When Nvidia’s market cap soared past USD 4 trillion, making it the most valuable company in the world, CEO Jensen Huang was in China, warmly praising the country’s homegrown automakers. A video montage of Huang name-checking nearly every major Chinese car brand went viral, with one netizen quipping that now that Nvidia is “invincible,” Huang’s gaze has locked firmly on customers.
But even in the glare of that success, cracks are showing. Many of the automakers Huang applauded are now actively seeking to reduce their dependence on Nvidia.
One of those companies is General Motors. After an internal review of Nvidia’s assisted driving stack, a GM executive reportedly described the system as “very scary,” a verdict that sharply contrasted with the public fanfare surrounding a new partnership announced by Huang in March, where Nvidia would help GM develop autonomous fleets.
Huang had also touted Nvidia’s expanding automotive ambitions, citing collaborations with Toyota, Mercedes-Benz, and others, and projecting USD 5 billion in revenue from the division by 2026. But GM’s internal findings, relayed directly to Nvidia’s autonomous driving team and reportedly briefed to Huang, have dampened enthusiasm within the automaker.
This isn’t the first time Nvidia’s automotive push has faced scrutiny. Months before GM’s review, Mercedes-Benz had reached a similar conclusion. In mid-2023, CEO Ola Kallenius and his VP of technology conducted an extensive intercity test drive in the US, comparing assisted driving systems powered by Nvidia and Chinese startup Momenta.
To their surprise, even on Nvidia’s home turf, Momenta’s system outperformed it, and the outcome raised eyebrows in Stuttgart and beyond. GM took note. Sources told 36Kr the company began reevaluating its reliance on Nvidia after realizing how quickly Momenta was catching up, and in some cases, pulling ahead.
After learning about this, Nvidia’s automotive head Wu Xinzhou voiced his dissatisfaction internally. It didn’t help that Momenta’s software, which upstaged Nvidia on its own soil, had been fine-tuned in under a month.
According to 36Kr, Mercedes-Benz has now shifted its assisted driving contracts for several China-based models from Nvidia to Momenta. Another Nvidia software customer, Jaguar Land Rover, is also seeking new partners. “Nvidia employees in China have basically stopped engaging directly with automaker projects,” one source said.
In China’s fiercely competitive automotive market, time is a luxury Nvidia no longer gets.
To be fair, Nvidia’s automotive software business is a small slice of its empire. Even including its in-vehicle computing chips, from Xavier to Orin and Thor, automotive revenue accounts for less than 2% of Nvidia’s last disclosed annual earnings.
Even if it excels, the upside is marginal. For comparison, Huawei’s business unit for intelligent automotive solutions made RMB 26.4 billion (USD 3.7 billion) in total revenue last year. Nvidia earns that in roughly ten days.
So why does this tiny business matter?
Because if you believe that artificial intelligence will one day inhabit physical machines, cars are the best prototype. They are essentially robots without arms, and their technological roadmap closely aligns with that of embodied intelligence.
With this vision in mind, Nvidia merged its automotive and robotics divisions over a year ago. Huang has said he believes companies like Xiaomi and BYD will eventually build excellent robots.
Among tech insiders, smart cars are often cited as the most practical early application of embodied AI. The automotive sector already has a well-established industrial ecosystem and clearly defined use cases. One founder even suggested that if autonomous driving fails to reach full maturity in the next few years, embodied AI could prove to be little more than a bubble.
Embracing autonomous driving is, in effect, embracing AI’s entry into the physical world.
And technology moves fast. Remember, Nvidia’s meteoric rise was fueled by the AI boom. Just two and a half years ago, before ChatGPT shocked the world, Nvidia’s market cap was less than a tenth of what it is today.
That’s why this seemingly marginal business is actually a very big deal.
Unfortunately for Nvidia, while carmakers once fought to be first in line for its latest chips, that’s no longer the case. With the launch of the Thor chip, Nvidia is now at risk of losing major customers in China.
