Industry

Industry — Understand the Playing Field

1. Industry in One Page

Horizon Robotics sits in the ADAS and autonomous-driving silicon-and-software stack for passenger vehicles — a "Tier-2" supply layer that sells AI processing chips (the Journey SoC family), perception/planning algorithms, and developer toolkits to automakers and the Tier-1 integrators who package them into cars. Buyers pay because every new car increasingly needs computer-vision compute for active-safety features (lane keeping, automatic emergency braking) and, at the high end, NOA — Navigate-on-Autopilot, the urban/highway hands-off feature that has become China's must-have spec sheet item below ¥200,000.

Profits exist because the silicon is hard (multi-year design cycles, automotive-grade qualification, software co-development with each OEM) and once a chip is "designed in" to a vehicle program, switching during the 5-7 year model life is prohibitively expensive. The cycle is driven by (a) Chinese auto unit volumes, (b) the NOA attach rate, and (c) feature democratization — the pace at which urban NOA falls from ¥300,000+ premium cars into the ¥100,000–¥200,000 mainstream. This is not a chip business; it is a car-program design-win business. Revenue lags design wins by 18-30 months, so today's commercial momentum reflects decisions OEMs made in 2023-2024.

Where Horizon plays: layer 3 (chip + software). Its closest competitive comparison set is Nvidia DRIVE, Mobileye EyeQ, Qualcomm Snapdragon Ride, Huawei MDC, and Black Sesame's Huashan series.

2. How This Industry Makes Money

The revenue model is a two-stream package: the OEM pays a per-vehicle hardware fee for each Journey SoC shipped (variable, ramps with vehicle production volume), and a license-and-services fee for the algorithms, toolchain, and per-program engineering work (often front-loaded around start-of-production, sometimes royalty-style). Pricing units differ sharply: hardware is per-unit at chip-level ASPs of roughly ¥350-¥2,000 depending on TOPS tier; software/services are per-program, sometimes per-feature, sometimes per-fleet-mile.

The cost structure is fixed-cost-heavy, capex-light. There are no factories — these are fabless semiconductor businesses outsourcing to TSMC. Capital spend is mostly office build-outs and test fleets. The dominant expense is research and development (RD) — the engineering teams that write perception models, port them to chips, and re-validate against every OEM's ECU architecture. For a sub-scale player like Horizon this means RD running above 100% of revenue while the design-win base ramps; mature peers (Mobileye) sit at ~60%, Nvidia at ~9% (because revenue is enormous and the auto group rides on a vastly larger gaming/data-center cost base).

Bargaining power tilts toward whoever owns the algorithm-to-silicon co-design. OEMs who try to do it themselves (Tesla, NIO, Li Auto, XPeng) need years and billions to match a specialist. OEMs who outsource find that switching chip vendors mid-program means re-writing the entire perception stack — so once Horizon wins a vehicle program, it tends to keep that program for the model's full life cycle. Tier-1 integrators (Bosch, ZF, DENSO) are partners, not gatekeepers; in Horizon's case 95%+ of 2025 shipments flowed through ecosystem partners, which is unusual for a chip vendor and a real moat-builder.

Software/IP carries 90%+ gross margins, hardware sits in the 40s. Blended margin therefore moves with the mix between license-and-services (front-loaded, lumpy, high-margin) and chip shipments (recurring, scaling, lower-margin). Horizon's blended automotive gross margin compressed from 79% in FY2024 to 67% in H1 2025 not because economics worsened but because hardware revenue grew 250%+ as Journey 6 entered mass production.

3. Demand, Supply, and the Cycle

Demand flows through three multiplicative drivers: vehicle units × ADAS attach rate × per-vehicle silicon content. China's new-vehicle ADAS attach rate is now over 60% (Digitimes, Nov 2025), with L2+ penetration at 55.8% in 2024 (up from 38% in 2023) and urban-NOA preinstall at 8.6% in 2024 — material headroom. Each generational step (basic ADAS → highway NOA → urban NOA → L3 hands-off) roughly doubles the silicon revenue content per vehicle. Horizon's per-vehicle hardware ASP rose above 75% YoY in 2025 even as unit shipments grew 39% — the mix-shift toward high-end Journey 6 SKUs is the actual story.

Supply has two real bottlenecks. First, leading-edge wafer access at TSMC for the higher-tier Journey 6 series (5nm/7nm) — geopolitically exposed if US export rules tighten further on China-bound automotive AI chips. Second, automotive-grade qualification (AEC-Q100, ISO 26262 ASIL-B/D, OEM-specific endurance testing) takes 18-24 months and acts as a structural barrier; new entrants cannot ship volume even with a working chip.

