Latest / Elon Musk Podcast / Nvidia is coming for Tesla

Transcript

Nvidia is coming for Tesla · Elon Musk Podcast

• 0:00
  NVIDIA just released an open source AI model for self driving
• 0:03
  cars. At CES 2026, Jensen Wang
• 0:06
  unveiled Alpa Mayo, a 10 billion parameter reasoning model that
• 0:10
  can explain why it makes every driving decision and the first
• 0:14
  production vehicle to use it will be the 2025 Mercedes-Benz
• 0:17
  CLA, launching in the United States in the first quarter of
• 0:21
  2026. It's a real You can buy it in a
• 0:23
  few months with Nvidia's full self driving stack built in.
• 0:28
  Jensen Wong called this the ChatGPT moment for physical AI,
• 0:32
  the point where machines begin to understand, reason and act in
• 0:35
  the real world. And the question is whether
• 0:36
  NVIDIA just leapfrog Tesla in the race to autonomous driving
• 0:40
  by giving every other automaker access to technology that Tesla
• 0:44
  has spent a decade developing in secret.
• 0:47
  Now, the technical specs are pretty good.
• 0:49
  Nvidia's platform runs on two Drive 4 chips delivering over
• 0:53
  2000 teraflops of compute power, roughly 1000 trillion operations
• 0:58
  per second. Tesla's Hardware 4 is believed
• 1:01
  to be in the few 100 tops range and mobilize upcoming a key
• 1:05
  ultra is rated at 176 tops. Nvidia's offering five to six
• 1:09
  times more processing power than anyone else on the market.
• 1:13
  They up IO model weights are available on hugging right now.
• 1:16
  You can check them out yourself and the simulation framework is
• 1:19
  on GitHub. NVIDIA released 1700 hours of
• 1:21
  real driving data for anyone to use.
• 1:24
  And in this episode we're going to be talking about what the
• 1:26
  demo looked like, who's partnering with NVIDIA and why
• 1:29
  this open source approach could change everything.
• 1:33
  And we're right into that Right after this very short break,
• 1:36
  NVIDIA demonstrated the system live in San Francisco.
• 1:39
  The transportation editor from The Verge got a ride in a
• 1:42
  Mercedes-Benz CLA equipped with Nvidia's Level 2 Plus system. 40
• 1:46
  minutes to the city traffic and the route included delivery
• 1:49
  trucks, cyclists, pedestrians, four way stops, traffic lights,
• 1:53
  double parked cars and unprotected left turns.
• 1:55
  The car handled all of it without disengagements and
• 1:58
  according to this, there were no crashes and no hiccups during
• 2:01
  the entire ride. They said that NVIDIA showed
• 2:04
  would go toe to toe with Tesla's Full Self driving under the most
• 2:07
  complex circumstances. That's a significant claim right
• 2:11
  now. Now, the demo showed specific
• 2:13
  behaviors that suggests the AI is reasoning rather than just
• 2:16
  pattern matching. At one point, the vehicle
• 2:18
  approached an intersection blocked by a truck.
• 2:20
  The NVIDIA system for a slow to let pedestrians cross then
• 2:23
  executed a wide maneuver around the obstacle.
• 2:25
  That kind of nuanced decision making, combining caution and
• 2:29
  assertiveness the way a human driver would, is what the Epamio
• 2:32
  model is designed to produce. The AI uses chain of thought
• 2:35
  reasoning, which means it generates a step by step
• 2:38
  explanation of its logic before executing each action.
• 2:41
  That internal monologue is not just for debugging them.
• 2:43
  It's designed to help the system handle rare edge cases by
• 2:46
  explicitly reasoning about cause and effect.
• 2:49
  Tesla's FSD does not do this. Tesla's neural networks are
• 2:54
  black boxes that produce outputs without explanations.
• 2:57
  The search strategy is also different from Tesla's.
• 2:59
  Nvidia's platform supports cameras, radar, lidar and
• 3:02
  ultrasonics, and a full 360° array.
• 3:04
  Tesla famously removed radar from his vehicles in 2021, has
• 3:07
  never used lidar, relying entirely on cameras for full
• 3:10
  self driving. Nvidia's arguing that redundancy
• 3:12
  is necessary for safety, and the demo vehicle used a combination
• 3:16
  of cameras and radar to perceive its surroundings.
