelligent control cerebellum," and E3-R execution control MCU. Its official website shows that the D9-Max has achieved mass production at leading humanoid robot manufacturers.
Zhang Xitong, Vice President of SemiDrive, stated at the CIFTIS in September that, in SemiDrive's view, for automotive-grade chip companies entering the robotics field, "70% to 80% of capabilities can be reused." However, changes in its own products also show that reuse is not merely relocation: when it comes to joints, dexterous hands, and motion control, redesigning around new workloads is still necessary.
Safety capabilities are also migrating. In June this year, NVIDIA launched Halos for Robotics, extending the Halos safety system previously accumulated in the autonomous driving field to humanoid and industrial robots. Agility Robotics became the first company to partner with NVIDIA and incorporate some of these capabilities into its own robot safety system.
XPeng brings chips and manufacturing from automotive into robotics; UBTECH moves upstream from robotics to participate in chip definition; SemiDrive repurposes automotive-grade capabilities into robotics products; and NVIDIA continues to extend its automotive safety system into robotics.
Technologies can be reused across industries, but complete machines cannot be simply copied.
Underlying capabilities are increasingly becoming like a set of public building blocks, but every time they reach a new terminal, they must be reassembled.
Full-Stack Has Boundaries, and Openness Also Has Costs
If the story stopped here, it would be easy to draw a conclusion: in the future, all companies will become increasingly "full-stack."
Reality is not that simple.
Suppliers must first face the boundaries of their own capabilities.
In 2018, GlobalFoundries suspended its 7nm FinFET development, further shifting resources to differentiated processes such as FDX, RF, and analog mixed-signal. Today's acquisition of MIPS and ARC is an extension along its own strengths into IP, software, and custom design, rather than a return to the competition across all advanced logic nodes.
Customers will not naturally accept an increasingly closed platform either.
In January this year, the automotive open-source collaboration promoted by VDA and Eclipse expanded to 32 companies, hoping to establish an open, interoperable software foundation; in May, Automotive Grade Linux launched the SoDeV reference platform, one of the core goals of which is to free software development from binding to specific hardware platforms as much as possible.
This creates direct tension with the full-stack expansion of semiconductor companies.
Suppliers hope to integrate IP, chips, runtimes, toolchains, and even models more tightly, which can reduce adaptation, shorten development cycles, and generate more revenue.
Terminal enterprises, however, see the other side: if chips, compilers, middleware, and safety systems are all bundled together, the next time they switch suppliers, it might not just be "changing a chip," but redoing the software and validation from scratch.
A complete solution is about efficiency, while a closed ecosystem comes with a cost.
RISC-V presents this set of contradictions even more directly.
Yankin Tanurhan, CTO of MIPS, warned that if a large number of proprietary extensions are made in advance before a unified standard is formed, companies may have to rework and readapt once the official standard is finalized.
However, RISC-V's entry into automotive is no longer just at the discussion stage. As a leading automotive MCU supplier, Infineon has announced that it will launch a RISC-V product family within the AURIX system; Quintauris, jointly established by Bosch, Infineon, Nordic, NXP, Qualcomm, and STMicroelectronics, is also promoting a RISC-V reference architecture for automotive-grade real-time computing, and MIPS has already started cooperating with it.
China's automotive industry is also beginning to make up for the lesson in standardization. It is understood that the first industry standard for RISC-V in the automotive field has started to be drafted, with dozens of companies participating.
The past question was whether RISC-V "could get into cars"; now, the more realistic question has become: can different companies get into cars according to the same set of rules.
Openness does not mean a lack of standards. On the contrary, the more open the ecosystem, the more it needs a set of standards that everyone is willing to follow.
Once robots truly enter large-scale mass production, they will also have to face this hurdle.
Using mature platforms today can build machines the fastest; but when shipment volumes rise tomorrow, power consumption, BOM costs, real-time control, and algorithm workloads will force companies to recalculate: what to continue buying, what is worth defining jointly, and what must be kept in their own hands.
Physical AI has not invalidated the traditional industrial division of labor, but it is forcing every company to find its own position anew.
Some are moving from manufacturing to IP, some from chips to models, some bringing automotive technologies into robotics, and others participating in chip definition in reverse from robotics. Meanwhile, standards, open source, and multi-vendor systems are continuously drawing new boundaries for this expansion.
The more complete the platform, the more customers need to clearly ask about the exit mechanism.
Physical AI is redrawing the boundaries of the semiconductor industry.
What is truly worth watching in the next stage is not necessarily who can complete the entire industry chain.
But rather—who has the ability to decide where the division of labor in this industry chain will take place next time.
--END--