Yu Kai on the Next Round of Intelligent Driving: Good Algorithms Are No Longer Enough

Gasgoo Munich-On the afternoon of September 15, inside Horizon Robotics' new headquarters in Lingang.

Yu Kai had just taken his seat before the media after a ceremony marking a milestone: 15 million mass-produced Journey series chips. The first question came straight to the point: When will Horizon Robotics surpass NVIDIA?

Yu Kai initially deflected the question. "I think we really need to dial down the focus on who we’re surpassing," he said. "What matters more is who we become."

But he didn’t shy away from the rankings.

He described Horizon’s past strategy as "encircling the cities from the countryside" — starting with low computing power, moving to mid-range, and finally entering the high-performance market. Across all intelligent driving chips, Horizon currently holds the top market share; but looking strictly at high-compute chips, "we are roughly second this year," he said, with his sights set on taking the lead next year.

The 15 million figure is primarily a validation of the past. It demonstrates that a Chinese intelligent driving chipmaker has successfully bridged the gap from small-scale design wins to mass automotive-grade production, while building a substantial foundation in the basic ADAS and mid-range computing markets.

But high-level intelligent driving is a different game entirely.

Yu Kai didn’t name the current leader in high-compute chips, but NVIDIA remains the rival that cannot be ignored.

NVIDIA’s DRIVE Orin delivers up to 254 TOPS of computing power, powering or underpinning the development of Toyota’s next-generation models, JLR’s new lineup starting in 2026, and the Volvo EX90. Volvo also plans to migrate to Thor, which has already secured partnerships with Chinese automakers like BYD, Geely, Xiaomi, Great Wall Motor, and SAIC. The goal is to further integrate intelligent driving, cockpits, and even central computing into a more consolidated architecture.

The next-generation DRIVE Thor isn’t just about stacking more computing power; it’s about consolidating different workloads — from assisted driving to digital cockpits — onto a single in-vehicle computing platform.

This explains why the 15 million milestone and being "second in high-compute" are not contradictory.

Horizon has already proven it can sell chips. Now, it must prove whether it can carry that accumulated scale advantage into the next round of competition for high-level intelligent driving.

In the next round, algorithms will of course still matter. But simply having a good algorithm is no longer enough.

Algorithms are hard, but they aren't enough

For the past few years, the most visible competition in intelligent driving has centered on algorithms.

Highway NOA, city NOA, map-free navigation, BEV, Transformer, followed by end-to-end, world models, and reinforcement learning. Every few months, the industry churns out a new set of tech buzzwords.

At that stage, the mere existence of a feature was the differentiator.

If a car could navigate city streets, handle unprotected left turns, or recognize construction zones, that was enough to serve as a core selling point at a launch event.

Today, the situation has begun to shift.

As of September 8, installations of Huawei’s Qiankun Intelligent Driving had surpassed 2 million units, covering 25 brands and more than 60 models. Meanwhile, Momenta’s partnership with Audi has evolved from high-level driver assistance to L3 production capabilities, with the Audi E7X becoming Momenta’s first model geared toward mass-produced L3 driving.

Horizon, for its part, repeatedly highlighted another shift: using a single Journey 6M chip offering 128 TOPS without relying on lidar to bring city NOA capabilities to vehicles priced around 100,000 yuan.

More and more models are touting city NOA, door-to-door navigation, complex intersection negotiation, and active safety features.

This doesn’t mean the algorithms across different companies have become identical.

What has truly changed is that "having these features" is increasingly difficult to sustain as a long-term differentiator.

Getting a city NOA system to run is not the same as having it run reliably during a daily commute. Similarly, navigating a complex intersection once is different from doing so consistently and smoothly over the long term with minimal driver intervention. Beyond that, the question becomes: how much computing power, how many chips, and how many sensors are required to deliver that same experience — and can it be scaled down to lower price segments?

Competition is shifting from "who has the feature" to "who can deliver the same feature more reliably and cheaply."

Yu Kai used an analogy involving mobile phone basebands.

With mobile technology today, users can barely perceive any distinct "personality" differences between basebands. Yet, ensuring stable performance in open environments — amidst skyscrapers, in subways, or in weak signal areas — remains incredibly difficult.

