"Physical AI is Built on MIPS."
A 40-year-old Silicon Valley name that did the unthinkable: it scrapped its own chip architecture and rebuilt the whole company around the open standard it used to compete against.
// The logo of a company that has been owned by SGI, Imagination, a venture fund, a bankrupt AI startup, and - as of 2025 - GlobalFoundries. It still answers to the same four letters.
Walk into the right autonomous-vehicle lab today and the brain steering the car may carry a name older than most of the engineers in the room. MIPS - four letters that once meant workstations, then game consoles, then very nearly nothing - is back, and it is running on processor IP built for machines that sense, think, act, and communicate in real time.
The company no longer sells the architecture it invented. That is the first surprising thing about MIPS in 2026. The second is that it does not seem to mind. Headquartered on Zanker Road in San Jose with roughly 770 people, MIPS now licenses compute subsystems built on RISC-V, the open instruction set that the rest of the industry rallied around. The pitch is blunt: the future of artificial intelligence is not only in distant data centers. It is in cars, robots, and factory lines, where a few milliseconds of latency is the difference between a maneuver and a collision.
"Physical AI is built on MIPS."
MIPS calls this category Physical AI: intelligence that has to operate in the physical world, on a deadline, with safety on the line. It is a smaller, harder market than chatbots. It is also one where MIPS's old obsessions - low latency, deterministic timing, power efficiency, functional safety - suddenly read like a product roadmap instead of a museum plaque.
For decades, the deal in chip design was simple and uncomfortable. If you wanted high-performance processor cores, you licensed them from a small number of gatekeepers, accepted their architecture, and paid for the privilege of being locked in. Proprietary instruction sets were a fine business - for the people who owned them.
MIPS had spent most of its life on both sides of that bargain. It built one of the foundational RISC architectures, watched the industry consolidate around rivals, and then got passed from owner to owner as the value of a closed instruction set eroded. The problem was structural: you could own a perfectly good architecture and still lose, because customers increasingly wanted control, not captivity.
"Customers want to move beyond proprietary legacy architecture lock-ins."
Then came AI at the edge - sensors, cameras, radar, motors - each generating torrents of data that had to be moved and acted on instantly. The old answer, ship it to the cloud and wait, does not work when a robot arm is about to swing or a vehicle is about to brake. Someone had to build the local nervous system. That someone needed an architecture customers could trust and customize. Closed wouldn't cut it.
MIPS was born in 1984 out of Stanford, commercializing the RISC research of John L. Hennessy - who would later become president of Stanford and win the Turing Award - alongside co-founders including John Moussouris, Skip Stritter, and Chris Rowen. For a while it was the toast of the industry: design wins at DEC, an acquisition by Silicon Graphics, a starring role inside the Nintendo 64 and the original PlayStation.
Then the long descent. Acquired, spun out, sold, and sold again. By 2020 its then-owner, Wave Computing, filed for bankruptcy. Most stories end there. This one made a U-turn.
In 2021 the company emerged from bankruptcy, kept the storied name, and announced it was abandoning the MIPS architecture in favor of RISC-V. A chip-IP company walking away from its own instruction set is roughly as on-brand as a winery announcing it has decided to sell water.
The bet got a face in September 2023, when MIPS named Sameer Wasson - an 18-year Texas Instruments veteran who had rebuilt TI's processor business around automotive, industrial, and embedded AI - as CEO. His read was that the prize wasn't the architecture; it was the roadmap, the subsystems, and the markets where real-time AI actually has to work.
"Give back control, for AI-centric compute."
Spun out of Stanford to commercialize the MIPS RISC architecture.
The first MIPS microprocessor reaches the market.
Silicon Graphics buys MIPS; chips later power the Nintendo 64 and PlayStation.
The PowerVR graphics company acquires MIPS.
Sold to Tallwood Venture Capital, then to AI startup Wave Computing.
Emerges from bankruptcy and abandons its own architecture for open RISC-V.
Former TI processor chief takes the wheel to drive RISC-V market penetration.