Where the cycle hits first: average chip ASP, then gross margin (mix), then license-and-services lumpiness (program-launch timing), then cash burn. The industry has no real downturn history yet — China's smart-driving stack only became a true mass market in 2023 — so professionals pattern-match to mobile-handset chip cycles (where ASP compression after a feature plateau eventually crushed margins). The bull case is that NOA democratization keeps per-car content rising for the next 3-5 years; the bear case is that OEM price wars force chip vendors to share more of the value before content-per-vehicle saturates.

4. Competitive Structure

This is not a fragmented market. Inside China's high-end urban-NOA compute platforms — the segment that matters for incremental revenue growth — the top three suppliers held roughly 89% combined share in 2025. Below the high-end tier, the basic L2 ADAS market is more crowded but lower-value.

Three distinct competitor types. Global silicon majors (Nvidia, Mobileye, Qualcomm) bring vast RD scale but face geopolitical risk in China and slower OEM-by-OEM customization cadence. Chinese vertical integrators (Huawei, BYD's in-house DiPilot, Tesla's FSD) capture the entire stack but only serve their own or affiliated brands. Independent Chinese merchants (Horizon, Black Sesame) sell to anyone, including OEMs that explicitly do not want to depend on Huawei or a competitor's in-house team — a position Horizon describes as the "greatest common divisor" of Chinese smart driving.

5. Regulation, Technology, and Rules of the Game

Three exogenous forces shape this industry's economics over the next 3-5 years: Chinese industrial policy that explicitly favours domestic ADAS/AD silicon, US export controls that limit access to leading-edge wafers, and regulatory approval of higher-autonomy levels (L3+) that unlocks higher per-car silicon content.

Two technology shifts also reshape economics. End-to-end neural-network driving stacks (replacing rule-based perception+planning pipelines) raised the compute bar — driving Journey 5 → Journey 6 transitions and a 4× TOPS jump (Journey 6 peaks at 560 TOPS). And the BPU (Brain Processing Unit) architecture race — proprietary neural accelerators tuned to specific automotive workloads — is the actual competitive moat, far more than raw TOPS numbers.

6. The Metrics Professionals Watch

Generic semiconductor ratios (P/E, EPS) are nearly meaningless here because the leaders run at large losses while scaling design wins. The right metric set is built around design-win velocity, content-per-vehicle, and RD efficiency.

Two metrics deserve a sharper eye. Hardware-vs-license mix flips as a chip family matures: the license/services revenue is front-loaded engineering paid by the OEM around start-of-production, then chip shipments compound during model life. A flip from license-heavy to hardware-heavy is bullish (programs are launching) but compresses headline gross margin. RD-to-revenue is the counterweight: it rises before scale, then falls fast once a chip family hits volume — Mobileye sits at ~60% today, mature analog peers below 20%, Nvidia at ~9%. The path from 137% (Horizon 2025) to under 50% is the crux of the equity story.

7. Where Horizon Robotics Fits

Horizon is the dominant independent merchant in China's ADAS+AD silicon stack — not the largest globally (Nvidia), not vertically integrated (Huawei, Tesla, BYD's DiPilot), and not focused on cameras/EyeQ-style L2 only (Mobileye). It occupies the lane that benefits most from China's policy push for domestic auto silicon while remaining acceptable to OEMs that refuse to lock themselves into a competitor's stack.

Horizon is best read as a scale-stage challenger in the most strategically important corner of the Chinese auto-tech stack: dominant in mainstream ADAS, gaining on Huawei in high-end NOA, and the only credible non-vertical-integrator alternative for OEMs that want NOA without buying it from a rival. The company sits where Chinese policy, OEM neutrality preferences, and feature democratization converge.

8. What to Watch First

1. Monthly Chinese new-vehicle production and L2+/NOA attach rates (CAAM data; OEM monthly delivery announcements). Penetration toward 70%+ for ADAS or 30%+ for NOA is a clear tailwind.
2. Sub-¥200,000 NOA model launches by BYD / Geely / Changan / Chery. Each launch is a real-world test of feature democratization speed; Horizon's 44% share in this bracket is the single biggest revenue lever.
3. High-end urban-NOA compute share trajectory: Nvidia / Huawei / Horizon. The Huawei–Horizon gap was 0.8 percentage points at the end of 2025. A crossover would be a category-defining inflection.
4. Quarterly Journey-series unit shipments and ASP trend. ASP up + units up = mix is winning; ASP up + units flat = slowing volume; ASP down at all = price war contagion.
5. US export-control announcements affecting automotive AI accelerators or TSMC's China business. Direct supply-side risk to Journey 6 high-tier SKUs; would force fab diversification (SMIC) at lower performance.
6. RD-to-revenue ratio at half-year and full-year prints. The path from 137% (2025) to under 50% is the operating-leverage thesis; any divergence resets the breakeven date.
7. L3 commercial-deployment approvals in select Chinese cities. Unlocks the next tier of per-vehicle silicon content; first commercial L3 SKUs expected late-2026 / 2027 if regulators move on schedule.