Latest / Elon Musk Podcast / SpaceX Engineers create amazing product

Transcript

SpaceX Engineers create amazing product · Elon Musk Podcast

• 0:00
  OK, picture this. You have successfully gathered
• 0:03
  the smartest minds in the history of the universe into one
• 0:06
  single room. You have Einstein, you have
• 0:08
  Curie, you have Turing, you have Hawking.
• 0:10
  Millions of them. Actually, in the context of the
• 0:13
  current tech landscape, these are your GPU's.
• 0:16
  They are the silicon brains that are ready to solve cancer,
• 0:20
  crack, fusion energy, and, I don't know, write the next Great
• 0:24
  American novel all at the same time.
• 0:26
  A room full of potential. Right, but there is a catch.
• 0:30
  There is always a catch. And it's a massive 1.
• 0:32
  The room is the size of a football stadium, and it is
• 0:35
  deafeningly loud. Everyone is shouting.
• 0:37
  And because sound travels relatively slowly, by the time
• 0:40
  Einstein hears what Turing shouted from the other side of
• 0:43
  the room, the man has passed. The calculation is stale.
• 0:47
  So instead of a super intelligence, you just have a
• 0:49
  very expensive, very loud cafeteria.
• 0:52
  That is a terrifyingly accurate metaphor for the single biggest
• 0:56
  bottleneck in artificial intelligence right now.
• 0:58
  We are, you know, we're collectively obsessed with the
• 1:01
  brains, the chips, the NVIDIA H1, hundreds, the black whales.
• 1:04
  The shiny objects. Exactly the shiny objects, but
• 1:08
  we have completely neglected the ears and the mouths, the parts
• 1:11
  that let them all talk to each other.
• 1:12
  Exactly. And that is where we are
• 1:14
  starting today. We are ignoring the processors
• 1:16
  for a moment to look at the hidden nervous system of the AI
• 1:19
  boom. We are going to talk about a
• 1:21
  piece of hardware that acts as the universal translator in that
• 1:25
  room of geniuses. And this little piece of
• 1:27
  hardware is called an optical transceiver.
• 1:30
  It sounds like a prop from Star Trek, but it is actually the
• 1:33
  specific component that might determine whether AI scales up
• 1:36
  to the level of AGI or stalls out because of well.
• 1:40
  Physics. It really is that critical.
• 1:42
  It's the plumbing. And when the plumbing fails, it
• 1:45
  doesn't matter how smart the people in the building are.
• 1:47
  To guide us through this, we are pulling from a fascinating
• 1:50
  TechCrunch article. It's dated February 17th, 2026.
• 1:54
  The title is SpaceX Vets Raise $50 Layer Series A for Data
• 1:59
  Centrelinks and it's written by Tim Fernholtz.
• 2:01
  This story really has everything.
• 2:03
  I mean, it has rocket scientists coming back to Earth to solve a
• 2:05
  terrestrial problem. It has a $50 million bet from a
• 2:09
  major VC firm, Drive Capital. And it has a manufacturing
• 2:12
  challenge that touches on everything from global
• 2:15
  geopolitics to the future of American industry.
• 2:18
  It's a huge story packed into one little startup.
• 2:22
  The protagonists here are a startup called Mesh Optical
• 2:25
  Technologies, and as the headline suggests, they aren't
• 2:28
  your typical Silicon Valley software kids.
• 2:31
  No, not at all. They are former SpaceX engineers
• 2:34
  who cut their teeth working on Starlink.
• 2:36
  Which is crucial context, right? Right.
• 2:38
  They spent years figuring out how to make satellites talk to
• 2:41
  each other in the vacuum of space using lasers.
• 2:44
  I mean, think about that environment.
• 2:45
  You can't send a repairman. It has to work, and it has to
• 2:48
  work perfectly for years. So they're trying to apply that
• 2:52
  same rigor, that same kind of, you know, bulletproof
• 2:54
  engineering to the data center. Exactly.
• 2:57
  So here is our mission for this deep dive.
• 3:00
  We are going to explain what an optical transceiver actually is,
• 3:03
  using analogies that even a business executive who has never
• 3:07
  touched a soldering iron can understand.
• 3:09
  And we'll breakdown why the founders background at SpaceX
• 3:12
  makes them uniquely qualified for this specific challenge.
• 3:15
  And then we are going to explore the massive, and I mean massive
• 3:19
  challenge of bringing lights out automated manufacturing back to
• 3:23
  the US. It is a story about light speed
• 3:27
  and the sheer difficulty of building physical things in a
• 3:30
  digital age. It's atoms versus bits.
• 3:33
  I love it. Let's jump right in Part 1, the
• 3:35
  tech because optical transceiver sounds complicated.
