Nvidia’s Physical AI Push Shows Why Self-Driving Cars Are Still Harder Than We Admit

Nvidia has unveiled a new technology platform aimed at pushing self-driving cars — and more broadly, “physical AI” embedded into machines — closer to real-world deployment. Jensen Huang framed it as a “ChatGPT moment for physical AI,” suggesting a step-change in how machines may learn to perceive and act in the physical world.

What makes the announcement notable is not that Nvidia is becoming a carmaker. It’s that Nvidia increasingly wants to be the platform layer underneath autonomous systems: the chips, software stacks, simulation tools, and models that other companies build on.

The announcement also pulled in an immediate, telling response from Elon Musk. Commenting after Nvidia’s new platform reveal, he wrote that it’s “easy to get to 99% and then super hard to solve the long tail of the distribution.” That “long tail” — rare but safety-critical scenarios — is one of the most consistent reasons autonomy has progressed in bursts, then stalled, for decades.

This is the real context behind Nvidia’s announcement: self-driving has always looked close in demos. The difficulty is getting it to work reliably, everywhere, for normal people, under real-world conditions.

Autonomous vehicles have been a human dream for longer than most people realise. As early as the 1920s, radio-controlled demonstration cars appeared — dramatic proof-of-concept stunts that also exposed the limits of the era. In one widely referenced 1925 demonstration, a radio-controlled car reportedly crashed into a sedan during a live event. The dream was visible early; so were the risks.

Mid-century efforts took a different path: rather than building intelligence into the vehicle, researchers tried to build it into the road. In the 1950s through 1970s, projects experimented with wire-guided or road-embedded systems that could steer vehicles along instrumented routes. These approaches could work under controlled conditions — but only where expensive infrastructure existed. They weren’t general solutions for ordinary streets.

By the late 1970s and 1980s, the direction shifted again toward self-contained robotic vehicles, using onboard sensors and computing. These systems were impressive for research, but too fragile for public roads. They struggled with weather, sensor noise, dense traffic, and unpredictable road layouts. They were also costly, and regulators had little basis for approving them. In short, the technology could “work,” but it couldn’t work consistently enough to be trusted.

The recurring reasons these systems didn’t become commercial products were structural:

• Dependence on special conditions (instrumented roads or controlled routes)
• Limited computing power for real-time perception and planning
• Weak generalisation: rules and sensors failed when reality deviated from expectations
• High cost of hardware and maintenance
• Safety and regulatory uncertainty: unclear accountability when things went wrong

This history matters because it explains today’s debate. Autonomy has advanced, but the core barrier hasn’t disappeared: the world is messy, and driving requires judgment under uncertainty.

For much of the 20th century, autonomous driving was constrained by two things: sensing and computation. Even when a system could “see,” it often couldn’t interpret the scene fast enough to react safely.

From the 1990s into the 2000s, advances in robotics and computer vision pushed autonomy from simple demonstrations to more realistic testing. The DARPA Grand Challenge era is often described as a turning point, because vehicles had to operate under more varied terrain and conditions than earlier lab systems.

But those systems were still heavily rule-driven. They relied on hand-coded logic, carefully tuned thresholds, and structured assumptions. They could look competent — until a scenario fell outside the designed rules.

After 2010, a more fundamental shift arrived: deep learning began to replace hand-built perception pipelines. Instead of coding what a pedestrian “looks like,” models learned patterns from massive datasets. At the same time, sensors got cheaper and more capable, and GPU-class computing moved closer to real-time use in vehicles.

This is the backbone of “physical AI”: systems that blend sensors, high-performance computing, and large models to interpret the physical world and act in it. It’s also why Nvidia is central: it has become the default supplier of the compute layer for many AI-heavy industries.

Tesla is widely seen as the first automaker to sell millions of vehicles globally with advanced driver-assistance marketed as Autopilot and Full Self-Driving (FSD). The important detail — often blurred in public conversation — is that these are generally classified as Level 2 systems: they can assist with steering and speed in many contexts, but the human driver remains responsible and must supervise.

Tesla’s AI role is not a single “autonomous mode.” It is a stack that tries to do three things well:

• Perception: detect lanes, vehicles, cyclists, pedestrians, signs, and signals
• Prediction: estimate what other road users may do next
• Planning/control: choose a safe path and execute steering, braking, acceleration

Tesla’s strategy is also shaped by scale. Instead of a small test fleet, Tesla has a large fleet of customer cars generating real-world data. Interventions — moments when the driver takes over — are especially valuable for training and iteration. This “fleet learning” loop is one of Tesla’s strongest structural advantages: it accelerates improvement compared with companies training on small, carefully curated datasets.