The threat doesn’t come just from bigwigs like Huawei and Momenta. China’s electric vehicle makers are now developing their own AI chips, following Tesla’s lead. Nio and Xpeng have already started deploying their chips in vehicles. Starting 2026, Li Auto plans to mass produce its own chips. Xiaomi CEO Lei Jun has also said that an in-house automotive chip is on the way.
Of course, chip development is notoriously difficult. That steep learning curve applies to Chinese carmakers and solution providers alike.
Nvidia’s Thor chip delays spark turmoil
Late last year, Li Auto’s suppliers were notified that its revamped L series, initially slated for release in March, would be postponed to May. Component deliveries were adjusted accordingly.
According to a core supplier, the delay was due to the late delivery of Nvidia’s Thor chip. Li Auto is one of Nvidia’s key automotive customers and was an early adopter of its Orin chip.
The Thor chip is the next-generation upgrade, offering up to 700 TOPS (trillion operations per second) of compute power. Li Auto’s 2025 L series models were meant to debut with this chip, which supports the company’s new vision-language-action (VLA) driving model.
But this wasn’t Thor’s first delay. Originally scheduled for mass production by the end of 2024, the chip’s delivery has been pushed back at least three times.
These delays come at a steep cost. Based on the sales gap before and after the L series refresh, Li Auto missed out on over 10,000 units, equivalent to RMB 6 billion (USD 840 million) in revenue, due to the launch being pushed from March to May.
Xpeng was the first to sniff out trouble. According to a company engineer, as of mid-2023, Xpeng still planned to prioritize Thor, with its in-house Turing chip merely serving as a backup.
But by early 2024, seeing signs of further delays, the company ditched its Thor roadmap and redirected resources to fast-track its own chip. Xpeng’s Turing chip is now shipping in its new G7 vehicles.
Carmakers had feared their own chips weren’t mature enough. But compared to the painful process of integrating Thor, many are now reconsidering.
One engineer described the ordeal of adapting to Thor:
“Even basic thermal control didn’t meet automotive requirements, and Nvidia no longer guarantees the 700 TOPS it advertised.”
After multiple iterations, the chip finally reached mass production grade. But actual performance reportedly hovers around 500 TOPS, well short of its original promise. Li Auto wanted to run a four-billion-parameter VLA model on it, but the reduced computing power makes that much harder.
Sources say Li Auto has since accelerated its chip program and plans to begin deliveries in Q1 next year, months ahead of schedule.
As more in-house chips get deployed, Nvidia’s share will likely shrink, several senior executives told 36Kr. Eventually, Nvidia chips might only be needed for export models.
Thor’s persistent delays have unintentionally given automakers the final push they needed to commit to internal chip programs.
RELATED ARTICLE
Years of grinding, now carmakers are breaking through
Developing a chip is a gamble for any automaker. While car development takes about 18 months, building a chip takes closer to four years, according to timelines followed by Nio, Li Auto, and Xpeng.
But with escalating geopolitical tensions and mounting concerns about tech restrictions, the fear of supply cuts hangs like a Damoclean sword over the industry. For the past four years, Chinese carmakers have been grinding through the pain.
There were plenty of missteps. Licensing costs for IP alone were a burden. As one executive put it, “Every chip you sell, you’re paying someone.” EDA (electronic design automation) tools, essential for chip design, are oligopolized by few key players, and negotiating with each is difficult.
Xpeng CEO He Xiaopeng once shared that the Turing chip underwent major redesigns. The company even had to pay a sizable penalty to its early partner, identified by 36Kr as Marvell.
Marvell serves as a production gateway to TSMC. As a key TSMC client, it offers frontend and backend services for automotive chips. Initially, Xpeng aimed for best-in-class specs. But costs ballooned, making the plan untenable.
Also, Marvell lacked experience in high-performance automotive chip design. The two companies parted ways amicably, but the compensation exceeded USD 100 million. Xpeng later partnered with Socionext instead.