• 3:38
  It sounds intimidating. It does, but the concept is
• 3:41
  actually very elegant if we strip away the jargon, I
• 3:43
  promise. OK, so let's do that for the
• 3:45
  business executive listening, who knows strategy, who knows
• 3:48
  panel statements, but maybe skipped electrical engineering
• 3:51
  in college. Help us visualize this.
• 3:54
  Let's go back to our room of geniuses, right?
• 3:56
  The loud cafeteria inside a computer chip.
• 3:59
  So inside that geniuses brain information travels as
• 4:02
  electricity. OK, it's electrons moving
• 4:05
  through tiny copper or traces, and that works great for very,
• 4:08
  very short distances, like millimeters inside a chip or
• 4:11
  maybe centimeters on a motherboard.
• 4:13
  It's super fast and efficient for that microscopic scale.
• 4:16
  OK, so electrons are the local dialect.
• 4:18
  They speak electricity. That's the internal monologue of
• 4:20
  the genius. That is a perfect way to put it.
• 4:22
  It's their internal thought process.
• 4:25
  But electricity has a physical limitation as soon as you try to
• 4:28
  push it fast over to long distance, and in modern
• 4:31
  computing, long distance can just be across a server rack
• 4:35
  barely a few feet. Just a few feet.
• 4:37
  Yeah, we're not talking miles. As soon as you do that, it meets
• 4:40
  resistance. The copper wire literally fights
• 4:43
  back. It fights back it.
• 4:45
  Gets hot, the signal degrades. It becomes noisy.
• 4:48
  It's like trying to shout across that football field.
• 4:51
  You lose your voice and the person on the other end only
• 4:53
  hears mumbling. The message gets garbled.
• 4:56
  So the local dialect doesn't travel well.
• 4:59
  You can't shout electricity across the room and expect
• 5:01
  anyone to understand you. Precisely.
• 5:03
  You lose both speed and clarity, but there is a universal
• 5:06
  language that travels perfectly over long distances with almost
• 5:10
  no loss. Light.
• 5:11
  Light photons. This stuff of fiber optics.
• 5:15
  Right. Light is incredibly fast.
• 5:17
  I mean, it's the speed limit of the universe.
• 5:19
  It generates almost no heat compared to electricity during
• 5:22
  transmission, and it can carry a massive amount of data without
• 5:26
  losing integrity. You can send a beam of light
• 5:29
  thousands of miles under the ocean and the message arrives
• 5:32
  perfectly clear. OK, so now we have a problem.
• 5:35
  The chips speak electricity, but the best way to connect them is
• 5:38
  with the light. They're speaking two different
• 5:39
  languages. Exactly.
• 5:40
  So you need a translator. And that's the transceiver.
• 5:43
  That is a transceiver. It is a tiny, incredibly
• 5:45
  sophisticated little box that sits at the edge of the
• 5:48
  computer. It takes the electrical signal
• 5:50
  from the GPU, the local dialect translates it into a pulse of
• 5:54
  laser light, zips it across the fiber optic cable to another
• 5:57
  computer, and then another transceiver at the other end
• 6:01
  catches that light pulse and translates it back into
• 6:04
  electricity so the receiving chip can understand it.
• 6:07
  So it's literally a translator box.
• 6:09
  Electricity in, light out, and on the other side, light in,
• 6:11
  electricity out. That is exactly what it does,
• 6:14
  and it has to do this billions, even trillions of times per
• 6:17
  second with near perfect accuracy.
• 6:19
  It's an amazing piece of engineering.
• 6:21
  For a long time we didn't really need this for short distances,
• 6:24
  right? We used something else, right?
• 6:27
  For a long time, we relied on radio frequencies, or RF, to
• 6:31
  move data over copper cables for these shorter distances.
• 6:35
  Think about your Ethernet cable at home.
• 6:38
  But Travis Brashears, the CEO of Mesh Optical, put it really well
• 6:42
  in the source material he said. I'm quoting here.
• 6:45
  The world has primarily focused on radio frequencies for a long
• 6:48
  time. We want to be at the precipice
• 6:50
  of transition from RF to photonics.
• 6:53
  Precipice of transition? That sounds dramatic.
• 6:56
  It is, though. He's talking about a fundamental
• 6:58
  shift in the physics of how we build computers, right?
• 7:00
  We're hitting the wall with copper and RF.
• 7:02
  You just can't push electrons any faster or denser without the
• 7:06
  the components. So the future is moving from
• 7:08
  pushing electrons through copper to shooting photons through
• 7:11
  glass. Yes, and not just for the long
• 7:14
  haul cables under the ocean, which we've done for decades,
• 7:17
  but for the short, incredibly fast connections between
• 7:20
  computers in the same room. The connections that form the
• 7:23
  brain of an AI. So if I'm that executive, I'm
• 7:26
  thinking, OK, I understand the box, it translates, but why does
• 7:31
  this matter now? Why is this suddenly a crisis?
