Latest / Connecting the Dots / Nvidia’s AI Empire: From Data Centers to Space

Transcript

Nvidia’s AI Empire: From Data Centers to Space · Connecting the Dots

• 0:00
  Today's tech news moves fast, but understanding
• 0:03
  it shouldn't be hard. That's why Snarfl Solutions
• 0:06
  Group brings you Connecting the Dots, where we
• 0:09
  break down what's happening in tech and why it
• 0:11
  matters to you. Now, let's dive in with your
• 0:14
  hosts. Welcome to Connecting the Dots, broadcasting
• 0:17
  live from the Snarfl Solutions Group headquarters
• 0:20
  right here in Sacramento, California. I am your
• 0:22
  host, and I'm joined by my fantastic co -host,
• 0:24
  ready to break down how today's tech news connects
• 0:27
  to your everyday life. Good morning, everyone.
• 0:30
  and it is a very warm Tuesday, March 17th, 2026.
• 0:35
  It really is. Yeah, we've got an anomalously
• 0:37
  early heat wave hitting the west and the southwest
• 0:40
  right now. Plus, there's some pretty serious
• 0:44
  fire concerns out in the plains, but locally
• 0:45
  here in Sacramento, the morning haze is actually
• 0:48
  clearing up. Oh, nice. Yeah, it's shaping up
• 0:50
  to be sunny, but an unseasonably hot high of
• 0:53
  89 degrees Fahrenheit today, so. You know, stay
• 0:56
  hydrated out there. Definitely. Keep the AC running.
• 0:59
  Let's take a quick look at the markets before
• 1:01
  we jump in. So the Dow is up a solid 210 points
• 1:04
  today. Always good to see. Yeah. S &P is up just
• 1:07
  slightly, about seven points. But Bitcoin actually
• 1:10
  took a bit of a hit. It's down about 1 ,150 points,
• 1:13
  sitting right around 73 ,720 right now. Gotcha.
• 1:17
  And I have to mention, since I know you were
• 1:19
  watching, the SF Giants dropped one to the Padres
• 1:22
  yesterday. One to three. Look. I don't want to
• 1:25
  talk about it. But hey, Jesus Rodriguez went
• 1:27
  two for four. So, you know, there's always a
• 1:29
  silver lining. Before we get into today's news,
• 1:32
  let's talk about how we actually survive these
• 1:34
  early mornings. Yes, absolutely. We want to give
• 1:37
  a shout out to our sponsor, Strong Coffee. If
• 1:40
  you want that premium coffee shop taste, but
• 1:42
  you, you know, actually need sustained fuel instead
• 1:46
  of just a caffeine spike that leaves you crashing
• 1:48
  at 2 p .m. This is it. Yeah, it's a staple for
• 1:51
  us in the studio. Strong Coffee. is this premium
• 1:54
  instant coffee that genuinely tastes like a traditional
• 1:58
  slow -dripped brew. It really does. But the best
• 2:00
  part is it's packed with 15 grams of protein,
• 2:03
  MCTs, and adaptogens like ashwagandha. Right.
• 2:06
  So it gives you incredibly clean, sustained energy
• 2:09
  without any of those awful jitters. To upgrade
• 2:12
  your morning routine, use the promo code SNARFL
• 2:14
  for 20 % off your entire purchase at StrongCoffeeCompany
• 2:18
  .com. All right. Let's transition over to the
• 2:20
  Morningstar. That's the SNARFL Tech article report
• 2:23
  for those of you listening for the first time.