• 3:18
  Now this approach is safer and more robust than camera only
• 3:21
  systems, especially in poor visibility or when objects are
• 3:25
  occluded. NVIDIA has partnered with sensor
• 3:27
  suppliers like Eva, High C and Sony and RB to create a
• 3:32
  validated ecosystem of hardware that works with its software
• 3:35
  stack, and automakers can mix and match components without
• 3:38
  worrying about compatibility. There's no walled garden for
• 3:41
  this thing. Now.
• 3:42
  I've been digging through the analytics of the show.
• 3:44
  I've noticed 37% of you are following this right now.
• 3:47
  And for you, I'm forever grateful.
• 3:48
  Thank you so much for joining this community.
• 3:50
  The other 63% of you haven't hit the follower subscribe button.
• 3:53
  And I'm going to tell you one thing that I think is really
• 3:56
  important. Supporting independent
• 3:58
  journalists is important in this day and age because we used to
• 4:01
  have people that would block everything.
• 4:04
  We had the gatekeepers. Independent journalists like
• 4:06
  myself, I've been doing this for, I don't know, 6-7 years and
• 4:10
  covering tech, Elon Musk, SpaceX, Space Flight, anything
• 4:14
  like that, like any sort of tech.
• 4:15
  I've been doing that for the last 6-7 years and I'm going to
• 4:18
  continue doing it for the next 10 years.
• 4:20
  And all I need from you is literally one second of your
• 4:23
  time, which is going to help the show tremendously.
• 4:26
  Hit the subscribe or follow button on whatever podcast
• 4:28
  platform you're listening or watching on right now.
• 4:30
  That's going to help me out tremendously and help out the
• 4:32
  show and continue the growth of the show.
• 4:34
  We've grown so much because of you and I can't wait to go to
• 4:37
  this next part of the journey because of you.
• 4:40
  Now. The partnership list is very
• 4:42
  impressive. Mercedes-Benz is first launching
• 4:44
  the CLA with what they are calling MB Drive Assist Pro
• 4:48
  capable of point to point city navigation under driver
• 4:51
  supervision, and Huang said the CLA is safest car in the world,
• 4:55
  sending its Euro NCAP 5 star rating.
• 4:58
  Jaguar Land Rover has a multi year deal to use Nvidia's
• 5:00
  platform for level 4 autonomy and future vehicles.
• 5:03
  Lucid is in granting the full stack into upcoming models.
• 5:07
  Uber is back in the autonomous vehicle game, working with
• 5:10
  NVIDIA to potentially deploy robo taxis or automated delivery
• 5:13
  vehicles. The Tier 1 suppliers are on
• 5:15
  board too. Bosch, ZF, Magna, and Quanta are
• 5:18
  all building electronic controls and units around Nvidia's
• 5:21
  architecture. That means car companies can buy
• 5:24
  pre integrated hardware from their existing suppliers and
• 5:27
  just drop in Nvidia's system without building everything from
• 5:30
  scratch. I'm an open source advocate.
• 5:32
  I love that Nvidia's doing this and the open source approach is
• 5:35
  what makes this different from everything else in the market.
• 5:38
  Tesla's full self driving software is entirely
• 5:40
  proprietary. You can't buy it, you can't
• 5:42
  license it. They're by giving it away.
• 5:44
  You can't study it, we don't know how it works.
• 5:47
  Mobilized technology is also proprietary.
• 5:49
  Waymo and Cruz develop full self driving stacks, but they use
• 5:51
  them exclusively for their own robotaxi fleets.
• 5:54
  NVIDIA is the first company to release a production grade
• 5:57
  autonomous driving. EI is open source.
• 5:59
  This is a clear bid to become the kind of like Android of
• 6:03
  autonomy while Tesla continues to keep its FSD stack completely
• 6:07
  closed, similar to Apple. And the implication is that any
• 6:10
  automaker without Tesla's decade of in house AI development can
• 6:14
  now get equivalent technology right off the shelf.
• 6:17
  It's free, you get it, all you have to do is buy the parts and
• 6:21
  build this software around it and you're good to go.
