QCraft’s co‑founder and CEO, Yu Qian, has framed 2026 as the moment autonomous driving moves from early commercial runs to broad consumer adoption. He predicts that by next year, ‘city NOA’ — driver assistance that can navigate urban roads — will begin to appear as standard equipment on cars priced around ¥100,000 (roughly $14,000), a price band that defines the bulk of China’s passenger market.
Yu traces the shift to a technological inflection: the industry is coalescing around end‑to‑end architectures, and the incorporation of VLA (Vision‑Language‑Action) models and so‑called world models will let driving systems learn from massive pools of real and synthetic data. That learning, he argues, will allow autonomous systems to reach safety levels several times better than human drivers and to scale rapidly across large fleets.
QCraft itself is using its momentum to underline the point. The company announced that its assisted‑driving stack has now been deployed on more than one million vehicles as of January 2026, and that its city NOA package — built around the Horizon (地平线) single‑chip SoC 6M — is in production and fitted to Li Auto’s refreshed L series with the AD Pro option. For Yu, the million‑vehicle threshold is not just a marketing milestone; it is the tipping point that creates a robust data closed‑loop and accelerates continuous system improvement.
That data argument sits at the centre of how vendors hope to turn technical advances into consumer value. Yu emphasises that users do not buy concepts such as VLA; they buy experiences that feel as safe and competent as an experienced driver. Hence, the differentiator in the new phase will be resource efficiency and product engineering: delivering standout performance with lower compute and power budgets rather than simply pursuing raw model size.
On technical debates that have split the industry — pure vision versus lidar, and incremental modular stacks versus full end‑to‑end learning — Yu adopts a pragmatic posture. He says pure‑vision systems can already deliver very good urban and highway NOA experiences and casts lidar as a ‘safety plugin’ that adds redundancy in the most challenging scenarios. He also voices scepticism about narrow L3 deployments that cannot be practically scaled: for Yu, usefulness depends on a wide and simple operational design domain (ODD).
Competitive dynamics are in play too. Yu welcomes Tesla’s Full Self‑Driving (FSD) moving into China as beneficial for market education and expansion, while warning that successful entrants must adapt to China’s complex road environment and local compliance requirements rather than merely transplanting foreign products. He predicts the market will evolve into an ‘‘one‑super, many‑strong’’ landscape by 2026: a dominant leader alongside several credible challengers, rather than a single global winner.
The implications extend beyond engineering. Yu forecasts city NOA takeover rates moving to a monthly cadence for users and foresees autonomous‑specific insurance premiums falling to less than half the cost of human‑driver policies, should the safety promises be borne out. He also highlights QCraft’s compatibility strategy — building stacks that work across EVs and internal‑combustion models — as a lever for international expansion where EV penetration remains lower than in China.
For global observers, the coming year will be telling. The combination of cheaper hardware, large deployed fleets, and learning‑first software could finally bring advanced driver assistance to ordinary cars at scale. But regulatory alignment, realistic performance claims, and the thorny business of liability and insurance will determine whether the promised safety gains translate into durable market transformation.