“It took He Xiaopeng’s persistence to push it through,” one source said. “Without him, the project wouldn’t have survived.”
Xpeng began designing its chip in 2021, just as transformer models were gaining traction in Silicon Valley. Guided by its US-based team, Xpeng embedded key operators into the chip. But even now, foundational support for large models remains incomplete.
Nio faced its own white knuckle moment. CEO William Li recalled that in 2023, just as frontend chip design was nearing completion, a key design partner suddenly exited the Chinese market. That partner was also Marvell.
With backend design in jeopardy, Nio had to assemble its own internal team and request access from TSMC to proceed.
Today, Nio’s chip team has over 600 people, mirroring the setup of a full-fledged semiconductor company. It spans everything from frontend and backend design to testing.
Automakers understand vehicles better than Nvidia does. Some engineers even say that Nio’s Shenji chip architecture is more rational than Nvidia’s Thor, and that Nvidia’s dominance in autonomous driving is overstated.
Thor, too, is Nvidia’s first foray into chips offering over 1,000 TOPS of compute. Even TSMC engineers reportedly discovered design flaws in the Blackwell-based die that connects dual GPUs, and these are flaws that could affect yield rates. Huang has publicly acknowledged this.
Nio, Li Auto, and Xpeng have all delivered their first-generation chips, with costs reportedly USD 300–400 million each. More spending is coming, as companies prepare for their second generation.
Why take on such an enormous task?
Cost reduction is one reason. Nio’s Li said that the Shenji chip can shave RMB 10,000 (USD 1,400) off a vehicle’s production cost.
But more importantly, tighter integration between algorithms and chips gives automakers long-term control. One Xpeng staffer noted that the company’s AI stack is now fully tailored to the Turing chip, including its upcoming foundational model.
Xpeng CEO He privately told colleagues, “Only after we built our own chip did we start seeing all the hidden benefits.” Since Xpeng follows a vision-based approach, it has been able to embed dual image signal processors (ISPs) into the Turing chip to enhance perception in challenging conditions such as night, rain, and glare.
Li Auto is betting on deploying large AI models inside vehicles. But its tech leads also said that even Nvidia sometimes overlooks key needs. For instance, bandwidth limits on Thor’s memory or neural processing modules can cause latency during execution.
“These problems only emerge during deployment,” one engineer said. “When you own the chip, you can fix them faster.”
Tesla’s lead in deploying a three-billion-parameter model came from its in-house chip. That’s the bar.
China’s EV firms are also racing to build high-level assisted driving software. Understanding the software is what enables better chips.
The next frontier is already in sight. Tesla’s AI5 chip is now in production, reportedly boasting 2,000–2,500 TOPS. Elon Musk has also hinted at a new model with 4.5 times more parameters than the current one.
China’s top EV makers have made AI their main strategy. Building their own chips is the most difficult but most necessary step in that journey.
And once you start building chips, it’s hard to stop.
Nvidia doesn’t operate on automakers’ timelines
In the carmaking world, timely delivery is everything. Last year, when Nio’s sub-brand struggled with battery shortages, even CATL, the dominant battery supplier, ramped up output a month ahead of schedule.
But Nvidia doesn’t play by those rules. In the GPU market, it’s the trendsetter. Everyone else syncs their roadmaps to fit it.
Its automotive chips follow the same pattern.
Thor is built on Blackwell, Nvidia’s next-gen AI architecture, using TSMC’s N4P process. This allows for denser transistors and lower power consumption. But that’s also where the problems begin.
N4P is designed for consumer electronics, not cars. TSMC’s automotive-grade four-nanometer process will only be ready this year.
Automotive-grade standards are stricter across the board, not just at TSMC’s foundries, but also in packaging and testing. “Testing for automotive chips costs three times as much because it’s done three times over,” one chipmaking veteran said.
In general, automotive-grade chips lag their consumer counterparts by two years. The more advanced the node, the longer the gap.
