China Finalizes Mandatory L3/L4 Autonomous Driving Standard, 2000 TOPS Set as Entry Threshold

China moves autonomous driving from voluntary to mandatory regulation

China's Ministry of Industry and Information Technology (MIIT) has opened public consultation on a draft mandatory national standard for intelligent connected vehicle autonomous driving systems, marking the first time L3 and L4 autonomous driving will be governed by enforceable technical regulations rather than voluntary guidelines. The standard is expected to take effect on July 1, 2027.

Under the new rules, any vehicle failing to meet the standard cannot be manufactured or sold. Non-compliance will constitute a legal violation. The shift from "recommended" to "mandatory" signals that China's autonomous driving industry has entered a phase of strict regulatory oversight.

Meanwhile, multiple Chinese cities have already begun issuing L3 autonomous driving road test and demonstration operation licenses, pushing compliant mass production closer to reality.

Three pillars converge at a critical inflection point

According to Peng Xueming, senior chief engineer at Desay SV, the three foundational pillars of autonomous driving — algorithms, data, and computing power — are simultaneously approaching a critical threshold that will define the industry's trajectory over the next five years.

Algorithm design has shifted decisively from rule-driven programming to AI-native approaches. End-to-end models, world models, visual-language-action (VLA) frameworks, and reinforcement learning are now the dominant paradigms, allowing vehicles to interpret environments and make decisions through learned patterns rather than hand-coded logic. This transition has driven exponential growth in computing requirements.

Data quality has similarly evolved. As models move from rule-based to data-driven training, the focus has shifted from sheer volume to high-quality data covering long-tail scenarios. Leading companies have built systematic data closed-loop iteration capabilities, widening the gap between early movers and late entrants.

2000 TOPS becomes the L3 benchmark

Peng Xueming stated that 2000 TOPS of computing power is the baseline threshold for AI-native Level 3 autonomous driving. For Level 4, the requirement jumps significantly — he estimated that a full vehicle AI agent will need to evolve toward over 6000 TOPS.

Several Chinese automakers are already equipping production vehicles with chip platforms in this range, even though current capabilities remain at L2 level:

• Li Auto's L9 Livis is equipped with its self-developed Mach M100 chip, delivering 1280 TOPS per chip, or 2560 TOPS with dual chips.
• XPENG's GX flagship model carries three self-developed Turing AI chips for a combined 2250 TOPS; the Robotaxi variant uses four chips for 3000 TOPS.

These choices reflect forward-looking platform strategy — building headroom for L3/L4 rather than optimizing solely for today's feature set.

"Big-big brain" architecture gains traction for L4

The path to L4 presents a unique challenge: no single chip on the market currently delivers sufficient compute. The industry's current approach for L3 has been a "big-small brain" architecture, pairing a high-compute main chip with a lower-compute secondary chip running independent software stacks for safety redundancy.

However, Peng argued that this approach falls short for L4, which demands both far greater main-domain compute and stricter safety redundancy. Desay SV is pursuing an alternative — a "big-big brain" architecture that uses two high-compute chips sharing a unified software stack, with fault isolation achieved through hardware partitioning alone. The primary unit handles global perception and core decision-making, while the secondary unit provides independent safety redundancy.

This design avoids the costly platform rebuild that would be required when migrating a "big-small brain" system from L3 to L4, where two separate algorithm sets, toolchains, and verification systems would need to be redeveloped.

Industry window is narrowing

Visionaries across the industry have published ambitious timelines. XPENG CEO He Xiaopeng predicted that L4 and even L5 autonomous driving could be realized within three to five years. Horizon Robotics CEO Yu Kai forecast 100 percent hands-off driving by 2028, L4 "eyes-closed" capability by 2030, and full "sleeping driving" by 2035.

While timelines differ, the consensus is clear: the next five years will determine which companies define the industry landscape, and computing architecture choices made today will shape that outcome.