• 6:23
  Now if Mercedes actually ships a car in Q12026 that has similar
• 6:28
  capabilities to Tesla's FSD, it is based on an open source
• 6:31
  system any automaker can buy. It could commoditize level 2
• 6:36
  plus autonomous systems. Advanced driver assistance would
• 6:39
  no longer be the secret sauce of a few companies, would be a
• 6:42
  feature any car maker can implement with Nvidia's help.
• 6:45
  That puts pressure on Tesla, which has never faced an equal
• 6:48
  capable rival system in production vehicles.
• 6:51
  It also challenges Mobileye, whose business model depends on
• 6:54
  being the dominant ADAS supplier.
• 6:56
  Qualcomm announced new AI features for cars at CES, but
• 6:59
  Snapdragon Ride platform is still significantly behind
• 7:02
  NVIDIA in raw performance and NVIDIA laid out it that
• 7:06
  compresses the timeline to widespread autonomy.
• 7:09
  By mid 2026, they expect Level 2 plus urban and Hwy. driving with
• 7:13
  automated lane changes, traffic light recognition, and point to
• 7:17
  point navigation. By the end of 2026, they plan to
• 7:19
  cover the entirety of the United States in terms of operational
• 7:23
  design domain and add autonomous parking.
• 7:25
  Late 2026 will bring a small scale Level 4 trial similar to
• 7:29
  Waymo's early ROBO taxi pilots and by 2027 NVIDIA expects
• 7:34
  partner robotaxi deployments to begin.
• 7:36
  By 2028, they're targeting personally owned Level 4
• 7:39
  vehicles and Level 3 hands off driveway and Hwy. driving into
• 7:43
  production cars. That is an aggressive schedule.
• 7:46
  It depends on regulatory approvals and the confidence of
• 7:48
  automaker partners though. But the message is very clear.
• 7:51
  NVIDIA is committing to rapid iteration and deployment of this
• 7:54
  open source software. Now, the business model is
• 7:57
  significant. Nvidia's automotive division
• 8:00
  currently generates about $600 million in revenue, compared to
• 8:03
  51 billion in their core AI and cloud business.
• 8:06
  Now, if this works, every vehicle that uses Nvidia's
• 8:08
  platform brings recurring revenue for chips, software
• 8:11
  licenses and cloud services. NVIDIA also introduced the Rubin
• 8:15
  Platform, a new data center system for training autonomous
• 8:18
  driving model. That means they capture value
• 8:20
  both in the vehicle, in the cloud.
• 8:22
  As fleets scale up, Jensen Wong said the goal is that someday
• 8:26
  every car in every truck will be autonomous.
• 8:28
  NVIDIA is working toward that future with the foundation laid
• 8:32
  out at CES 2026. They have made a strong case
• 8:35
  that will make the key architectures of that future.
• 8:39
  I think it's probably the best announcement at CES 2026 as far
• 8:43
  as a regular person goes, as far as automakers go, because
• 8:47
  everybody uses a car, right? And if you have ever used Uber,
• 8:50
  you know how important it is to get your car there on time and
• 8:53
  also get you to where you're going.
• 8:54
  But also, it's not the perfect system.
• 8:57
  Once people are taken out of that equation and we can have
• 8:59
  robotaxis and we'll have to work with those robotaxis or work as
• 9:03
  those robotaxis, it's not just going to be Tesla anymore.
• 9:06
  People will be able to compete and NVIDIA is helping them.
• 9:09
  So this is a strategic frame too.
• 9:11
  Listen to this. Jensen said that Chet TBT is
• 9:14
  similar to this, and he believes this is the moment that
• 9:17
  reasoning AI will do for robots and vehicles what large language
• 9:20
  models did for text. For us, the ability to explain
• 9:23
  decisions is not just about debugging anymore.
• 9:26
  Governments want this. They want to understand why an
• 9:28
  autonomous vehicle made a particular choice before they
• 9:31
  approve it for public roads. And by exposing that reasoning
• 9:34
  trace, NVIDIA is addressing concerns about black box AI in
• 9:38
  safety. Critical application regulators
• 9:40
  and researchers can inspect how the system drives at each
• 9:44
  decision, and that transparency can make NVIDIA powered vehicles
• 9:47
  easier to validate and certify than systems that cannot explain
• 9:51
  themselves. Now, the challenges are
• 9:53
  absolutely real here. There's a lot of work to do.