However, the challenges remain visible. Investigations, lawsuits, and high-profile incidents have repeatedly raised questions about how these systems are marketed and how drivers understand their limits. The mismatch between names like “Full Self-Driving” and supervised reality increases the risk of over-trust — a human factor problem as much as a technical one.

You asked for year-by-year sales and profits to understand the trend. Public reporting does not cleanly separate autonomy/FSD profit from the rest of Tesla’s business (FSD is not consistently broken out as a standalone annual profit line), but overall revenue and net income trends help contextualise Tesla’s scale during the period when Autopilot/FSD expanded.

Here is a small table inside the article (as you requested) to show the trend:

Year	Revenue (USD bn)	Net income (USD bn)
2016	~7.0	~–0.7
2017	~11.8	~–2.0
2018	~21.5	~–1.0
2019	~24.6	~–0.9
2020	~31.5	~0.7
2021	~53.8	~5.5
2022	~81.5	~12.6
2023	~96.7	~15.0

These numbers show Tesla’s business scaling rapidly — but they do not prove autonomy is solved. They show that Tesla managed to ship and sell a product that combines EV manufacturing with increasingly capable driver-assistance software.

On your request for “specific sales & profits of Tesla that a model reinforces AI”: Tesla does not report profits by “AI model” or publish a year-by-year breakdown isolating FSD profit as a clean line item in the way you’re asking. That’s a transparency limitation of public financial reporting, not a lack of relevance. What we can say responsibly is that Tesla monetises software features and packages (including FSD) alongside vehicle sales, but the exact profit contribution is not publicly itemised in a way that allows a precise annual “FSD profit series.”

A helpful way to understand today’s moment is to separate two eras:

Earlier “low-AI / rule-heavy” systems

These relied on deterministic logic and narrow automation:

• Lane keeping based on simple lane detection
• Adaptive cruise control using radar thresholds
• Hard-coded rules for following distance, speed changes, and basic steering constraints

They could be reliable in constrained scenarios, but they failed sharply when conditions changed:

• Unusual road markings
• Complex merges or negotiation with human drivers
• Construction zones
• Rare obstacles or poor weather

Modern “high-AI” systems

These rely far more on learned perception and prediction:

• Neural networks trained on large datasets
• Better object detection in crowded scenes
• Improved response to variable road geometry
• Faster iteration cycles from real-world data and simulation

This enables more ambitious features — highway navigation, automated lane changes, and limited city-street automation — but it introduces a new reality: the system can appear highly capable while still failing in rare edge cases. That’s the long-tail problem Musk referenced, and it’s why “99% competent” is not enough for unsupervised driving.

Driving isn’t a single task. It’s a constant stream of micro-judgments under uncertainty: interpreting intent, predicting behaviour, reacting safely, and making trade-offs quickly.

The “long tail” includes events that are statistically rare but safety-critical, such as:

• A pedestrian behaving unexpectedly
• A partially obscured traffic signal
• Debris that looks like a shadow
• An emergency vehicle approaching from a confusing angle
• Construction that contradicts lane markings

These cases are hard because they don’t repeat often enough to be “solved” by simple scaling, and because failure carries high consequences. This is why genuinely driverless systems tend to begin in constrained areas (geofenced robotaxis, fixed trucking routes) rather than open-ended global deployment.

Tesla’s approach is vertically integrated: vehicles, software, and increasingly custom compute are built to serve Tesla’s own fleet strategy.

Nvidia’s speciality is different: it builds high-performance AI compute and developer platforms that many companies adopt. That matters because autonomy is increasingly a compute + data + simulation problem, not only a vehicle-manufacturing problem.

Nvidia’s platform strategy is to offer:

• Training infrastructure (data-center scale)
• In-vehicle inference hardware
• Simulation and synthetic data generation
• Foundation-model style approaches that can “reason” through scenarios

In plain language: Nvidia wants to supply the AI “operating system” for physical machines — cars, robots, factories — and let others compete on branding, manufacturing, and services.

This is how Nvidia could “overpower” Tesla in influence without competing in consumer vehicle sales: by becoming the default autonomy platform adopted across multiple automakers.

Tesla faces challenges on several fronts:

Regulatory and legal pressure
Investigations and lawsuits have raised persistent questions about crashes, system limits, and marketing language. Even when systems improve, legal scrutiny can slow deployment or constrain how features are described.

Technical reality
Reaching reliable unsupervised autonomy is hard regardless of approach. If a system is safe in 99% of cases but fails unpredictably in rare scenarios, it is still not safe enough for Level 4/5 freedom.

