On September 28, AMD announced a definitive agreement to acquire World Labs, the spatial intelligence startup founded by Fei-Fei Li, in an all-stock deal valued at about $8.2 billion. The transaction, still subject to regulatory approval, is expected to close by the end of this year. World Labs focuses on spatial intelligence and world models; after the deal closes, Li is set to join AMD as executive vice president and chief scientist. A chip company known for CPUs and GPUs is now folding a model-research team into its technology portfolio, according to Chinese outlet D1EV.
On the other side of the industry, GlobalFoundries is also extending beyond wafer manufacturing. In August 2025 it completed its acquisition of MIPS, and this June it took over Synopsys' ARC Processor IP business. RISC-V processor IP, software tools, custom design and manufacturing are being bundled into a single "software-to-silicon" framework. AMD is moving from chips to models; GlobalFoundries is moving from manufacturing to IP and software. Different directions, but they point to the same shift: AI is stepping out of the digital world and into the physical one.
One big chip can't contain Physical AI
MIPS CEO Sameer Wasson describes Physical AI as a closed loop of "perception, reasoning, execution and communication": sensors see the real world, processors understand and reason, control systems turn results into motor, steering or robotic-joint actions, and nodes keep communicating. The common mistake is to think AI upgrades mean simply swapping in a bigger GPU or NPU. Wasson argues the opposite: large centralized processors will remain, but lots of low-power intelligence will also spread across sensors, MCUs, actuators and communication nodes. Centralized computing and distributed intelligence will coexist for a long time.
Over the past few years, vehicle electrical and electronic architectures have shifted from distributed ECUs toward domain controllers, central computing and zonal control - but radar, cameras, high-speed communication, power management and safety MCUs have not disappeared. Faisal Saleem, global senior vice president at GlobalFoundries, said the number of cars covered by its chips has stayed around 9 million over the past five years, but the number of chips per vehicle keeps rising. A large model can understand "bring the cup over," but the motor loops, force control, balance and collision detection across dozens of joints cannot wait for a large model to think slowly. One handles high-level tasks, the other handles millisecond-level actions - they are not the same workload.
That is why MIPS stresses "workload-native" design. Facing Physical AI, it does not try to have one processor do everything; it splits computing needs into AI inference, real-time response and safety-critical computing. Meanwhile the raw-compute race has not stopped - its NPX6 AI accelerator claims 3,500 TOPS. Compute is increasingly a ticket to entry rather than the answer itself. The shift even reaches power delivery. Asked about 48V architectures, Saleem said that as central CPUs, ADAS and zonal controllers draw more power, moving cars from 12V to 48V is "inevitable," though the transition will take time.
Chipmakers chase the whole system
As AI starts controlling the real world, making the whole system run is harder than making a stronger chip. Suppliers are no longer content to sell just one small piece. At a recent MIPS media briefing, Wasson said the company wants to move from an IP supplier to a full-stack technology solution provider, with custom chips a key part of future business. IP licensing remains MIPS's core, but much of the new market and new revenue will come from chip manufacturing and post-tapeout customer delivery, he said. Selling an IP captures one slice of the chain; continuing into software, custom design and final chip delivery gets a supplier deeper into the customer's development process.
By its own count, MIPS is the world's largest RISC-V IP supplier and the second-largest processor IP supplier after Arm. Chips using its technology ship at about 3.5 billion units a year, with over 200 million using the same microarchitecture already in ADAS. It has more than 500 customers, including 9 of the world's top 12 automakers and all of the top 5 robotics and industrial automation firms. GlobalFoundries' acquisitions of MIPS and ARC follow the same logic: it stays close to manufacturing but now pulls processor IP, software tools and custom-chip capability into its system. AMD pulls world models and spatial-intelligence research into a chip company from the other end. It is not that chip firms suddenly all want to be "full-stack" - it is that customer problems no longer stop at one chip.
The same logic is playing out in robotics. On September 8, Xpeng's automated robot production line went live, with its IRON humanoid completing automated final assembly and walking off the line on its own. According to Xpeng, IRON has 76 degrees of freedom and three Turing AI chips with 2,250 TOPS of effective compute. The quality system and manufacturing capability built for cars are being carried into the robot line - a typical internal reuse play. UBTech, for now, relies on mature computing ecosystems, but this June it co-founded Xixuan Chuangzhi with companies including Moore Threads, registered at 100 million yuan (about $14.1 million), to develop dedicated chips for embodied-intelligence edge computing, with tapeout planned for the second half of 2027 and production in 2028. Semidrive has formed a robot chip lineup with its R1 "brain," D9 "smart control cerebellum" and E3-R execution control MCU; its D9-Max is already in mass production at a leading humanoid robot maker, according to its website. Semidrive vice president Zhang Xitong said at the September CIFTIS fair that 70-80 percent of automotive-grade chip capabilities can be reused in robotics - but reuse is not a simple move, and joints, dexterous hands and motion control still need redesign around new workloads.
Safety capability is migrating too. In June, Nvidia launched Halos for Robotics, extending the Halos safety framework it built for autonomous driving to humanoid and industrial robots. Agility Robotics is the first company to work with Nvidia to fold parts of that capability into its own robot safety system. Xpeng brings chips and manufacturing from cars into robots; UBTech moves from robots upstream to help define chips; Semidrive reworks automotive-grade capability into robot products; Nvidia extends its automotive safety system to robots. Underlying capabilities increasingly look like shared building blocks, but every new terminal requires reassembly.
The story is not simply that everyone becomes more "full." In 2018, GlobalFoundries halted 7nm FinFET development, redirecting resources to differentiated processes like FDX, RF and analog mixed-signal. Buying MIPS and ARC today extends its strengths into IP, software and custom design rather than rejoining the race in every advanced logic node. And customers will not naturally accept an ever more closed platform. In January, the VDA- and Eclipse-backed automotive open-source collaboration expanded to 32 companies aiming for open, interoperable software foundations; in May, Automotive Grade Linux launched its SoDeV reference platform, with a core goal of freeing software development from binding to specific hardware platforms. Suppliers want to tie IP, chips, runtime, toolchains and even models closer together - it reduces adaptation work, shortens development cycles and earns more revenue.