• 9:55
  Automakers traditionally want to control their own stack.
• 9:58
  Handing over the core driving AI to NVIDIA means giving up a lot
• 10:02
  of that control. But the presence of Mercedes,
• 10:04
  JLR and Lucid on board suggest a shift towards collaboration.
• 10:08
  If the big players can do it, if they jump on board, I mean
• 10:11
  Mercedes come out now. Public acceptance will require
• 10:14
  flawless safety records though. NVIDIA focus on safety,
• 10:17
  including their Halos framework for validation and functional
• 10:20
  safety is aiming at building trust with government.
• 10:23
  The competitive response will also shape outcomes here.
• 10:26
  And also they're building trust with the public.
• 10:27
  They want you and me to feel comfortable in a car with NVIDIA
• 10:31
  stacks. Kind of Tesla opens parts of its
• 10:34
  stack. Or a mobile eye undercuts NVIDIA
• 10:36
  on cost. The market could go wild here.
• 10:39
  NVIDIA also released ELPA SIM, which is an open source
• 10:42
  simulation framework for testing autonomous vehicles in virtual
• 10:45
  environments. And this is cool because
• 10:47
  engineers can create realistic scenarios with configurable
• 10:50
  traffic, whether in sensor modes that run the AI through edge
• 10:53
  cases, without putting anyone at risk.
• 10:54
  They don't need an actual car to do this.
• 10:56
  It's going to be training the AI on the infinite amounts of
• 11:00
  situations that we could hit on the road.
• 11:02
  Then they're going to run the AI through edge cases without
• 11:05
  putting any people in a vehicle. The 1700 hours of driving data
• 11:09
  released alongside it covers diverse graphical locations and
• 11:13
  conditions, including rare and complex situations that are hard
• 11:16
  to capture a normal testing. Like just bad drivers.
• 11:19
  You see them every day. You see a bad driver or you'll
• 11:21
  see somebody who, this is a green light and they're on their
• 11:23
  cell phone for a while. You have to hog at them.
• 11:25
  Those kind of things happen all the time.
• 11:27
  But you can do this virtually so nobody gets hurt, nobody gets
• 11:30
  mad. So by providing both a simulator
• 11:34
  and real world data sets and videos, enabling what they call
• 11:37
  a self reinforcing development loop, developers train models on
• 11:40
  a wide range of situations, test them in simulation, and
• 11:44
  continuously improve the system in all of this.
• 11:47
  Which is crazy. It's public.
• 11:49
  You and I can check it out. Alpa SIM ALPASIM is on GitHub.
• 11:55
  The data is on a hugging face. We can check it all out.
• 11:58
  The barrier to entry for autonomous vehicle development
• 12:00
  just drops significantly. Test like it has some absolute
• 12:04
  killer competition. NVIDIA has planted their flag.
• 12:07
  The technology that only a few companies had access to is
• 12:10
  beginning to be available for everybody.
• 12:14
  Now this is going to be a crowded race here.
• 12:17
  Every vehicle manufacturer has access to this and why would
• 12:20
  they spend years developing their own from here on out when
• 12:24
  they can just plug in NVIDIA stuff and then plug in some
• 12:26
  sensors, plug in their own cameras, plug in their own
• 12:28
  stacks of other things. It seems like a no brainer to
• 12:31
  me. They're going to be not the
• 12:33
  Apple, but they're going to be the Android of vehicles.
• 12:36
  They're going to make so much money.
• 12:37
  They're going to be selling their cloud services, their
• 12:39
  sensors, their chips, etcetera. This is going to be so much
• 12:42
  money for NVIDIA. I can't wait to see what happens
• 12:44
  in the future because one, I want to take a ride in this
• 12:47
  thing. I want to see how great it
• 12:49
  actually performs compared to a Tesla.
• 12:51
  And two, just think about how easy it's going to be when you
• 12:54
  need to go someplace and you don't actually have to own a
• 12:56
  car. Any that's going to be great.
• 12:57
  That's going to be 1 asset that doesn't depreciate when you
• 13:00
  drive it off the lot. It's going to be democratization
• 13:04
  of driving.