Billed as the first high-performance AI-enabled RISC-V automotive CPU.
GF announces it will acquire MIPS; it will run as a standalone business inside GF.
MIPS organizes its current portfolio - branded Atlas - around the loop every autonomous machine runs: sense, think, act, communicate. Instead of selling a single core, it sells the building blocks of a real-time compute subsystem that chipmakers and OEMs drop into their own silicon. The flagship pieces are unapologetically technical.
The industry's first high-performance, AI-enabled RISC-V automotive CPU, aimed at the low-latency, data-hungry demands of ADAS and autonomous vehicles.
Think / ComputeA power-efficient in-order multiprocessor for SoCs, fusing multi-threading with a triple-issue pipeline - scalable to 64 clusters, 512 cores, 2,048 threads.
ScaleData-movement engines for low-latency, high-reliability processing with functional safety, coupling tightly with edge AI engines.
Sense / DataA virtual-platform tool to evaluate and optimize IP, subsystems, and SoCs - testing the chip before the chip physically exists.
Tools"Each I8500 core fuses multi-threading with a highly efficient triple-issue pipeline."
The business model is the same one MIPS has lived by for forty years: it does not run factories. It designs and licenses the intellectual property - the cores, subsystems, and tooling - and earns license fees and royalties when its customers ship silicon. What changed is the foundation underneath. The blueprints are now RISC-V, which means customers can extend and customize them with their own instructions rather than waiting on a vendor's permission.
A reboot is a story until someone signs a check. In 2025 several did. In June, MIPS and Cyient Semiconductors announced a collaboration on custom RISC-V-based intelligent power solutions for AI power delivery, industrial robotics, and automotive, built on the Atlas portfolio. MIPS had also tapped Siemens' Veloce proFPGA platform to bring up the P8700.
Bars scaled relative to the 2,048-thread maximum. Translation: more parallelism than most engineers will ever schedule - which is exactly the headroom autonomous systems are starting to demand.
Then the big one. In July 2025, GlobalFoundries - one of the world's major chip manufacturers - announced a definitive agreement to acquire MIPS, explicitly to accelerate its AI and compute capabilities. The plan is for MIPS to keep operating as a standalone business inside GF, pairing its RISC-V processor IP with GF's manufacturing reach. After a decade of being passed between owners who didn't quite know what to do with it, MIPS was bought by one that did.
"GlobalFoundries to acquire MIPS to accelerate AI and compute capabilities."
The most ambitious thing MIPS wants is, in a sense, the most modest: to make adopting Physical AI simple. Not flashy. Reliable. The company's stated aim is to deliver open, modular, safety-capable compute subsystems so that engineers building robots, vehicles, and industrial systems can assemble real-time intelligence the way you'd assemble anything else - from trusted parts, on an open standard, without betting the design on one vendor's goodwill.
"Sense. Think. Act. Communicate."
It is a deliberately unglamorous mission for a company that once lived in living rooms via game consoles. But unglamorous is the point. Safety-critical compute is judged on what doesn't happen - the crash avoided, the deadline met, the worst case that stayed hypothetical.
If AI keeps escaping the data center and moving into the physical world - and the momentum behind robotics, autonomous vehicles, and industrial automation suggests it will - then someone has to supply the local intelligence those machines run on. That intelligence has to be fast, safe, power-thrifty, and open enough to customize. MIPS has spent forty years accidentally training for exactly this exam.
Backed by GlobalFoundries' factories and anchored to the open RISC-V standard, the company is no longer fighting to own an architecture. It is fighting to be the default subsystem inside the next generation of thinking machines. Whether it wins is unknown. That it gets to compete again, after the obituaries were half-written, is the part nobody saw coming.
"Beyond proprietary lock-in: an open standard, a 40-year roadmap, and a second act."
Return, for a moment, to that autonomous-vehicle lab. The car edges forward, reads the road, and brakes a half-second before a human would have noticed the reason. The name on the silicon doing the thinking is older than the lab, older than RISC-V, older than the idea of Physical AI itself. It just learned a new architecture - and quietly went back to work.