Human factors and trust
A system that works well most of the time can create driver complacency. Misunderstanding the limits of assisted-driving features can add risk even as other risks decrease.

Competitive pressure
Other autonomy companies may progress faster in narrow domains (fixed routes, geofenced areas) even if they lack Tesla’s global scale.

For most people, the near-term impact will be gradual rather than dramatic. The most visible benefits of “physical AI” in cars may arrive as quieter improvements: smoother lane keeping, better collision avoidance, more consistent behaviour in routine traffic, and more useful safety alerts.

The risk, however, is that capabilities improve faster than understanding. If a system looks confident, people may assume it is autonomous even when it is not. So public impact is two-sided: better assistance, but higher stakes for communication, design, and driver expectation management.

For professionals — automakers, fleet operators, regulators, insurers — the shift is broader. Autonomy is increasingly shaped by simulation, data governance, and platform ecosystems. The winners may be those who can build not only better models, but better validation, monitoring, and accountability frameworks.

Nvidia’s direction strengthens a “platform” view of autonomy: multiple companies building on a shared compute and simulation backbone. That could reduce duplication of effort across the industry, and it may accelerate the adoption of stronger training and testing standards.

At the same time, platform dominance raises questions:

• Who sets the safety assumptions inside the models?
• How do regulators audit systems built from third-party stacks?
• How do we ensure transparency when systems become more complex?

These aren’t reasons to slow innovation — they’re reasons to design governance in parallel with capability.

If you zoom out, Nvidia’s message is that autonomy is evolving from a feature race to an infrastructure race. The next phase may look less like a single “robotaxi moment” and more like layered deployment:

• Driverless operation first in controlled or restricted environments
• More autonomy in logistics and fixed-route trucking
• Increasingly capable assistance features in consumer cars
• Greater reliance on simulation and synthetic data to pressure-test edge cases

Progress may still be uneven — breakthroughs in some contexts, stubborn limits in others — because the world doesn’t standardise itself for models.

The most realistic signal is not “driverless everywhere soon,” but “AI increasingly embedded in the systems that move people and goods,” with autonomy advancing where it can be validated safely.

Progress in autonomous vehicles has rarely been limited by ambition; it has been constrained by how unpredictable the physical world actually is.

Tesla proved that AI-assisted driving can scale to millions of cars. Nvidia is betting that the next leap will come from treating autonomy as shared infrastructure — models, simulation, and compute that many companies can build on — rather than a single company’s breakthrough.

The companies most likely to shape this future may not be the ones promising the fastest transformation, but those building systems designed to learn cautiously, explain their decisions, and improve without assuming the world will simplify itself.

The real story is not who “wins autonomy” first. It’s who manages the last 1% responsibly.

Primary source:

• BBC News — Nvidia unveils self-driving car tech as part of physical AI push
  https://www.bbc.com/news/articles/c0jv1vd571wo

Secondary sources (as provided in your research list):

• Wikipedia — History of self-driving cars
  https://en.wikipedia.org/wiki/History_of_self-driving_cars
• Computational and Experimental Systems journal article — Deep learning for autonomous driving systems
  https://www.oaepublish.com/articles/ces.2024.83
• A Standard Driven Software Architecture for Fully Autonomous Vehicles (research PDF)
  https://jstvssr.github.io/assets/pdf/standard-driven-software-architecture.pdf
• ST-P3: End-to-end Vision-based Autonomous Driving (ECCV paper PDF)
  https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136980522.pdf
• Evolution of Decision-making Systems in Autonomous Vehicles (Journal of Systems and Control)
  http://www.coscipress.com/journal/JSC/article/1a6ffbc57552fbd990e4ef86a9dd2e76
• Software System of Autonomous Vehicles: Architecture and Design (Wharton research thesis PDF)
  https://fisher.wharton.upenn.edu/wp-content/uploads/2020/09/Thesis_Nova-Qiaochu-Guo.pdf
• Vision-Based End-to-End Planning for Autonomous Driving (Stanford CS231n project report PDF)
  https://cs231n.stanford.edu/2025/papers/text_file_840589978-final_report.pdf
• Deep learning’s impact on autonomous vehicle evolution (UBT conference paper)
  https://knowledgecenter.ubt-uni.net/cgi/viewcontent.cgi?article=4512&context=conference
• Sequoia Capital — How End-to-End Learning Created Autonomous Driving 2.0
  https://sequoiacap.com/podcast/how-end-to-end-learning-created-autonomous-driving-2-0-wayve-ceo-alex-kendall/
• arXiv — Evolution of Artificial Intelligence and Learning Algorithms
  https://arxiv.org/html/2402.17690v2