Seventeen Days of a Chip: The Journey of a Group of ZEKU Veterans

Stock News10-30

At 15:00 on October 29 at the Suzhou International Conference Hotel, the temperature was 19 degrees and the air was slightly cool, as the X7 Ultra large-model chip launch event for XHEART was underway.

On stage, CEO Li Zonglin — wearing denim jeans and a black T-shirt marked with the "XHEART" logo — slowed his voice when he reached a particular detail.

"That colleague from the XHEART team was asked by HR before departure when he would return. He said: 'I don't know either.' So he bought a one-way ticket first." The audience fell silent for a moment, then applause broke out. Among those present were Turing Award winner Jack Dongarra, TSMC Vice General Manager Chen Ping, SAIC Chief Engineer Zu Sijie and MOMENTA-W (06880) CEO Cao Xudong. That ticket with no return date was the first footnote to a choice made by a group of chip veterans in the winter of 2023.

Part One: Those Who Dispersed

The story goes back three years. When OPPO's ZEKU was disbanded in an 18-minute all-hands meeting, Li Zonglin stood at the front and left behind the line that would be quoted repeatedly afterward: "The past three years were the most colorful experience of my life." He was ZEKU's COO and, earlier, general manager of MediaTek's wireless communications business unit, having led the Helio processor and Dimensity chip series — one of the strongest teams in the mobile chip industry.

ZEKU had four years, nearly 3,000 people, and a 4nm flagship SoC that had just taped out, when a single order sent everyone home. But the veterans did not exit the stage. In the summer of 2023, Li Zonglin led more than a dozen core members including software department director Jia Mingjun and senior SoC director Yu Guojun, together with nearly a hundred engineers, in transferring as a whole into the MOMENTA-W (06880) system; in December of the same year, the chip project became independent and XHEART was registered and established in Suzhou. First came the team, then came the company. That order determined the unusual speed it would later achieve.

"Starting over actually made us more cautious. What exactly do we want to do this time? A few years from now, will the chip we make still be what the market needs?" Li Zonglin said. Fate came quickly. They met MOMENTA-W (06880) CEO Cao Xudong. "We talked about technology, about the market, and about what will change in intelligent driving over the next few years." After several long conversations, two judgments gradually became clear. First, large models will change intelligent driving, and will also change the way chips are designed. In the past, intelligent driving systems were pieced together from multiple small models — one for perception, one for prediction, another for planning — and a great deal of information was lost in transmission. "The era of small models was essentially a human-brain-driven system," Li Zonglin said. End-to-end large models allow massive data to be truly utilized efficiently, but what about the chip? "Can computing power be brought into play in a real system? Can a chip designed today keep up with tomorrow's models and achieve the required capability with reasonable energy consumption?" All three questions had to be answered together. The answer was six words: large models define chips. It is not that models accommodate hardware, but rather that hardware works backward — the model tells you where computation time is spent, and the chip buries transistors there. Second, automobiles are the starting point where this transformation will benefit more users first. The idea was sexy, but execution was lean — a capital winter, a small team, and the long-term investment required by chips meant every trade-off was a matter of life and death. They still took this road few travel, with only one guiding thought: understand large models first, then design chips.

Part Two: Seventeen Days

Mobile phone chips and automotive-grade chips are two opposite sets of muscle memory. Mobile phones compete on generality and shipment volume, with hundreds of millions of units a year spreading costs thin, and iteration as fast as fashion; automotive-grade is the reverse — a single model sells only a hundred thousand or so units a year, one chip must endure ten years, verification is measured in years, and failure rates must approach zero. What MediaTek and ZEKU left these people was not any single technology, but a timetable: which step will stall for how long, who it will stall with, and which step must leave buffer room. Advanced processes from 6nm to 4nm, low power consumption, and software-hardware co-design were all the same skill set. The X7 Ultra therefore grew into an "unconventional" form: a super-large-core NPU with terabyte-level internal interconnect; hardware accelerators specially designed for Linear, Softmax and Transpose; and zero-copy across the entire network. Measured NPU utilization reached 98.7%, bandwidth consumption dropped 80%, and energy efficiency doubled.

The result was astonishing. "In less than 24 hours, we successfully lit up the chip," Li Zonglin said. But lighting up was only the first step. "What we need to deliver is a chip that can truly enable stable and reliable operation on real vehicles." The first test vehicle hit the road within days, but the team decided to redefine "getting on the vehicle" using stricter mass-production standards — the complete system composed of the chip, underlying software and autonomous driving model had to perform on real roads at a level close to mass production. "Down to the turn at every intersection, the driving trajectory planned by the model had to remain consistent with mass-production standards." This standard was extremely harsh for a newly born chip. But the team worked day and night. Eventually the test vehicle drove continuously and stably for several hours on urban roads with no takeover throughout — daytime, nighttime and heavy rain, and the system withstood every test. From the chip's birth to a closed loop on the vehicle, it took only 17 days. On day 21, the test vehicle entered real traffic and opened up the experience to customers.

