Gasgoo Munich-"If you can't become a large-model company, you have no chance." That was the verdict DeepRoute.ai CEO Zhou Guang recently reiterated in an interview.
This isn't the first time he's said it.
Back in 2024, he predicted the autonomous driving industry had to pivot to large-model-driven systems — or face an existential mismatch. At the time, most listeners were skeptical.
"When I first said it, people probably didn't think it would happen. After GPT-6, everyone should realize: it's happening."
What brought this forecast back into the spotlight was a recent real-car test by GPT-6 Astra. For the first time, it offered visible proof that a general-purpose large model can control a physical vehicle through driving tasks.
Zhou puts the window at two to three years.
Fail to pivot to the large-model paradigm by then, and an autonomous driving company risks being priced out of the game.
Some see those words as a wake-up call; others dismiss them as industry hyper-anxiety.
A 134-Meter Slow Motion Reveals Two Coordinate Systems
In September, independent evaluator DrivingBench ran a public test. It tasked four general-purpose large models — GPT-6 Astra, Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol — with directly controlling a 2022 Toyota Corolla via tool interfaces on a closed course. The route covered starting, cone weaving, reversing, and parking, spanning 134.7 meters.
GPT-6 Astra completed the course on its second attempt in 5 minutes and 22 seconds, averaging roughly 1.5 km/h. The other models logged completion rates between 6% and 45%.
Around the same time, GPT-6 Astra — without any specialized robotics training — also topped the RoboDojo simulation leaderboard.
A 1.5 km/h crawl on a closed course with a simple route and multiple attempts offers virtually no reference value for real-world driving. DrivingBench itself conceded the tasks were "all very simple" — no complex traffic flows, no random pedestrians, no continuous urban driving.
Yet the symbolism is potent: for the first time, a general-purpose large model with no specialized driving training controlled a real car through a complete set of actions via tool interfaces.
That had never happened before September 2026, and it gave Zhou’s 2024 forecast its first visible footnote.
Place this against the reality of China’s autonomous driving roads, and the gap between the two coordinate systems snaps into focus.
General-purpose large models today: closed courses, low speeds, simple routes, multiple attempts. Mass-produced self-driving systems in China: urban roads, average speeds of 20 to 30 km/h, top players hitting 100-kilometer MPI (miles per intervention), and daily commutes by millions of users.
Still, the other side of the market landscape deserves attention.
In the first half of 2026, Huawei held 18.1% of China’s high-tier autonomous driving market, Tesla 13.2%, and Momenta 12.2%. More than a dozen other players split the rest.
The market remains fragmented. If the prediction that "only a few will survive" holds true, the industry shakeout has barely begun. Conversely, no single player has established absolute dominance — leaving room for newcomers and those pivoting.
Zhou’s forecast, meanwhile, leans more on the slope of technological evolution than on a snapshot of current capabilities.
When GPT-2 and GPT-3 debuted, the outside world shrugged, Zhou noted — yet just two years later, the landscape was unrecognizable. This evolutionary pattern breaks from the past.
That means general-purpose large models start from a low baseline on driving tasks. Whether their progress curve will overtake traditional self-driving pipelines is the industry’s core fault line.
No Self-Driving Firm is Standing Still
To gauge whether "traditional self-driving company" still holds true, look at the technology roadmaps publicly disclosed for 2026.
Huawei Qiankun opted for the WEWA 2.0 architecture. At its April 2026 tech summit, the company launched ADS 5, officially billed as a "full upgrade from rule-driven to an AI agent for autonomous driving." Its cloud-based world engine introduced a "multi-agent game" mechanism, using online reinforcement learning to generate, learn, and verify simultaneously — a 10x leap in training efficiency. On the vehicle side, the world behavior model applied safety risk field theory to generate real-time risk heatmaps.
In September, Jin Yuzhi revealed Qiankun had raised its full-year R&D budget from 18 billion RMB to over 19 billion RMB. By August, Huawei Qiankun’s self-driving systems had surpassed 2 million installations.
Momenta’s public roadmap layers a world model atop its "one flywheel, two legs" strategy. When it launched the R7 world model in April 2026, CEO Cao Xudong declared: "Prediction is the cornerstone of intelligent evolution; the core of physical AI lies in a deep understanding of the world’s fundamental physical laws."
Per its prospectus and public filings, Momenta has amassed over 13 billion km of real-world driving data and more than 100 million golden training segments. It draws on real data from 800,000 mass-produced vehicles for world model pre- and post-training.