"There is a category of technology that is highly converged and difficult to differentiate, but extremely hard to do well."

He believes autonomous driving fits this mold. "It operates in an open world with absolutely no script." It is in this context that Yu Kai predicts: "In the end, there will likely be no more than two or three global suppliers for core autonomous driving technologies."

The prediction of "two or three" players isn’t new.

Yu Kai has spoken repeatedly about industry consolidation, and executives from Momenta, SiEngine, Black Sesame Intelligent, and several automakers have offered similar forecasts.

What’s truly worth watching now is why the industry is finding it increasingly easy to reach this consensus.

The reason likely goes beyond just the difficulty of algorithms.

The real trouble begins only after the algorithms are up and running.

When asked how costs are squeezed to fit city NOA into 100,000-yuan vehicles, Yu Kai responded quickly: "Horizon never engages in that kind of low-quality cost-cutting."

He explained that the Journey 6M offers 128 TOPS not by simply slashing chip prices, but through software-hardware co-optimization. This allows it to achieve performance that previously required much higher computing power, all within a relatively limited compute and cost structure.

He then used a term with his own personal flair — "outvolution."

"Involution is homogeneous competition within a boundary, relying on meaningless low-quality price cuts; outvolution is about expanding boundaries through technology."

The term "outvolution" carries an obvious corporate bias, but the problem it addresses is real.

Once city NOA moves from the 300,000 and 200,000-yuan price brackets down to 100,000 yuan, competition can no longer focus solely on algorithm performance. Cost, computing efficiency, model adaptation, production validation, and — most critically — the ability to replicate at scale all become issues simultaneously.

Doing one project well doesn’t guarantee success across ten.

Different automakers have different chassis, braking, steering systems, sensor configurations, and development processes. If winning a new project requires adding a dedicated engineering team, then as scale grows, costs could rise in lockstep.

Yu Kai’s response to this was blunt: "Our real flexibility lies in how we empower customers: through technology IP licensing, not by throwing manpower at the problem."

He listed partners and cited the Volkswagen-CARIZON partnership to emphasize that massive engineering integration and project delivery aren’t handled solely by Horizon. "There is always a partner layer in between," he noted.

By his account, about 95% of shipments are fulfilled through partners. Horizon wants to be a "technology enabler," not an "engineering service provider" that relies on endlessly stacking engineers to complete projects.

This actually gets closer to the core of the next round of competition than the 15 million figure itself.

Technology must go deeper, but the business must become easier to replicate.

Without this, even with superior algorithms and more clients, a company risks becoming increasingly bloated.

In the platform war, the path matters more than ambition

That’s why Yu Kai repeatedly referenced two classic combinations from the tech industry that day: Wintel (Windows + Intel) and ARM + Android.

In his view, about 80% of automakers will ultimately rely heavily on specialized suppliers, while the top 20% will insist on vertical in-house development.

For the former, Horizon can directly supply chips and software; for the latter, it can offer chip and software IP licensing, allowing automakers to continue development on their own.

Yu Kai calls the first the Wintel model and the second the AA model. His ultimate goal is to be the "definer of computing platforms for the automotive era."

Horizon isn’t the only one chasing this goal.

NVIDIA has expanded from Orin to Thor, DriveOS, and Hyperion, turning its computing chips into a comprehensive development foundation. Qualcomm’s Snapdragon Ride Flex is also merging cockpit and ADAS onto a single SoC. Just this July, BMW selected Qualcomm as a key supplier for future digital cockpits and next-generation ADAS/autonomous driving systems, extending the partnership into the next decade.

Huawei is taking a different path.

Its capabilities span intelligent driving, cockpits, vehicle control, in-vehicle optics, and vehicle-cloud connectivity, offering a more complete technology stack than most suppliers. Yet, at Huawei’s Qiankun media day in July, Yinwang CEO Jin Yuzhi deliberately described the company’s role as an "electronic screw" in the automotive industry.

The point of that remark wasn’t modesty; it was about drawing boundaries.

While building deep technical capabilities, Huawei repeatedly emphasizes that its position remains that of a supplier of incremental parts for intelligent connected vehicles, not competing with automakers to "build cars."