• 2:25
  Yes. And today is unofficially NVIDIA Day. It
• 2:28
  really is. Yeah. All three of the top stories
• 2:31
  we're exploring today stem from NVIDIA's GTC
• 2:34
  2026 conference. And there's this incredible
• 2:38
  through line here. We're looking at how this
• 2:40
  one company is basically redesigning the physical
• 2:43
  world of computing. Right. Starting down here
• 2:45
  in Earth's data centers and literally blasting
• 2:48
  off into outer space and finally bringing that
• 2:51
  intelligence down to the actual streets you drive
• 2:53
  on. It's wild. So let's jump right into that
• 2:56
  first story. NVIDIA unveiled a brand new liquid
• 2:59
  -cooled rack system built on their MGX modular
• 3:02
  architecture. Yeah. But the thing that really
• 3:05
  stood out to me in the report, the big shocker,
• 3:07
  is that this multi -million dollar rack isn't
• 3:10
  packed with their famous GPUs. Right, which is
• 3:12
  their whole thing. Exactly. Instead, it holds
• 3:14
  256 of their new custom Vero CPUs. Yeah, central
• 3:19
  processing units, not graphics. Which is confusing,
• 3:22
  right? I mean, everyone has been screaming about
• 3:23
  GPUs for the last three years. They are the engine
• 3:26
  of the AI boom. So why on earth is NVIDIA suddenly
• 3:29
  pivoting back to CPUs? Isn't that a massive step
• 3:32
  backward? It definitely looks like one at first
• 3:35
  glance, but it's all about the type of AI we
• 3:38
  are using now. We're shifting from just generative
• 3:41
  AI to agentic AI. OK, break that down for me.
• 3:44
  What's the difference? So generative AI is basically
• 3:47
  answering questions. You ask it to write a poem,
• 3:49
  it predicts the next word, and boom, you have
• 3:51
  a poem. GPUs are amazing at that. Right. But
• 3:54
  an agent, agentic AI actually takes action, uses
• 3:57
  tools. If you ask an agent to compile a report
• 4:00
  on a company's finances, it has to execute a
• 4:02
  SQL query, grab the data, run a Python script,
• 4:06
  and compile code. And GPUs struggle with that?
• 4:09
  They are terrible at it, honestly. GPUs do thousands
• 4:11
  of math problems at the exact same time. But
• 4:13
  running code is sequential. Step one has to finish
• 4:16
  before step two can start. Ah, I see. For currently,
• 4:19
  the GPU has to pause, ask a traditional CPU to
• 4:22
  do the sequential task, wait for the answer,
• 4:25
  and then start up again. OK, so it's a massive
• 4:27
  bottleneck. If the GPU is this brilliant architect
• 4:30
  dreaming up the blueprint for a house, the CPU
• 4:33
  is the general contractor actually swinging the
• 4:35
  hammer to build it. That is the perfect analogy,
• 4:38
  and right now the contractor is moving too slowly.
• 4:41
  So Nvidia built the Vera CPU to swing the hammer
• 4:43
  faster. Got it. But what makes Vera so much better
• 4:46
  than the standard CPUs already out there, like
• 4:49
  Intel or AMD? It comes down to memory bandwidth
• 4:51
  and low latency. The Vera chip has 88 custom
• 4:55
  Olympus ARM cores. But the secret sauce is the
• 4:57
  memory. They are using LPDDR5X. Wait, hold on.
• 5:01
  Isn't LPDDR5X mobile memory like the low -power
• 5:04
  stuff you find in smartphones and high -end laptops?
• 5:06
  It is, yeah. Why would they put laptop memory
• 5:08
  into a massive enterprise data center rack? Because
• 5:11
  of how fast it moves data over very short distances.
• 5:14
  When an agent is frantically doing tasks, it
• 5:16
  needs data instantly. This memory setup delivers
• 5:19
  1 .2 terabytes per second of bandwidth. Wow,
• 5:21
  1 .2 terabytes. Yeah, per second. It basically
• 5:24
  eliminates the bottleneck. And when you pack
• 5:26
  256 of these into a liquid -cooled rack, you're
• 5:30
  looking at over 22 ,000 cores working together
• 5:33
  to take action on your behalf. So they're reinventing
• 5:36
  the data center down here on Earth to support
• 5:38
  AI that does our busy work. That makes total
• 5:41
  sense. Yeah. But our second story takes this
• 5:43
  idea and launches it into orbit. Literally. Yeah.
• 5:47
  NVIDIA announced the space one Vera Rubin module.
• 5:50
  They are putting AI compute into space. The report
• 5:54
  says this module delivers up to 25 times more
• 5:57
  AI compute for space -based inferencing compared
• 6:00
  to their H100 ship. Which is just a staggering
• 6:03
  amount of power to put in a satellite. It is.