• 7:33
  Because the sheer volume of conversation has exploded.
• 7:36
  It's not one genius talking to another anymore.
• 7:39
  It's a million geniuses all trying to talk to each other at
• 7:42
  the exact same time. The vision is interconnecting
• 7:44
  everything, right? I think Brashears mentions that.
• 7:46
  He does. He explicitly says we want to
• 7:49
  interconnect everything, and not just computers.
• 7:51
  But that's where we're starting. This is step one in a much
• 7:54
  larger vision. OK, so that's the what?
• 7:57
  It's a high speed translator for the internet's internal traffic.
• 8:00
  Now let's talk about the why. Because usually plumbing is
• 8:03
  boring. You only notice it when it
• 8:04
  breaks. Why is Thrive Capital dumping
• 8:07
  $50 million into this right now? Because the plumbing is about to
• 8:11
  burst. In fact, it's already leaking.
• 8:13
  That sounds ominous. It's the scale of AI.
• 8:17
  We have to wrap our heads around the numbers here because they
• 8:19
  are just staggering. When we talk about scaling a web
• 8:22
  app, maybe you add a few more servers to handle more users.
• 8:25
  Easy enough. Yeah, When you scale an AI
• 8:28
  model, you are building a supercomputer that acts as one
• 8:31
  cohesive brain. All the parts have to work in
• 8:34
  perfect concert. The source material specifically
• 8:37
  mentions a 1,000,000 GPU cluster, which just a few years
• 8:42
  ago would have sounded like science fiction.
• 8:44
  Completely. And here is the key stat from
• 8:47
  the article that blew my mind. Breshear says someone will brag
• 8:51
  about a million GPU cluster. You have to multiply by 4 to 5
• 8:56
  for the number of transceivers in that cluster.
• 8:58
  Wait, hold on, let's do that math. 1,000,000 GPU's?
• 9:01
  You're telling me you need 4 to 5 million optical transceivers?
• 9:05
  Correct. 4 to 5 million of these high tech translator boxes for
• 9:09
  one single installation. Why?
• 9:12
  That seems counterintuitive? Why does one chip need 5
• 9:15
  translators? Because of how deep learning
• 9:17
  works, these models, like the ones powering the chat bots we
• 9:21
  all use, aren't trained on a single computer.
• 9:24
  They're too big. They are trained by slicing the
• 9:27
  brain into thousands of pieces and spreading it across
• 9:30
  thousands of chips. But those chips have to
• 9:32
  constantly talk to each other. It's a constant, furious
• 9:35
  conversation. Hey, I learned this pattern.
• 9:37
  What did you see? I updated this weight.
• 9:39
  You need to adjust that. So the chatter is constant.
• 9:41
  It's not just one chip sending another a file to download, it's
• 9:46
  active real time collaboration. It's relentless.
• 9:50
  It's a many to many conversation.
• 9:52
  Every GPU needs a high speed link to many other GPU's.
• 9:56
  If those links are slow, if the translators can't keep up, the
• 9:59
  whole system stalls. You have a million Ferraris
• 10:01
  stuck in traffic. And at that point, you spent
• 10:03
  billions on GPU's that are just sitting there waiting.
• 10:06
  Exactly. You're just burning electricity.
• 10:08
  That's why the market is exploding.
• 10:09
  Companies are desperate for faster, more efficient plumbing.
• 10:12
  And it is a massive market. The article brings up a really
• 10:15
  concrete example. It mentions A supplier called
• 10:18
  AOI. Yes, Applied Optoelectronics, an
• 10:21
  established US supplier. A big player.
• 10:24
  And they won a contract just last year worth $4 billion with
• 10:28
  AB, with AB $4 billion. And that was just for AWS data
• 10:32
  centers. 1 customer, one contract.
• 10:37
  Wow, that gives you a sense of the pot of gold at the end of
• 10:39
  this rainbow. We are talking about a
• 10:43
  foundational component for the entire AI industry.
• 10:46
  OK, so the money is there, the demand is there, everyone needs
• 10:49
  this. But what is Mesh Optical doing
• 10:52
  differently? Why not just buy from AOI or the
• 10:54
  other existing guys? There must be a catch.
• 10:57
  There is. The existing solutions are
• 10:59
  hitting a wall in efficiency. This is where we get into the
• 11:01
  weeds of the engineering a bit, but stay with me, it's
• 11:03
  important. Let's do it.
• 11:05
  Mesh is core innovation. Their secret sauce is a new
• 11:08
  design that removes a specific, commonly used component from the
• 11:12
  transceiver. The source calls it a power
• 11:14
  hungry component. Right, they're referring to
• 11:16
  something called a DSP, a digital signal processor or
• 11:19
  similar components called retimers.