In late October, Cao stated publicly that scaling L4 requires roughly 100 billion km of data. Momenta listed on the Hong Kong Stock Exchange in July 2026.
Image Credit: Tesla
Tesla is pursuing data scale and end-to-end models in parallel. Supervised FSD became officially available in China on May 21, 2026. Overseas markets shifted to a $99/month subscription in February; mainland China still offers only a 64,000 RMB one-time purchase, with no timeline for a subscription tier.
DeepRoute.ai is betting on foundation models. It bypassed the world-model route in favor of foundation models. Zhou’s rationale: "Generative tasks don’t necessarily demand massive parameter counts." He offered a vivid analogy: humans may have evolved for a long time before developing language and requiring a 1,500cc brain; most animals get by on roughly two milliliters, yet they still operate within a world model.
He views VLA as a subtask of foundation models, not the endgame. The real bottleneck is the "intelligence ceiling" — simple models fail in long-tail scenarios requiring causal reasoning. Public data shows DeepRoute.ai ranked second industry-wide in third-party insurance registrations in August 2026.
Line up these four roadmaps, and a common thread stands out: no top-tier self-driving player in 2026 is clinging to traditional modular, rule-driven architectures. Whether it’s Huawei’s WEWA, Momenta’s R7 world model, Tesla’s end-to-end FSD, or DeepRoute.ai’s foundation model, the public messaging all points toward large models.
The real debate isn’t whether to make the leap — it’s about three unresolved questions:
Parameters and architecture: Which is more efficient — a general-purpose foundation model, a vertical world model, or the VLA extreme? Data flywheel weighting: At what scale does intelligence naturally emerge, or must the right paradigm come first? Commercialization pacing: Scale up installations and iterate later, or perfect the model before chasing volume?
These are open industry questions, and for now, there are no public answers.
The Window’s End: 2027 Is the Pivotal Year
Zhou’s "two-to-three-year window" happens to overlap with several fixed industry milestones.
The first is the mandatory L3 national standard. On July 30, 2026, China released GB 44721-2026, "Safety Requirements for Intelligent Connected Vehicle Autonomous Driving Systems," taking effect July 1, 2027, for L3 and L4 vehicles. As China’s first mandatory L3/L4 safety standard, it turns what was once an aspirational policy into a hard regulatory gate.
The second is the scaled deployment of robotaxis. Zhou pointed to a trend: "Scalable robotaxis are moving fast — just look at Tesla’s rollout speed. Some companies spent a decade to field one or two thousand cars; others are rolling them out in large batches right out of the gate." He argues that MPI climbing from 100 km to 1,000 km, then to 10,000 km, marks the key thresholds for L4 scaling. "Today, no matter the L4 system, there’s a human in the loop, layered with cloud safety strategies. At the 10,000-km level, you can basically start scaling."
Image Credit: VOYAH
The third is validating the self-driving revenue model. FSD’s 75% overseas subscription rate, the take-up rate for the 64,000 RMB buyout in China, and domestic automakers’ option rates are answering a fundamental question: Will consumers keep paying for high-tier self-driving features? The answer shapes the business model for every player in the field.
Put these milestones together, and the industry consensus is clear: 2027 is the pivotal year.
As for how many years "traditional self-driving companies" have left, a preliminary answer may not emerge until mid-2027. By then, the arrival of L3-compliant models, the expansion of robotaxi operating cities, the production performance of next-gen models, and the driving capabilities of the next GPT release will together paint a much clearer industry landscape than today’s.
For the past two years, the self-driving industry has obsessed over whose end-to-end system is more advanced, whose urban NOA feels smoother — rivalries that, at their core, were just infighting within the same technical paradigm. Zhou’s latest forecast yanks the industry’s gaze from "who does it better" back to "should we be switching tracks entirely."
The irony is that everyone is already switching tracks — just with different form and pacing. Huawei is upgrading to WEWA, Momenta is building the R7 world model, Tesla’s FSD is already an end-to-end large model, and DeepRoute.ai is going all-in on foundation models.
The real debate was never about whether to pivot to large models. It’s about model size, data scale, and the pace of market deployment.
As Zhou put it, the steam-engine era of autonomous driving has arrived.
The industry is waiting for a new answer. Notably, until it arrives, everyone still at the table is already restacking their chips.