• 6:05
  But I have to push back a little here. Why do
• 6:07
  we need a data center floating in space? Doesn't
• 6:09
  it make way more sense to just beam the satellite
• 6:11
  weighted down to those liquid -cooled Variracks
• 6:13
  we just talked about? You would think so, but
• 6:15
  the problem is the data pipeline. Satellites
• 6:18
  taking high -resolution imagery for, say, weather
• 6:20
  tracking or disaster response, they are collecting
• 6:23
  petabytes of raw visual data. And trying to beam
• 6:26
  those massive raw files down to Earth via radio
• 6:30
  frequency data links is incredibly slow. Ah,
• 6:33
  so it's a bandwidth issue. Completely. Imagine
• 6:35
  a satellite spots the early signs of a wildfire
• 6:38
  in a remote area. If it has to send gigabytes
• 6:41
  of raw photos down to Earth to be processed,
• 6:44
  it could take hours before authorities are alerted.
• 6:47
  So it's kind of like if you have a digital camera,
• 6:50
  but instead of seeing the photo on the screen
• 6:52
  right away, you have to mail the physical memory
• 6:55
  card across the country to a photo lab. Exactly.
• 6:58
  And by the time the lab develops the photo and
• 7:00
  tells you there's a fire, the force is burned
• 7:02
  down. Wow, okay. That puts it in perspective.
• 7:04
  Yeah, so what NVIDIA is doing is putting the
• 7:06
  photo lab inside the camera. These orbital data
• 7:09
  centers process the data locally in space. The
• 7:12
  AI analyzes the imagery up there, identifies
• 7:15
  the fire, and instead of sending a giant photo,
• 7:17
  it just sends a tiny text packet down to Earth.
• 7:19
  Like fire located at these exact coordinates.
• 7:22
  Precisely. It turns disaster response from a
• 7:25
  historical analysis into near real -time intervention.
• 7:28
  That is incredible. But wait, reading through
• 7:31
  this star report, there's a massive physical
• 7:33
  problem here. Jensen Hong mentioned it himself
• 7:36
  at the conference. He said, and I quote, In space,
• 7:39
  there is no convection, only radiation. Yes,
• 7:43
  the thermodynamics problem is fascinating. Right,
• 7:45
  because down on Earth, a server rack gets hot
• 7:47
  and you use fans to blow air over it, or you
• 7:50
  use liquid to carry the heat away into the air.
• 7:53
  You're using the air molecules to transfer the
• 7:55
  heat. Exactly. But space is a vacuum. There's
• 7:58
  no air. So how do you cool a supercomputer that's
• 8:02
  25 times more powerful than an H100 when there's
• 8:05
  literally no air to run a fan? Why doesn't it
• 8:08
  just melt? It's an engineering marvel. Companies
• 8:10
  like Star Cloud and Ather Flux are pioneering
• 8:12
  these orbital setups to solve exactly this. They
• 8:16
  use passive cooling based entirely on radiation.
• 8:19
  Okay, how does that work? Since you can't use
• 8:21
  convection, you know, blowing air, they have
• 8:23
  to convert the thermal energy generated by the
• 8:25
  processors into infrared light. Wait, they turn
• 8:27
  the heat into light? Basically, yeah, they use
• 8:29
  these massive, specially angled fins on the outside
• 8:32
  of the module. The fins face away from the sun,
• 8:36
  pointing into deep space. The heat travels to
• 8:38
  the fins and simply radiates away into the freezing
• 8:40
  vacuum as infrared waves. That is mind -blowing.
• 8:43
  And the best part is the whole system is powered
• 8:46
  by unfiltered solar energy up in orbit. So it's
• 8:48
  a completely self -sustaining, passively cooled
• 8:51
  data center. Unbelievable. So we've got AI managing
• 8:54
  data centers on Earth, and we've got it tracking
• 8:56
  wildfires from space. But let's bring it back
• 8:59
  down to the literal asphalt for our third story
• 9:01
  today. Right, the streets. NVIDIA announced that
• 9:04
  major global automakers are adopting their drive
• 9:07
  Hyperion platform for Level 4 autonomous vehicles.
• 9:10
  We're talking about BYD, Geely, Isuzu, and Nissan.