• 11:22
  Think of these as little amplifiers and error checkers
• 11:24
  for the signal. They cleaned up the mumbling I
• 11:26
  mentioned earlier. OK, that sounds useful.
• 11:28
  Why would you want to remove it? Because it uses a ton of power
• 11:31
  and generates a ton of heat, it's a necessary evil in older
• 11:34
  designs. Mesh believes their architecture
• 11:37
  is so clean and efficient that they don't need it, or at least
• 11:40
  they can use a much simpler, lower power version.
• 11:43
  So by stripping that out, what's the impact?
• 11:45
  They claim they can reduce the power usage of a GPU cluster by
• 11:48
  3% to 5%. You know, I can hear the
• 11:50
  listener saying 3%. That's it.
• 11:52
  I tip more than that for coffee. It sounds tiny.
• 11:55
  I know, right? In the consumer world, 3% is a
• 11:58
  rounding error. You wouldn't even notice it on
• 12:00
  your phone's battery life. Exactly.
• 12:01
  But in the world of hyperscalers, Amazon, Google,
• 12:04
  Microsoft, 3% is a fortune. An absolute fortune.
• 12:08
  Because these data centers use as much power as a a small city.
• 12:11
  A medium sized city in some cases.
• 12:13
  We were talking about gigawatts of electricity.
• 12:15
  A 3% of 5% reduction across a massive cluster saves 10s of
• 12:20
  millions of dollars a year in electricity bills.
• 12:22
  But honestly, the money is almost the secondary benefit.
• 12:24
  What's the primary one? Heat.
• 12:26
  Heat is the enemy. Heat is the ultimate enemy of
• 12:29
  computing. Every single Watt of power you
• 12:32
  use generates heat. If you generate heat, you have
• 12:35
  to run massive air conditioning systems to cool it down, which
• 12:38
  takes more power. It's this vicious cycle I see.
• 12:42
  If mesh can reduce the heat generated by the plumbing by 5%,
• 12:46
  that means you can pack the GPU's closer together, or you
• 12:49
  can run them harder and faster without the melting.
• 12:51
  It's a win for density and a win for performance.
• 12:54
  Philip Clark, the partner at Thrive Capital.
• 12:56
  He pointed this out too in the article.
• 12:58
  He did. He said mesh is solving the
• 13:00
  immediate term need for better interconnects to keep scaling
• 13:04
  AI. This isn't a science project.
• 13:06
  For 10 years from now. The hyper scalers have Rd. maps,
• 13:09
  they know the power and heat limits that are about to hit
• 13:11
  next year and they are desperate for solutions.
• 13:13
  They need this hardware yesterday, literally.
• 13:15
  So we have a massive demand, a bottleneck in physics, and a
• 13:18
  solution that saves critical power and heat.
• 13:21
  Now let's look at the team, because this isn't just three
• 13:24
  grads from a coding boot camp. This is the SpaceX DNA.
• 13:28
  This is one of the most compelling parts of the story
• 13:30
  for me. The founders Travis Brashears,
• 13:33
  the CEO, Chairman Ramos, the President, and Serena Grown
• 13:37
  Heberly, the VP of product. They all work together at
• 13:40
  SpaceX, not just at the same company, but on the same
• 13:42
  projects. Yes, specifically they worked on
• 13:45
  optical communications links for Starlink.
• 13:48
  Starlink is the satellite Internet constellation, right?
• 13:51
  The one with thousands of satellites.
• 13:54
  That's the one. Thousands of satellites orbiting
• 13:56
  Earth beaming Internet down to us.
• 13:58
  But crucially, those satellites also have to talk to each other
• 14:02
  in space. How do they do that?
• 14:04
  With lasers, that is an optical link.
• 14:07
  They are shooting beams of light from one satellite to another
• 14:10
  across hundreds of miles of empty space to relay data around
• 14:13
  the globe. So they were building laser
• 14:15
  plumbing for space. Precisely.
• 14:17
  And let's just think about space for a moment.
• 14:19
  It is the harshest engineering environment imaginable.
• 14:23
  You have extreme temperature swings, boiling hot in the sun,
• 14:26
  freezing cold in the earths shadow.
• 14:28
  You have constant radiation bombarding the electronics.
• 14:31
  And most importantly, you cannot send a repair technician to low
• 14:35
  Earth orbit. If the Wi-Fi breaks in space,
• 14:38
  nobody can come and reboot the router.
• 14:40
  Exact. You can't just jiggle the cable.
• 14:42
  So the engineering culture at SpaceX is all about extreme
• 14:46
  reliability, vertical integration, and aggressive
• 14:49
  scaling. You don't just build 1 perfect
• 14:51
  satellite, you build thousands of them cheaply and quickly.
• 14:54
  And that mindset is what they're bringing to this new problem.