This may look different from Horizon’s focus on IP licensing and partner ecosystems, but fundamentally, both are addressing the same issue:

Where exactly should a supplier draw the line?

If the technology is too shallow, automakers can replace the supplier at any time.

If the technology goes too deep, it may encroach on the core interests of automakers, Tier 1 suppliers, or even existing ecosystem partners.

The real challenge of platform competition isn’t about doing as much as possible, but having enough control while knowing what not to take.

Horizon hopes its licensing model allows automakers with in-house R&D to retain control over upper-layer development; Huawei uses its "no car manufacturing" stance to address automaker concerns about brand and vehicle definition; while NVIDIA and Qualcomm aim to anchor their positions in underlying computing, software, and development platforms.

Momenta, similarly, is working to reuse its algorithmic capabilities across different vehicle models and computing platforms.

It seems everyone is expanding their capabilities, yet everyone is also drawing lines for themselves.

This is the most subtle aspect of the next round of so-called "platform competition."

Everyone wants clients to depend on them, but no one wants clients to feel they might one day become the biggest competitor.

The table is shrinking, but the stakes are rising

As more top players begin fighting for platform dominance, another change is underway: the number of new players is shrinking.

These two trends are not in conflict; on the contrary, platform-based competition itself makes intelligent driving an increasingly expensive game.

Developing an algorithm set requires investment; developing a next-generation chip requires another. Toolchains, automotive-grade validation, mass production engineering, overseas R&D, and data loops all demand continuous funding. In the past, a startup could crash the table with a single algorithm or one client. To stay long-term, the fixed costs are becoming higher and higher.

Meanwhile, the hottest direction in capital markets has clearly shifted toward large models, data centers, and embodied intelligence, and the intelligent driving sector is seeing talent flow toward robotics.

"The intelligent driving industry has actually already highly converged."

In his view, an early-stage industry constantly spawns new companies, but later on, players naturally dwindle and talent flows to the next hot track. He describes the current state of intelligent driving as "still waters running deep."

Later, discussing the capital markets, he was even more direct.

Yu Kai admitted that capital and attention are currently being drawn to data centers and large models, while the automotive industry itself continues to endure fierce competition and profit pressure.

But then he said something interesting: "I think it’s actually a good thing not to be a hotspot."

The reason? "Without hot money coming in, no new teams will emerge."

This remark carries a distinct perspective from an industry leader.

For companies that haven’t found their footing, a cooling industry means harder financing, valuation pressure, and customers continuing to squeeze costs.

But for companies that already have scale, orders, and enough cash on hand to continue R&D, a less hyped industry means fewer new entrants.

So "consolidation" doesn’t necessarily happen at a single dramatic moment.

More commonly, the seats at the table just get more expensive.

New companies can’t get in, fringe players can’t hold on, and the survivors must keep betting more money on next-generation technologies.

Yu Kai is relieved the industry lacks hot money, yet he emphasizes that as long as cash flow is healthy, companies should launch a "saturation attack" — even aiming to be "the craziest investor in the industry."

It seems contradictory, but it fits the current reality of leading intelligent driving companies.

As the industry consolidates, competition doesn’t disappear.

It just shifts from "anyone can try" to "fewer and fewer can afford the cost of staying here long-term."

After the 15 million milestone, someone asked Yu Kai if he was happy.

"Yes, I was happy."

And then?

"I was happy for two seconds, and then I moved on."

Because, "the harder challenges are definitely still ahead."

That statement applies to the entire intelligent driving industry as well.

In the first half, algorithms were the easiest way to pull ahead. Whoever launched city NOA first, or got end-to-end AI into production cars first, had a chance to step forward.

Today, algorithms still determine whether a company qualifies to stay at the table.

But it is shifting from a bonus feature to an increasingly clear barrier to entry.

Once across that threshold, the competition becomes about chips, cost, toolchains, engineering, ecosystems, and customers — and whether you have enough money and patience to hold it all together over the long term.

Yu Kai predicts that in the end, only two or three may remain.

Whether that final number is two, three, or even more, no one knows yet.

But one thing is certain: the next round of intelligent driving is no longer a contest of algorithms alone.