In 2027, Horizon Robotics and the Momenta faction will go head-to-head in the chip sector.
This collision, therefore, is far less like a launch event, but more like two production lines catching sight of each other's headlights in advance at a curve.
On October 29, when XHEART unveiled its X7 Ultra, what truly silenced the audience was the next slide on the PPT: X9, 3000 TOPS, 700 GB/s, 2027.
This is no regular upgrade. The 3000 TOPS computing power is more than 5 times that of Horizon Robotics' current flagship J6P, and the 700 GB/s memory bandwidth is also more than three times that of J6P. CEO Li Zonglin made a very bold statement: the X9 will still follow the rhythm of "mass producing one generation, developing one generation, and pre-researching one generation", continue to follow the path of "large model defining chips", and carry out hardware-software co-design around the next-generation world model. Translated into plain language: XHEART has no intention of being a runner-up in the current urban NOA cost war, and it wants to move the entire game table directly to 2027.
The problem is that the year 2027 is not marked by XHEART alone. Horizon Robotics' next-generation flagship Journey 7 series is also scheduled for 2027. The goal of J7P is more straightforward: its computing power will "greatly exceed NVIDIA Thor-X", it will be equipped with the 4th generation BPU "Riemann", and it will debut synchronously with Tesla's next-generation AI5. In other words, the time gap that X9 intends to seize is never intended to be reserved by Horizon Robotics.
This collision, therefore, is far less like a launch event, but more like two production lines catching sight of each other's headlights in advance at a curve.
On one side, there is the mass production river that Horizon Robotics has waded through for six years. In the first half of 2026, it took 31.94% of the full-level intelligent driving chip market share of Chinese self-owned brand passenger vehicles, followed by NVIDIA with 29.38%. It took five years for Journey chips to grow from 0 to 10 million units, and only 12 months to increase from 10 million to 15 million units. The current J6P, with 560 TOPS computing power and about 205 GB/s bandwidth, has been installed in EXEED ET5, Deepal L06, iCAR V27, and Fengyun T9L; its HSD urban assisted driving solution has been deployed on four different categories of car models in less than half a year. The most powerful advantage of Horizon Robotics is not a single specific chip, but that it has turned high-level intelligent driving into a product that can be continuously reduced in cost: from 200,000-yuan-class models down to 150,000-yuan-class, even lower. The Horizon Cell "magazine system" is like a set of pluggable base, on which carmakers can make differentiated adjustments, but the base itself is controlled by Horizon Robotics.
On the other side, XHEART's genes are more like Momenta extending its reach into the silicon wafer field. Li Jun once served as the head of the chip department at Momenta, Li Zonglin used to work at OPPO's Zeku, and earlier worked on MediaTek's Dimensity series; Cao Xudong and Xia Yan hold indirect shares through the controlling shareholder. During the planning period of X7, the team worked side by side with Momenta: figuring out which operators need acceleration, which segment of memory bandwidth should be reserved for the model, and deriving the design from the algorithm requirements. Its logic is very smooth: once the end-to-end large model is widely applied, the architectural redundancy of general-purpose chips will become an extra tax; theoretically, people who understand models can make chips that put every bit of computing power and bandwidth to the most critical use.
But between "theoretically" and "in mass production", there are vehicle specification compliance, adaptation, verification, supply consistency, and pages of checklists that the procurement departments of carmakers fill in. Momenta can provide XHEART with model insights and early vehicle access channels, and X7 has already been adopted in the Momenta R7 solutions for SAIC Volkswagen ID.ERA 9X and MG 07; but once chips move towards large-scale deployment, the question is no longer "do you have a solution", but "how many models are you on, what is your shipment volume, and how deep is your customer structure". XHEART has raised a total of about 5 billion yuan in financing, with investors including Renault, Chery, and Changxin Technology, which sends a very strong signal. So strong that only four days after the disclosure of its strategic round, Chery released its own Ziwei chip; NIO, XPeng, Li Auto's self-developed chips have also been mass-produced and installed on vehicles, and XPeng's Turing chip even began external supply. Carmakers can vote for you with equity investment, but they can also keep you out of the market with their orders.
This is where the sense of gunpowder comes from. By lifting the parameters to 3000 TOPS and 700 GB/s, X9 is essentially stating publicly: the next round of competition will not be "whoever is cheaper will get NOA widely deployed", but "who can feed larger world models on vehicle-spec chips". The quantitative expectation put forward by Cao Xudong is even more aggressive: the scale of the end-side model will be expanded by another 100 times. It sounds like a technical declaration, but in the market it is a fight for definition power - if the world model really becomes the backbone of intelligent driving, then the chips, compilers, toolchains, and data closed-loop customized around it will all become new moats; whoever masters the backbone will no longer just be a supplier, but hold part of the "soul" just like Tesla.
Horizon Robotics certainly will not stand by and watch. The most frequent words Yu Kai has said over the years are roughly about seizing the right market rhythm. While J6P is still seeing rapid shipment growth, its Starry Sky cockpit-driving integrated chip is already on the way, and the "Riemann" BPU of J7P is also targeted at the next-generation model load. It no longer only speaks with scale and cost, but also starts to fight back with the language of "model defining architecture". For XHEART, this is the most uncomfortable situation: it originally wanted to bypass the frontal battlefield of J6P and plant its flag at the next commanding height; but when it looks up, it finds that Horizon Robotics has already come up from the heavy track of mass production, software stack, and vehicle model network, with the engineering muscle tempered by 15 million units of shipments in hand.
Therefore, the collision in 2027, on the surface, is 3000 TOPS vs "exceeding Thor-X", 700 GB/s vs "Riemann"; fundamentally, it is the first showdown between the two different hardware-software integration paths. Horizon Robotics first built a scalable chip platform, then developed HSD to define experience, lock in customers, and spread R&D costs; the XHEART/Momenta faction first gained model understanding and mass production solutions, then let the chips grow out according to this set of understanding. The former is heavy, but every link has been tested on real vehicle models and real sales; the latter is pure, with a beautiful architecture story, but it must convert this purity into the certainty, cost and delivery timeline that carmakers dare to place orders for.
This is where the intensity of the competition lies. They are competing for the same group of Chinese carmakers, the same window from urban NOA to world model, and the same group of customers who have been educated by "self-developed chips"; even their narratives have begun to fight closely - you say large model defines chips, I say BPU is oriented to next-generation load; you say the end-side model is 100 times larger, I say the magazine system has already deployed high-level intelligent driving to 150,000-yuan-class models. No one can scare the other away with a single parameter. The cruelty of the chip industry never lies in the launch event, but in how many vehicle models you have, how much cost advantage you retain, and how fast your iteration cycle is when your opponent arrives at the same platform.
X9 is a declaration of war, and also an examination paper. Horizon Robotics has proved that it can complete the full cycle in six years; now XHEART needs to prove that it can break into this cycle from Momenta's algorithm side, and will not be crushed head-on by Horizon Robotics' scale, software ecosystem and customer trust by 2027. 2027 has not arrived yet, but the track has narrowed in advance: one side has run a long way covered in mud, the other rushes to the intersection with new shoes and a map. When the collision happens, the parameters will step back, and the first sounds you hear will be the braking sounds, the order sounds, and the question from carmakers - "Can you deliver stably next year, for real?"
This article is from the WeChat Official Account "Co-opetition AI", author: Co-opetition, published with authorization from 36Kr.