Nvidia Automotive Compute: The Internal Battle for Resources
Here's the thing: when your own head of automotive is fighting for compute resources inside Nvidia, you know there's a problem. Wu Xinzhou's reported dissatisfaction isn't just a whisper; it's a flashing red light. The mainstream narrative talks about Nvidia's challenges in China, the carmakers ditching them, and the Thor chip delays. But that's just the surface. The real issue is a fundamental misalignment of priorities and a company culture that can't pivot fast enough, especially when it comes to Nvidia automotive compute.
Maintainers are getting drowned in garbage PRs, and we need to stop pretending AI is a magic wand for open source. (I've seen PRs this week that literally don't compile because the bot hallucinated a library). This isn't about some abstract market shift; it's about engineering teams getting starved of the very thing their company sells: compute. This internal struggle directly impacts the delivery and performance of Nvidia's automotive compute solutions.
The Orin Era: A False Sense of Security
For a while, Nvidia's DRIVE Orin chip was everywhere in Chinese EVs. From 2022 to 2025, it was the go-to for ADAS, powering over a million BYD vehicles, along with Nio, Xpeng, and Li Auto. Orin was solid for AI inference, hitting 254 TOPS, and it had the ASIL-D safety cert. It felt like Nvidia had this market locked down for its automotive compute offerings.
But the automotive division, even with its projected $5 billion revenue by 2026, still makes up less than 2% of Nvidia's total earnings. That's a rounding error when you're talking about a company that's printing money in data centers. This tiny slice of the pie means the automotive team, despite having 200 people in China, lacks real authority. Decisions come from the 2,000-strong US team, which focuses on technical challenges, not carmaker production deadlines. Jensen Huang isn't checking in on this business often. That's a recipe for disaster for Nvidia's automotive compute ambitions.
Thor's Hammer Fell Flat
The next big thing was Thor, the upgrade to Orin. Advertised at 700 TOPS, built on Blackwell, using TSMC's N4P process. Sounds great on paper, right? The reality? A mess for Nvidia automotive compute.
Thor's mass production, initially slated for late 2024, has been delayed at least three times. Li Auto's 2025 L series models, which were supposed to launch with Thor in March 2026, got pushed to May. Why? Thermal control didn't meet automotive requirements. Nvidia no longer guarantees 700 TOPS; it's closer to 500 TOPS. And there are reported design flaws in the Blackwell-based die, affecting yield rates. Jensen Huang even acknowledged issues with the Blackwell architecture, impacting the promised performance of Nvidia automotive compute.
Here's the kicker: the N4P process is for consumer electronics. Automotive-grade 4nm from TSMC isn't expected until 2026, and those chips usually lag consumer versions by two years. Nvidia tried to rush a consumer-grade process into an automotive product, and it broke. That's a classic engineering misstep, prioritizing speed over the specific, brutal requirements of the automotive world for reliable compute.
Mercedes-Benz saw this firsthand. In a mid-2023 intercity test drive, Momenta outperformed Nvidia. Then, in a February 2024 Shanghai urban NOA demo, Momenta needed almost zero intervention, while Nvidia's system braked abruptly and accelerated unpredictably. General Motors executives reportedly called Nvidia's assisted driving stack "very scary." When your partners are saying that, you've got a serious problem with your automotive compute solution.
The Great Exodus: Carmakers Go In-House
About Thor's delays is about control. Carmakers are tired of waiting, tired of the "strategic partners, not client-vendor" line when support is lacking. They're building their own chips, moving away from relying solely on external providers for their automotive compute needs.
- Li Auto was an early Orin adopter but accelerated its in-house chip program. They plan to deliver their own chips in Q1 2027, months ahead of schedule, because Thor's reduced compute makes their vision-language-action (VLA) model harder to run.
- Nio has Shenji, an in-house chip that engineers describe as more rational than Thor. It can cut vehicle production costs by $1,400. They built a 600-person chip team and secured TSMC access after a key partner exited China.
- Xpeng ditched Thor in early 2024 and fast-tracked its Turing chip, which is now shipping in new G7 vehicles. They started design in 2021, paid over $100 million in penalties for redesigns, but they own their silicon now.
- Tesla has been doing this since 2019 with its AI chips, and their AI5 chip is reportedly hitting 2,000–2,500 TOPS. Waymo, Cruise (pre-shutdown), and even Xiaomi are all developing custom silicon.
This trend is a direct response to the frustration. Carmakers want cost reduction, tighter algorithm-chip integration, faster problem resolution, and long-term control. Geopolitical concerns in China also play a part, with the government explicitly discouraging Chinese OEMs from buying Nvidia chips in September 2025. This shift profoundly impacts the market for Nvidia automotive compute.
Social discussions, especially on Reddit, get this. People understand the immense compute, memory bandwidth, and sensor needs for autonomous driving. They also get the unique challenges of power consumption and thermal management in a car. The move to custom silicon isn't a surprise; it's a logical engineering response to a vendor that can't deliver consistent automotive compute solutions. This trend is further highlighted by recent reports from industry analysts detailing the acceleration of in-house chip development.
The Future: Niche Player or Irrelevant?
Nvidia's automotive division is stuck. It's too small to command the internal resources it needs, and its core product development is lagging. The company's "We're strategic partners" stance rings hollow when partners are forced to patch defects at the vehicle controller level, undermining the value of their automotive compute offerings.
The industry is moving towards vertical integration. Carmakers are becoming chip designers, just like Tesla showed them. Nvidia will still have a place, likely in niche areas or as a component supplier for those who can't afford the four-year, $300-400 million investment in custom silicon. Aurora, for example, is still committed to DRIVE Thor for its 2027 commercial trucking deployment. But for the broader passenger vehicle market, especially in China, Nvidia is losing its grip on the automotive compute landscape.
The days of a single vendor dominating automotive compute are over. The market is fragmenting, and the future belongs to those who control their own stack, from the algorithms down to the silicon. Nvidia's internal compute war is a symptom of a much larger problem: they're not built for the automotive industry's pace or its demands. They're built for data centers, and that's where their focus, and their future, remains, leaving their automotive compute division in a precarious position.