And because automotive volumes are lower, foundries naturally prioritize consumer chips. These timelines are structural, not negotiable.
A single delay can cost automakers heavily. In any other supplier relationship, that would trigger a crisis review. But inside Nvidia, there’s barely a ripple. Why? Because Nvidia isn’t a car supplier at heart. Automotives make up only a small chunk of its gargantuan business.
To be clear, Nvidia is trying. Its teams reportedly worked through Christmas, but its focus is on solving technical challenges, not meeting carmakers’ production deadlines.
Had Nvidia prioritized delivery timelines, Thor could have been built on a more mature node. After all, cars prize stability over top-tier specs.
Then there’s resource allocation. Multiple engineers told 36Kr that when they encountered issues using the Thor chip, Nvidia’s support was lacking. “Some defects were patched at the vehicle controller level—by us,” one said.
Even Huang doesn’t check in on the automotive business much. That’s a tough mismatch for carmakers used to assertive, attentive suppliers.
What Nvidia lost was a window of opportunity in China
When it comes to autonomous driving software, Nvidia’s hardware DNA has repeatedly clashed with the strengths of startups that started off in software development, like Momenta.
In February 2024, Nvidia’s automotive head Wu flew to Shanghai with several vice presidents and senior directors for a six-week development sprint. But even after the blitz, Nvidia’s assisted driving experience still lagged behind Momenta’s.
According to a company source, Mercedes-Benz had requested a demo of urban navigate-on-autopilot (NOA) functionality in Shanghai. During the test, Momenta’s system required nearly zero driver intervention. Nvidia’s? It reportedly braked abruptly, accelerated unpredictably, falling short of mimicking human driving.
Wu, a former head of Xpeng’s assisted driving division, is known for his relentless execution. At Nvidia, he kept his routine of testing in cars daily. But even he hasn’t managed to close the gap.
Corporate culture may be the biggest obstacle. While Wu built a 200-person team in China, 80% of Nvidia’s autonomous driving unit—over 2,000 people—remains US-based. “The China team has no real authority,” one source claimed. “Even for urgent edge cases, decisions all come from the US. People joke that the China team is a puppet.”
By contrast, China’s top players build software under intense pressure, often in closed-loop development cycles, with quick feedback and faster turnaround for clients.
At Nvidia, no one’s getting fired. A 36Kr source said that any employee with more than three years at the company is likely already a multimillionaire, assuming equity wasn’t sold too early. “People don’t feel the need to hustle anymore,” the source added.
That makes it harder to cater to carmakers.
In one project meeting with Mercedes-Benz, a US-based Nvidia staffer reportedly slammed the table and said, “Remember, we’re strategic partners. We’re equals. This isn’t a client-vendor relationship.”
China’s software startups couldn’t afford that tone. Startups like Momenta, QCraft, and DeepRoute.ai are still fighting for survival.
“Two to three out of every six months are spent in full-throttle mode,” one Momenta employee said. That may not suit every employee’s lifestyle, but it does help keep the company alive in the assisted driving race.
Nvidia, for its part, is trying to inject new life into its workforce. Earlier this year, Huang hired Kristin Major from HP as senior vice president, widely seen as a move to rekindle morale internally.
At Nvidia’s GTC conference in Paris this June, Huang said everything that moves will be robotic in the near future. Accordingly, the next frontier will be cars.
That prediction is dead on. According to insiders, Qualcomm’s automotive revenue has grown from just 1.2% of its total two years ago to nearly 10% today. The company is rapidly adapting its latest technologies for automotive-grade chips.
But Nvidia’s tightly integrated CUDA and NVLink ecosystems, which form its moat in the GPU world, don’t translate as well to the automotive sector. Several carmakers told 36Kr they are unsure whether Nvidia will continue investing in this space, or eventually pull out.
Robotics may be the long game, but the first real challenge is already here: the smart car.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Li Qin and Li Anqi for 36Kr.