• 9:14
  And Level 4 means the car can handle all the
• 9:16
  driving under specific conditions. No steering
• 9:19
  wheel needed. Right. And Uber is getting in on
• 9:21
  this, too. They're launching a global fleet of
• 9:23
  Nvidia powered robo taxes across 28 cities by
• 9:26
  2028. And they're starting right here in California,
• 9:29
  L .A. and San Francisco in 2027, which is next
• 9:32
  year. The technology is maturing so fast it's
• 9:36
  becoming a scalable commercial reality. It really
• 9:39
  is. And, you know, just to be completely neutral
• 9:42
  and objective here, we have to note the geopolitical
• 9:45
  side of this report. Nvidia is supplying this
• 9:48
  cutting edge platform to BYD and Geely, which
• 9:51
  are China's top automakers. Yeah, that's a significant
• 9:54
  detail. Right. And this is happening amid the
• 9:57
  ongoing U .S.-China trade tensions and the recent
• 9:59
  Trump administration approvals for H -200 ship
• 10:02
  sales. Now, we aren't taking any sides on the
• 10:05
  tariffs or the tech race. We're just reporting
• 10:08
  the facts from the source material. But it clearly
• 10:10
  shows that this autonomous revolution is totally
• 10:13
  global. Oh, absolutely. The market doesn't really
• 10:16
  recognize borders when it comes to solving this
• 10:18
  problem because, honestly, teaching a car to
• 10:20
  drive itself safely is incredibly difficult.
• 10:23
  Right, because AI makes mistakes. We've all seen
• 10:26
  AI chatbots hallucinate and make up facts, but
• 10:28
  if a car hallucinates a stop sign, that's a life
• 10:31
  or death situation. Exactly. And that's where
• 10:33
  NVIDIA's new Halos OS comes in. It provides the
• 10:37
  safety guardrails. Halos operates by separating
• 10:40
  the probabilistic AI from the deterministic safety
• 10:43
  systems. Okay, hold on. Probabilistic versus
• 10:45
  deterministic. Clarify that for everyone sure
• 10:48
  the AI that decides how to navigate traffic is
• 10:51
  probabilistic It's making highly educated guesses
• 10:54
  based on what its cameras see right, but the
• 10:57
  deterministic layer the hills OS operates on
• 11:00
  strict rigid math It calculates the physics the
• 11:04
  speed of the car the friction of the tires the
• 11:07
  exact distance to the car in front of you Oh,
• 11:09
  so if the AI makes a probabilistic guess that
• 11:11
  is physically unsafe the deterministic layer
• 11:14
  steps in and overrides it immediately. It's like
• 11:17
  having a hyper -vigilant driver's ed instructor
• 11:19
  sitting in the passenger seat. That's exactly
• 11:21
  it. The AI is the student driver making the decisions,
• 11:24
  but the instructor has their own brake pedal
• 11:26
  operating on pure physical reality, ready to
• 11:28
  intervene if the student makes a dangerous move.
• 11:31
  Perfect way to picture it. And to make the student
• 11:33
  driver better, NVIDIA introduced an open reasoning
• 11:36
  model called Alpamayo 1 .5. This lets developers
• 11:40
  steer the AI's behavior using simple text prompts.
• 11:44
  Really? Text prompts? Not coding? Yeah. Say the
• 11:48
  car encounters a really weird edge case, like
• 11:51
  a construction zone where traffic is routed through
• 11:54
  a parking lot filled with traffic cones. It's
• 11:56
  a total nightmare scenario for an AI. Right.
• 11:59
  Instead of rewriting thousands of lines of code,
• 12:02
  engineers can just type a prompt into AlpaMio.
• 12:04
  When you see cones arranged in this specific
• 12:06
  pattern, Follow the outside edge at five miles
• 12:09
  per hour. And the AI internalizes that reasoning.
• 12:12
  But wait, how do they test that? They can't just
• 12:14
  throw the car back out onto the highway and hope
• 12:16
  it doesn't crash into the cones, right? No, definitely
• 12:18
  not. They use a simulator called Omniverse Nurak.
• 12:21
  And this uses a technique called 3D Gaussian
• 12:23
  Splatting. Oh, a 3D Gaussian Splatting. I love
• 12:26
  that name. It sounds like a messy art project.
• 12:27
  What does that actually mean? It's actually revolutionary.