• 14:58
  The article mentions a specific aha moment for them.
• 15:02
  It wasn't just hey let's go do a startup.
• 15:05
  It came out of a problem they were trying to solve at SpaceX.
• 15:07
  It did. This is a classic scratch your
• 15:09
  own itch story. They were designing a new
• 15:12
  generation of space F satellites that were very compute hungry.
• 15:17
  They needed to move a lot of data inside the satellite itself
• 15:20
  between different processors. So naturally they did what any
• 15:24
  engineer would do. They look at the existing market
• 15:26
  for optical transceivers to see if they could just buy the parts
• 15:29
  off the shelf, and they saw limitations.
• 15:31
  That's the very polite engineering way of saying this
• 15:34
  stuff isn't good enough. It's too slow.
• 15:36
  It uses too much power and it's not reliable enough for what we
• 15:39
  need. I love that it's the ultimate
• 15:41
  engineer's move. We checked the store, didn't
• 15:45
  like the merchandise, so we decided to build our own
• 15:47
  factory. It's classic innovation by
• 15:49
  necessity. They couldn't find what they
• 15:51
  needed for the rigor of space, so they decided to build it
• 15:54
  themselves. And then they had the bigger
• 15:56
  realization. Wait a minute, the data centers
• 15:58
  down on Earth have the exact same problem, just at a much,
• 16:01
  much bigger scale. They are taking that space grade
• 16:04
  mindset where failure is not an option and efficiency is
• 16:07
  everything, and applying it to the data center floor.
• 16:09
  And that SpaceX DNA informs their entire strategy.
• 16:13
  It's not just about the design of the product, it's about how
• 16:16
  they plan to build it. Which brings us to the biggest
• 16:19
  hurdle in this entire story, the real crux of the challenge.
• 16:22
  The manufacturing. The manufacturing this.
• 16:24
  Is Part 4 the challenge? And frankly, this is the part
• 16:28
  that makes me nervous for them, because designing a cool gadget
• 16:31
  on a computer is one thing. Building millions of them is a
• 16:34
  whole other ball game. And building millions of them in
• 16:37
  the United States is a whole other level of difficulty on top
• 16:40
  of that. It's like playing the game on
• 16:41
  ultra hard mode. Let's look at their goals from
• 16:43
  the article. They want to be manufacturing
• 16:45
  1000 units per day within the year.
• 16:48
  Is an absolute Sprint. And then the real goal is to
• 16:51
  qualify for bulk orders in 2027 and 2028.
• 16:54
  That means ramping up to 10s of thousands a day to feed the
• 16:57
  hyperscalers. And the founders say the main
• 17:00
  challenge, the thing that keeps them up at night, is executing
• 17:03
  lights out automated manufacturing techniques.
• 17:06
  Let's unpack that term. Lights out.
• 17:08
  It's a very evocative phrase. It is.
• 17:10
  Ideally it means you can literally turn the lights off in
• 17:12
  the factory and leave because there are no humans working on
• 17:15
  the assembly line. It's just robots building
• 17:17
  things, 2047. Exactly.
• 17:19
  It implies total automation, high precision, high speed, and
• 17:24
  theoretically 0 human error. And here is the brutal reality.
• 17:29
  The United States, as an industrial base, has largely
• 17:32
  forgotten how to do this for optoelectronics.
• 17:35
  The source is pretty blunt about this.
• 17:36
  It says this expertise is heavily concentrated in China.
• 17:39
  It is the China problem, as it's often called over the last 20-30
• 17:44
  years. The entire supply chain, the
• 17:46
  engineering talent, the specific institutional know how for mass
• 17:51
  producing these tiny complex optical components.
• 17:56
  It all moved to Asia and primarily to China.
• 17:58
  There is an anecdote in the article that really, really
• 18:01
  drives this home for me. It's about a European equipment
• 18:04
  supplier. The German form story.
• 18:07
  This stuck with me too. It's so telling.
• 18:09
  Tell us about that, because it's a perfect illustration of the
• 18:11
  problem. O Mesh Otical is in the process
• 18:14
  of buying the complex machinery they need to set U their factory
• 18:18
  in Los Angeles. They go to a German firm that
• 18:21
  makes the best in class manufacturing gear.
• 18:24
  The robots that build the transceivers.
• 18:26
  The standard intake form for that German company, the one you
• 18:28
  have to fill out to become a customer, asks for a Chinese
• 18:32
  company registration number. Wow, not just company
• 18:35
  registration number, specifically a Chinese one.
• 18:37
  Just think about the baked in assumption there.
• 18:40
  The form assumes by default that if you are a company buying this
• 18:44
  highly specialized equipment, you must be a Chinese company,
• 18:48
  because who else in the world would be buying it at scale?
