Horizon Rolls Out Full Auto AI Stack
By Chen Wei

Image / pandaily.com
China’s car AI just got a turnkey stack: Horizon’s Starry chip, KaKaClaw OS, and HSD V1.6.
On April 22, Horizon Robotics announced a trio of products that together complete what the company calls a full-stack strategy—from chips to software, and from smart driving to a smart cockpit. The centerpiece is the Starry 6P, a 5nm automotive-grade chip offering 650 TOPS of BPU computing power and 273 GB/s of memory bandwidth, designed to support both AI cockpit functions and autonomous driving models in a single, centralized architecture. Horizon touts a unified memory approach that replaces traditional domain-separated systems, claiming a 50% reduction in hardware footprint and a RMB 1,500–4,000 per-vehicle cost saving (roughly $210–$560). Development cycles shorten dramatically—from about 18 months to 8 months—and the delivery time for an integrated cockpit-driving system is cut by 56%.
The Starry chip sits alongside KaKaClaw, Horizon’s agentic vehicle operating system, and HSD V1.6, the full-scenario assisted driving system. In Horizon’s framing, this trio moves the company beyond component supply toward enabling a complete, end-to-end vehicle intelligence stack. The company says the Starry 6P has already secured intended mass-production partners, signaling a shift from R&D showcase to real-world installation timelines, even as OEMs weigh the implications of a vertically integrated platform.
Key Chinese terms translate into policy-relevant context. The Starry chip is described as built on a 5nm automotive-grade process—an industry shorthand for cutting-edge silicon designed to meet stringent auto-grade reliability standards. TOPS, or tera-operations per second, is used to quantify AI inference throughput; in Horizon’s figures, 650 TOPS signals substantial on-chip AI capability. The truly transformative claim is the “unified memory architecture,” which Chinese engineers frame as a way to centralize compute and reduce system complexity on the vehicle. The KaKaClaw OS, described as an agentic operating system, embodies a Chinese narrative about vehicle learning and decision-making as a software-dominant, safety-conscious stack. Taken together, the Starry-KaKaClaw-HSD package is positioned as a self-contained ecosystem rather than a modular kit.
From a supply-chain vantage, this represents a notable bet on a single-stack approach in a market long accustomed to multi-vendor ADAS supply chains. If Horizon’s ecosystem gains broad OEM adoption, the company could become a focal point for both chip supply and software updates, potentially compressing the time-to-market for new features. Yet the absence of named, publicly disclosed mass-production partners leaves some uncertainty about scale, interoperability with existing vehicle platforms, and the degree to which Tier-1 suppliers will align around Horizon’s stack.
Industry watchers should watch a few critical signals. First, deployment timelines and partner announcements will reveal how aggressively OEMs commit to a fully Horizon-driven architecture versus mixing Horizon silicon and software with third-party ADAS modules. Second, price and total-cost-of-ownership claims will hinge on the real-world efficiency of the unified memory approach and the costs of integrating KaKaClaw with legacy vehicle architectures. Third, policy and provincial incentives for domestic chip ecosystems could accelerate Horizon’s adoption if vehicle makers see a clearer path to localization and after-sales support.
For global manufacturers evaluating supply options, Horizon’s rollout underscores how quickly China is maturing from a component supplier to a platform provider. If Horizon’s full-stack approach proves scalable and interoperable across multiple vehicle lines, buyers should price in not just chip costs but the downstream benefits—simplified integration, faster feature rollout, and a potential shift in who ultimately controls the vehicle compute stack.
Two to four practitioner takeaways: expect OEMs to demand tighter performance guarantees around AI latency and reliability; anticipate a premium on centralized compute for new features but pressure on service networks to support Horizon-based platforms; and monitor for regional incentives that shape which automakers commit to this stack versus a blended, multi-vendor approach.
Sources
Newsletter
The Robotics Briefing
A daily front-page digest delivered around noon Central Time, with the strongest headlines linked straight into the full stories.
No spam. Unsubscribe anytime. Read our privacy policy for details.