• 12:31
  Normally, if you want to simulate a driving environment,
• 12:33
  you have to build a 3D world out of polygons,
• 12:36
  like a video game. It's time consuming, and it
• 12:38
  never quite looks perfectly real. Right. It looks
• 12:40
  like a computer graphic. Exactly. But with Gaussian
• 12:44
  splatting, they take actual 2D dash cam footage
• 12:47
  from a real drive, and the AI mathematically
• 12:50
  reconstructs it into a 3D environment using millions
• 12:54
  of blurry, tiny blobs of color. the Gaussian
• 12:57
  splats. Wow. So it turns a simple, flat video
• 13:00
  into a photorealistic, fully 3D world that the
• 13:05
  simulated car can drive through. Yes. They can
• 13:07
  take that weird construction zone footage, turn
• 13:09
  it into a 3D simulation, and then test the car's
• 13:12
  new text -prompted reasoning millions of times
• 13:15
  in the simulator before it ever hits the real
• 13:17
  street. That is just incredible technology, but
• 13:19
  it brings up a big question for me about NVIDIA's
• 13:21
  business strategy. Okay, what's that? Well, NVIDIA
• 13:23
  makes tens of billions of dollars selling those
• 13:25
  AI data center chips we talked about. But if
• 13:28
  you look at their auto division, it only brought
• 13:29
  in $592 million recently. Yeah, a fraction of
• 13:33
  their total revenue. A tiny fraction. So why
• 13:35
  put all this effort into Halos OS and Gossin
• 13:38
  Splatting and these massive partnerships? Is
• 13:41
  this just a side project for them? Not at all.
• 13:44
  Jensen Huang has been very clear about this.
• 13:46
  He calls autonomous vehicles the first multi
• 13:50
  -trillion dollar robotics industry. Wow! Multi
• 13:53
  -trillion! Yeah. He's playing the long game here.
• 13:56
  By providing the foundational safety guardrails
• 13:59
  and the simulation engines now, NVIDIA is positioning
• 14:02
  themselves to power the entire global robo taxi
• 14:05
  fleet when it inevitably scales up. So they're
• 14:08
  building the infrastructure for the future. Exactly.
• 14:10
  Well, before we wrap up and connect all these
• 14:12
  dots, we have to mention our second sponsor today,
• 14:14
  Old Glory. Yes. If you are looking for a gift
• 14:17
  for the techie friend in your life, you have
• 14:20
  to check them out. Totally. They're an iconic
• 14:22
  music and sports fan merch store. They have an
• 14:24
  amazing selection of gear. And right now, if
• 14:26
  you use the promo code SNARFL at checkout, you
• 14:29
  get 15 % off your entire purchase. That's a great
• 14:31
  deal. Yeah. Just go to oldglory .com and use
• 14:34
  code SNARFL. So looking at everything we discussed
• 14:39
  today really paints a clear picture This one
• 14:41
  company's technology is reshaping reality across
• 14:44
  every level. It really is at work Those Vera
• 14:47
  CPUs are going to be powering the AI agents doing
• 14:50
  your tedious tasks. Right. And from up in the
• 14:53
  sky, those space one modules are going to be
• 14:55
  monitoring the climate and catching disasters
• 14:57
  in real time. And on your commute, the Hyperion
• 15:00
  platform is literally going to be driving your
• 15:02
  Uber in just a couple of years. Intelligence
• 15:04
  is moving everywhere. It's at the edges of our
• 15:06
  universe, from satellites down to the cars on
• 15:09
  the street and the devices in our pockets. Yeah.
• 15:12
  Which actually leads to a really provocative
• 15:13
  thought to end on today. We've spent the last
• 15:15
  decade obsessed with the cloud, right? This idea
• 15:19
  that all our computing happens in some massive
• 15:21
  centralized data warehouse somewhere. Right,
• 15:24
  relying on the cloud for everything. But if all
• 15:26
  this immense AI intelligence is moving to the
• 15:29
  literal edges processing natively in orbit or
• 15:32
  directly inside your car, how long until the
• 15:35
  whole concept of a centralized cloud becomes
• 15:37
  completely obsolete? Oh, wow. Yeah, if every
• 15:40
  device is its own supercomputer, the cloud just
• 15:42
  becomes a storage drive. Exactly. It's a wild
• 15:44
  shift to think about. Something to mull over
• 15:47
  this week. We're here to help at Snarful, so
• 15:49
  reach out on our website if you have any more
• 15:51
  questions. That's a wrap on today's top stories.
• 15:58
  Remember, tech doesn't have to be complicated,
• 16:00
  and we're here to help connect the dots. If you
• 16:03
  got questions or want to learn more, visit us
• 16:05
  at snarful .com. Thanks for listening, and we'll
• 16:08
  catch you next time.