• 18:51
  That is a sobering detail. It shows just how entrenched
• 18:55
  that global supply chain dominance is.
• 18:57
  It's not just a talking point. It's built into the paperwork.
• 19:00
  It really is. It's not just that China has the
• 19:03
  factories. The entire global ecosystem of
• 19:06
  suppliers and experts assumes China is the only place this
• 19:09
  kind of manufacturing happens. But Mesh is trying to change
• 19:12
  that. And this is where the
• 19:13
  geopolitical angle comes roaring in Thrive Capital.
• 19:16
  Philip Clark again. He didn't mince words about the
• 19:19
  national security aspect of this.
• 19:20
  No, he didn't. And just to be clear for our
• 19:23
  listeners, this is the perspective of the investor in
• 19:25
  the source text he wrote to TechCrunch, and I'm
• 19:28
  paraphrasing, but the gist was if AI is the most important
• 19:32
  technology of our generation and they believe it is, then having
• 19:35
  the critical CapEx parts, the foundational hardware running
• 19:38
  through misaligned competitive countries is a massive strategic
• 19:42
  problem. Misaligned countries that is VC
• 19:45
  speak for. We don't want our entire AI
• 19:48
  backbone dependent on China. That's exactly what it means.
• 19:52
  It's a supply chain risk. The article notes that trade
• 19:54
  restrictions haven't hit this specific market yet, but
• 19:58
  everyone sees the writing on the wall.
• 20:00
  If geopolitical tensions rise, you don't want your access to
• 20:04
  optical transceivers suddenly cut off.
• 20:06
  That would strangle your domestic AI progress overnight.
• 20:10
  So Mesh sees a strategic advantage in building a supply
• 20:13
  chain outside of China. They're positioning themselves
• 20:16
  as the secure domestic alternative.
• 20:18
  They do. They are trying to get ahead of
• 20:20
  the dilemma to offer a solution before it becomes a full blown
• 20:24
  crisis for the US tech industry. And they are doing it by Co
• 20:27
  locating their design and production.
• 20:29
  Right. Their design team is in Los
• 20:30
  Angeles. Their production line is being
• 20:32
  built in Los Angeles. They're in the same building.
• 20:34
  Why does that matter in a world of Zoom and global logistics?
• 20:37
  Why not design in LA and manufacture in, I don't know,
• 20:40
  Vietnam or Mexico? Because when you are trying to
• 20:43
  invent a completely new manufacturing process, this
• 20:46
  lights out automation. You need the design engineer
• 20:50
  standing next to the robots on the factory floor.
• 20:53
  If there is a tiny 3% efficiency gain to be found, you find it by
• 20:57
  watching the robot, quinking the code and iterating instantly.
• 21:01
  You don't find it by emailing a factory halfway across the world
• 21:04
  and waiting two weeks for a new prototype to ship back.
• 21:07
  The feedback loop has to be instantaneous.
• 21:09
  So it's about the speed of iteration, the speed of learning
• 21:12
  it's. Speed.
• 21:13
  And they believe ultimately it's about cost.
• 21:16
  They're making a bet that if they can crack the code on
• 21:19
  automation, they can actually produce these things cheaper in
• 21:22
  the US than they could by outsourcing, simply because they
• 21:25
  eliminate the shipping, the delays, the tariffs, and the
• 21:28
  communication overhead. That is the Holy Grail of re
• 21:31
  shoring, isn't it? Bringing manufacturing back not
• 21:33
  out of some sense of patriotism, but because it's actually Better
• 21:36
  Business it. It the only way it works long
• 21:38
  term. Patriotism is great for a press
• 21:40
  release, but unit economics rule the world.
• 21:43
  I want to pause on the lights out aspect for a second.
• 21:45
  We say robots building robots, but what does that actually look
• 21:49
  like for something this small and delicate?
• 21:52
  That's a great question because we aren't talking about the
• 21:55
  giant robotic arms you see welding car frames in a Tesla
• 21:59
  factory. These transceivers are small,
• 22:01
  maybe the size of a stick of gum.
• 22:04
  The components inside are microscopic.
• 22:06
  So we're talking about incredible precision.
• 22:08
  Extreme precision. Remember, one of the key steps
• 22:12
  is to align a tiny laser with a fiber optic cable.
• 22:16
  The core of that fiber optic cable is about the width of a
• 22:19
  human hair. You have to shoot a beam of
• 22:22
  light perfectly into that hair width target, and you have to do
• 22:25
  it while everything is hot and vibrating slightly, and then you
• 22:28
  have to lock it in place with a microscopic dab of epoxy.
• 22:31
  And doing that by hand is slow and.
