An engineer at a Chinese automaker has a familiar problem. The cockpit chip already runs navigation, music and three displays. Now the product team wants a language model that answers requests inside the car without waiting for a network. The current chip cannot do both jobs at once, so the usual fix is to add a second computer behind the dashboard, with its own memory, power rail and cooling.
SemiDrive (芯驰科技), a Nanjing-based automotive chip designer, says its X10 removes that second computer. The chip is built on a 4nm automotive process and Arm v9.2 architecture, and the company says it can run the cockpit and a local large model (大模型) on one piece of silicon. It is a claim about system cost as much as speed, and it has begun to reach the market.
What is actually in the X10
SemiDrive publishes a specific set of parameters. According to figures on the company's own website, reproduced in November 2026 by the trade publication EEFocus, the X10 delivers:
- CPU performance of 300 KDMIPS and GPU performance of 3,300 GFLOPS.
- An NPU rated at 80 TOPS, the figure that matters for running a model locally.
- Memory bandwidth of 154 GB/s, which the company says is enough to deploy on-device models of up to 9B parameters.
- Architectural support for dynamic expert routing in mixture-of-experts (MoE) models.
The memory-bandwidth number is the one an engineer will look at first. A language model spends much of its inference time moving parameters from memory into the compute units; if the memory channel is narrow, a large NPU sits idle waiting for data. Running AI alongside navigation and multi-screen display makes that scheduling problem harder, not easier.
SemiDrive's answer is integration. The X10 combines traditional cockpit computing and model inference in one chip, replacing what the industry calls the "cockpit SoC plus external AI box (座舱SoC+外挂AI Box)" arrangement. The company also supplies standardised model invocation interfaces and pre-adaptation for mainstream open-source models, so that developers do not have to port everything from scratch.
The cost argument, and its limits
The company's case rests less on raw speed than on the bill of materials. Based on SemiDrive's published schemes, EEFocus calculated that the X10 can cut DDR memory capacity by about 25 percent and lower total system cost by CNY 1,500–3,000 (approximately US$208–417 / HK$1,630–3,260, at US$1≈¥7.2 and HK$1≈¥0.92). Because power is controlled through process and architectural choices, the chip can be cooled by air rather than by a more complex liquid loop, the company says.
Those are company calculations, not the result of an independent teardown. No customer has published a before-and-after bill of materials. The 25 percent and the cost range should be read as a claim awaiting verification in a real vehicle programme.
Where the pressure is real is in the price band. Gasgoo Research Institute data cited by EEFocus puts SemiDrive's cockpit chip installations at 353,000 units in the first three quarters of 2026, in the domestic passenger-car market priced CNY 100,000–300,000 (approximately US$13,900–41,700 / HK$109,000–326,000). That was the highest among mainland Chinese local suppliers in that scope, at about 26 percent of local suppliers' installations. The China Passenger Car Association says the CNY 100,000–200,000 band (approximately US$13,900–27,800 / HK$109,000–217,000) was 41.9 percent of domestic retail sales in the same period, the largest single price band by volume. Automakers entering that market have to add AI without adding much to the invoice.
From a cockpit chip to a platform
SemiDrive says cumulative shipments of all its automotive-grade (车规级) chips have passed 15 million units across more than 170 vehicle models. It supplies central control hosts for SAIC Volkswagen, Dongfeng Honda and Chery; Gasgoo counted more than 110,000 chips for the SAIC Volkswagen Lavida alone in January–September 2026. The company describes itself as the only domestic chip maker with mass production across all five core domain-control scenarios — cockpit, assisted driving, chassis, body and powertrain. That "only" is the company's own description, not an independent ranking, and it should be read that way.
At the 2026 Beijing Auto Show, SemiDrive announced Strategy 2.0, extending its automotive technology into robots. The R1, D9 and E3-R series target, respectively, large-model inference, motion planning and real-time control, and joint and torso actuation for embodied intelligence (具身智能). The company says the D9 series is already in mass production with leading embodied-intelligence manufacturers, and it has worked with partners including Radxa Computer and GALBOT on controllers and toolchains.
That expansion is a bet that a chip company which has learned to meet automotive functional-safety and reliability standards can sell the same discipline into robotics. It is a strategy, not a disclosed revenue line. The core X10 business remains the near-term test: whether automakers sign volume agreements and ship cars with the single-chip cockpit.
According to Li An, chief scientist at BrainNet (脑机网), the test for a new computing platform is not its peak figure on a data sheet but whether an ordinary engineering team can reuse its software across several products. For the X10, the toolchain and the migration path from the older X9 matter as much as the 80 TOPS.
Honest limitations
This article draws on SemiDrive's own website and on a 10 November 2026 EEFocus article that republishes data from Gasgoo Research Institute. The X10 parameters, the approximate 25 percent memory saving, the CNY 1,500–3,000 cost range and the air-cooling claim all originate with the company or with calculations based on its published schemes; none has been confirmed by an independent teardown or by a named customer. Installation figures cover a specific price band and statistical scope, not the whole market. SemiDrive is private, so there are no audited financial statements, no disclosed X10 price and no named list of X10 customers. The "only domestic supplier" phrase is the company's. No comparison with competing cockpit system-on-chips from other vendors is offered here, and no claim is made about the chip's functional-safety certification status beyond what the company states.
What readers can do now
- If you are buying a car with an AI cockpit, ask whether the assistant runs on the car's chip or in the cloud, and whether the vehicle carries a separate computing box. The answer determines what still works with no signal.
- If you evaluate automotive chips, look at memory bandwidth and the software toolchain alongside TOPS. A high NPU number paired with a narrow memory path is a specification, not an experience.
- Treat integration and cost-saving figures as company claims until a named vehicle programme publishes its own before-and-after numbers.