"If someone else were telling this story, I would be skeptical." The speaker was TSMC Vice General Manager Chen Ping. He recalled the scene three years earlier when he first heard Li Zonglin and Cao Xudong describe their vision. "He talked about large-model chip architecture, and Xudong talked about L3 and L4. At the time I thought, this sounds like something out of a science fiction film." But today, Chen Ping walked into the venue and, at the entrance of the temporarily built XHEART hall, saw those models already in mass production. "In three years, such an advanced chip has already landed on so many vehicle models. If you had told me three years ago, I would not have believed it no matter what." He paused, then smiled and said, "So I must congratulate our partners."

The data is the best footnote. Eight months after mass production, cumulative shipments exceeded 200,000 units, with 10 mass-produced models carrying it and orders from more than 40 models. MOMENTA-W (06880) cumulative mass production surpassed 1.4 million units, with mileage reaching 20 billion kilometers — equivalent to 25,000 round trips to the moon and back for Apollo. This is the first inflection point in the intelligent driving industry where "the data flywheel spins faster than chip iteration": based on the X7 Ultra, the MOMENTA-W (06880) world model R7 Ultra expanded its parameter scale fivefold and improved safety fivefold, moving from "understanding the world" to "simulating all things."

At the launch event, a media test video was played: a self-driving car reversed in a narrow section to yield, patiently waited after perceiving the movement of an oncoming vehicle, started instantly after an e-bike passed, and turned the steering wheel while moving, all smooth throughout. "From multi-point U-turns and negotiation when meeting oncoming cars to reversing to yield and extreme parking — our feeling is that it is top-tier, even able to surpass a seasoned human driver," Cao Xudong said. SAIC's Zu Sijie put it bluntly: "Consumers are not buying a computing power number, they are buying the actual experience." Even the Roewe priced at 119,800 yuan can handle a six-way intersection and 20-centimeter extreme parking.

Part Three: Science Fiction Enters Reality

After Li Zonglin, Cao Xudong also took the stage to share his own feelings. He quoted a line from Kevin Kelly's "The Future of AI": "The brain uses a large amount of memory to build a world model... The brain uses the model of memory to continuously predict future events." "The ability to predict the future is the key to intelligence." And large-model chips are now taking this predictive ability from the digital world into the physical world.

Near the end of the roundtable, the host asked: "In the era of large models, what kind of chip do we really need?" Turing Award winner Jack Dongarra answered: "Not the one with the highest peak performance, but the one that delivers the correct answer on time within constraints of power consumption, cost and latency." Zu Sijie added four words — "all-around warrior." Cao Xudong added two more dimensions: "Bigger, safer." XHEART revealed its hand on the spot: the next-generation X9, targeting 3000 TOPS and single-chip bandwidth of 700GB/s, to be launched in 2027. Cao Xudong nailed the timetable on the blackboard: from 2022 to 2026, intelligent driving capability improves tenfold every two years; after large models define chips, it accelerates to tenfold every year; 2028 is the L3/L4 tipping point; 2030 brings safety 100 times that of humans.

As the launch event ended, Li Zonglin did not talk about valuation. He only said: "This is not just a breakthrough in one chip, it is a look at what the intelligent driving experience can be." Looking back, what that one-way ticket bought three years ago was not the success of some product, but a revolution in paradigm. In the past, chip companies competed on computing power and algorithm companies competed on models, each running its own track. XHEART and MOMENTA-W (06880) welded them into one: models pull chips forward, chips support models, mass-produced vehicles return data, and data trains models again. A few days earlier, XHEART had promoted nearly 5 billion yuan in financing, with Changxin making an exceptional investment and Renault placing its first bet, while the shareholder list was crowded with automakers, memory leaders and state-backed funds. What they bought was not a piece of silicon, but a team that had run through the entire process of a flagship large chip. On the day ZEKU disbanded, Li Zonglin said the "most colorful period" had ended. No one expected that the same group of people would spend more than two years writing the engineering and mass-production discipline accumulated through failure back into an automotive-grade large-model chip. Scattered, they are stars across the sky; gathered, they are a fire. When a one-way flight lands, and when a chip begins to be born for "predicting the world," these veterans have also written themselves back into the opening story of the endgame of intelligent driving.

Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.

Comments

We need your insight to fill this gap
Leave a comment