• 22:33
  Difficult. It's incredibly slow, and it's
• 22:35
  prone to air. If a human technician has a
• 22:37
  shaky hand after their morning coffee, the laser misses the
• 22:41
  target and the part is garbage. Automation allows for what's
• 22:44
  called active alignment. It's where the robot holds the
• 22:47
  laser. A sensor measures the light
• 22:49
  coming out the other end of the fiber, and the robot software
• 22:52
  uses that feedback to adjust the lasers position by nanometers
• 22:56
  until the signal is perfect, then it locks it in.
• 22:59
  And it does this in a fraction of a second.
• 23:01
  Right, to hit that 1000 units a day target and eventually 10s of
• 23:05
  thousands, those robots have to be a blur of motion performing
• 23:09
  these microscopic surgeries over and over again.
• 23:12
  Florida State. So when they say lights out,
• 23:14
  it's not just about saving money on electricity for the room
• 23:17
  lights, it's about creating a process that humans physically
• 23:21
  cannot perform at the required speed and quality.
• 23:24
  Exactly. Humans become the bottom neck in
• 23:27
  the manufacturing process. Just like copper is the
• 23:29
  bottleneck in the data transmission process.
• 23:32
  Mesh is trying to remove the human element from the assembly
• 23:35
  line to reach the necessary scale and precision.
• 23:38
  It's removing friction everywhere.
• 23:40
  Friction in the wire, Friction in the factory.
• 23:42
  That is the theme of this whole story, reducing friction to
• 23:45
  increase speed. Let's synthesize this.
• 23:47
  We are moving into part 5 here, connecting all these dots.
• 23:50
  Let's look at the narrative arc we've uncovered, because it's a
• 23:52
  really powerful one. We started with a fundamental
• 23:55
  bottleneck in physics, the AI geniuses.
• 23:58
  The GPU's are so fast that the current plumbing can't keep up.
• 24:03
  Right, The copper wires and the old radio frequency tech are
• 24:05
  hitting a wall. The conversation in the loud
• 24:08
  cafeteria is breaking down. Enter the solution, a team of
• 24:12
  engineers from SpaceX who spent years solving this exact
• 24:16
  problem, just in the much harsher environment of the
• 24:18
  vacuum of space. They bring a mindset of extreme
• 24:21
  reliability and a clever new design that saves critical power
• 24:26
  and reduces HEAT, which is the number one enemy of performance.
• 24:29
  But they run head first into a wall, a geopolitical and
• 24:32
  industrial wall. The manufacturing capability to
• 24:35
  build this thing at scale doesn't really exist in the West
• 24:38
  anymore. So they have to build that too.
• 24:40
  They have to reinvent the factory floor while they
• 24:42
  reinvent the component that's built on it.
• 24:44
  It's a double challenge. It's product innovation plus
• 24:47
  process innovation happening at the same time.
• 24:49
  Which is incredibly risky. Most startups would only tackle
• 24:52
  one of those, but if it works, the payoff is absolutely
• 24:55
  enormous. You don't just own a product,
• 24:57
  you own the entire means of production.
• 25:00
  Let's go back to that quote from Travis Brochures one more time.
• 25:03
  We want to interconnect everything.
• 25:04
  That is the bigger picture here. We are fixated on AI right now
• 25:08
  because that is the immediate term demand that's driving the
• 25:12
  $50 million investment. But think about what happens
• 25:15
  when this kind of optical communication becomes cheap,
• 25:19
  power efficient, and ubiquitous. It's not just about faster chat
• 25:23
  bots or better image generators. No, it's about a fundamental
• 25:27
  shift in how all machines communicate.
• 25:29
  We are moving from a world where computers talk to each other via
• 25:32
  electricity to a world where they talk to each other via.
• 25:35
  Light. Light is the new standard for
• 25:37
  internal communication, not just for long distance.
• 25:40
  Exactly. Imagine self driving cars,
• 25:43
  factory robots, airplanes, even your home appliances.
• 25:46
  If they can all start communicating with each other
• 25:49
  with the speed and efficiency of light, the latency, the delay
• 25:53
  all but disappears. The power consumption drops
• 25:56
  dramatically. The bandwidth becomes
• 25:58
  effectively infinite for most applications.
• 26:00
  It's like upgrading the nervous system of the entire planet from
• 26:03
  copper wires to a fiber optics. That is a great way to put it.
• 26:06
  It's a planetary scale nervous system upgrade.
• 26:09
  So for our listeners, specifically the learners out
• 26:12
  there who love to stay ahead of the curve, what is the key take
• 26:15
  away from this deep dive? The take away is this.
• 26:18
  Don't just look at the smarts and don't just focus on the
• 26:21
  software or the LMS. When you read about the next big
• 26:24
  AI breakthrough or the next super chip from NVIDIA, ask
• 26:27
  yourself the boring question how is the data moving?
• 26:31
  Look for the plumbing. Look for the plumbing.
• 26:33
  Look at the optical links. That is where the physical
• 26:35
  constraints are, and in technology the biggest
• 26:38
  constraints are always where the biggest opportunities hide.
• 26:41
  That's where the next billion dollar companies are being
• 26:43
  built. That is fascinating.
• 26:45
  It's a reminder that even in our increasingly digital world,
• 26:49
  atoms still matter. You still have to build the
• 26:51
  physical thing that moves the photons from A to B.
• 26:54
  Absolutely. The cloud is not a cloud.
• 26:55
  The cloud is made of metal, silicon and glass running in a
• 26:59
  giant power hungry building. I want to challenge one part of
• 27:01
  this before we go. We talked a lot about the SpaceX
• 27:04
  DNA being a huge asset, but is there a potential downside to
• 27:08
  that culture? That's a fair question.
• 27:09
  What's the risk? SpaceX is famous for its move
• 27:13
  fast and break things ethos. They literally have a highlight
• 27:16
  reel of their rockets blowing up.
• 27:18
  Rapid unscheduled disassembly. Right.
• 27:20
  They test a failure, they build a rocket, launch it, watch it
• 27:22
  explode, learn why, and then build another one better in a
• 27:25
  few weeks. Can you apply that culture to a
• 27:27
  supply chain for Amazon or Google?
• 27:30
  If you ship a million bad transceivers to AWS, you don't
• 27:33
  get a second chance. Yo, you get sued and your
• 27:35
  company dies. Yeah, you're absolutely right.
• 27:38
  A rocket blowing up on a remote test stand is valuable data.
• 27:42
  A data center in Ohio going dark because your transceivers failed
• 27:46
  is an economic catastrophe for your customer.
• 27:48
  So that is the tension they have to manage.
• 27:50
  They have to keep the innovation, speed and the
• 27:52
  agility of SpaceX, but combine it with the Six Sigma 0 defect
• 27:57
  reliability of a traditional boring industrial manufacturer.
• 28:01
  And that transition from scrappy startup mode to trusted
• 28:05
  industrial supplier mode is where so many promising hardware
• 28:08
  companies fail. It's known as the Valley of
• 28:11
  Death. Scaling from the first working
• 28:13
  prototype to the first million units off the assembly line,
• 28:17
  that is where the wheels usually fall off.
• 28:19
  It is So the $50 million from Thrive Capital, that's not just
• 28:23
  for R&D, that's basically the fuel to build the bridge to get
• 28:26
  them across that valley before the money runs out.
• 28:28
  High stakes. Extremely high stakes, but as we
• 28:30
  said, the rewards on the other side are astronomical.
• 28:33
  OK, before we sign off, I want to leave the listener with a
• 28:35
  final thought, a bit of a provocation to chew on.
• 28:37
  Go for it. We talked about this lights out
• 28:40
  manufacturing. The idea of a fully automated
• 28:42
  factory running in the dark mesh is betting their whole company
• 28:46
  that they can pull this off in the USA.
• 28:48
  Very bold bet. If they succeed, if these three
• 28:52
  engineers from SpaceX can actually figure out how to mass
• 28:55
  roduce these incredibly complex recision electronics in Los
• 28:58
  Angeles and do it cheaper than the established layers in China,
• 29:04
  what does that mean? Does that prove that the whole
• 29:06
  narrative about US manufacturing being dead is wrong?
• 29:09
  Exactly, Is this the start of a renaissance, a blueprint for how
• 29:12
  to bring high tech manufacturing back?
• 29:15
  Or is it just a special exception?
• 29:16
  A1 off? That's only possible because the
• 29:18
  product is so strategically valuable for AI that you can
• 29:22
  justify the massive initial investment.
• 29:25
  That is the billion dollar question, isn't it?
• 29:27
  If they crack the code, maybe we see lights out factories for
• 29:30
  other complex electronics. Maybe some consumer electronics
• 29:32
  manufacturing comes back. But if they fail.
• 29:34
  That it just reinforces the dominant idea that complex
• 29:37
  hardware belongs in Asia and software and design belong in
• 29:40
  the US. Which is a very dangerous
• 29:42
  dichotomy if you care about supply chain security and
• 29:44
  national resilience. It absolutely is.
• 29:46
  And consider this as a final, final thought.
• 29:50
  We are moving toward a world where the speed of light isn't
• 29:54
  just a constant in physics textbooks, it is rapidly
• 29:57
  becoming the standard speed for internal machine thought.
• 30:01
  When the speed of thought literally equals the speed of
• 30:03
  light, things are going to get very interesting very quickly.
• 30:07
  Indeed they are. And on that note, we are going
• 30:10
  to wrap it up for today. It's been a pleasure digging
• 30:12
  into this one. Thanks for diving deep with us.
• 30:14
  This has been the deep dive. We'll see you on the next one.
• 30:17
  Stay curious.