Latest / Elon Musk Podcast / AI UPDATE: The AI Industry Is a $300 Billion House of Cards Built on Stolen Books and Broken Promises

Transcript

AI UPDATE: The AI Industry Is a $300 Billion House of Cards Built on Stolen Books and Broken Promises · Elon Musk Podcast

• 0:00
  Welcome back to the AI update at Stage 0.
• 0:02
  Glad to be here. So it is mid February 2026 and
• 0:08
  look, if you have been trying to keep your head above water with
• 0:11
  a new cycle this week, you are probably exhausted.
• 0:14
  Yeah, completely drained. Because we aren't just talking
• 0:16
  about a busy week here, we are talking about a week where the
• 0:19
  tectonic plates of the entire industry just shifted.
• 0:24
  It really, really does feel that way.
• 0:26
  You know, for the last few years, we've basically been in
• 0:28
  this what you could call the chatbot era.
• 0:31
  You type a prompt, you get a paragraph, maybe a recipe or an
• 0:34
  e-mail draft. But that era, it's officially
• 0:37
  over. As of this week, we have fully
• 0:40
  entered the era of the industrialization of
• 0:42
  intelligence. Industrialization is definitely
• 0:45
  the right word because looking at the stack of sources we have
• 0:47
  for this deep dive, we aren't just looking at software updates
• 0:50
  anymore. We're looking at GW scale
• 0:52
  infrastructure, nation states fighting over processor
• 0:55
  allocation, and, well, valuations that are starting to
• 0:57
  look like the GDP of small European countries.
• 1:00
  Yeah, the scale is just unprecedented.
• 1:02
  And at the center of this storm, we have this massive high stakes
• 1:06
  dual between the two Titans of the industry, Open AI and
• 1:10
  Anthropic. That is the headline, and to
• 1:13
  really make sense of this week for you, we need to pull apart 3
• 1:16
  distinct threads from our sources that all just collided
• 1:20
  at one. First, you have the brains.
• 1:23
  So the release of these new reasoning models like GPT 5.2
• 1:26
  and Claude 4.6, these things don't just talk anymore, they
• 1:30
  think. Right, which is a huge paradigm
• 1:32
  shift. Exactly.
• 1:33
  Then second you have the Braun, the physical reality of AI.
• 1:37
  We are talking about multi billion dollar data centers with
• 1:41
  sci-fi names like Stargate and Project Rainier.
• 1:44
  Consuming insane amounts of electricity.
• 1:46
  Literally enough to power millions of homes.
• 1:48
  And finally, and honestly, this is the part that sticks with me
• 1:51
  the most from the research, the conscience.
• 1:53
  Yeah, the human cost. Because while the money is
• 1:56
  flowing and the tech is scaling up, the human cost is becoming
• 1:59
  impossible to ignore. We saw a wave of resignations
• 2:02
  this week from the top safety leads at these labs, effectively
• 2:05
  warning us that the world is in peril.
• 2:07
  That part is, yeah, it's heavy and we're going to get there, I
• 2:11
  promise. But let's start with the shiny
• 2:13
  new toys, the brains. It felt like an absolute blitz
• 2:17
  this week. Open AI really came out
• 2:19
  swinging. They absolutely did.
• 2:20
  They have officially retired the legacy GB 24-O models.
• 2:24
  That entire architecture is now considered completely obsolete.
• 2:27
  Wow. Yeah, the new flagship is GPT
• 2:30
  5.2. But what is really crucial here
• 2:33
  is how they've split the product line.
• 2:36
  They've basically acknowledged that chatting and thinking are
• 2:38
  two entirely different sports. OK, help me unpack this a bit
• 2:41
  because I opened my app this morning and I noticed I have
• 2:43
  choices now. I've got instant and I've got
• 2:45
  thinking. What is actually happening under
• 2:47
  the hood there? Is it just, you know, a
• 2:49
  marketing rebrand? It's definitely not just
• 2:50
  marketing. Think of it kind of like the
• 2:52
  human brain. Psychologists often talk about
• 2:54
  System 1 and System 2 thinking. Right, Daniel Kahneman stuff.
• 2:57
  Exactly. So system 1 is fast instinctive
• 3:00
  thought, like knowing 2 + 2 is 4 or finishing a really common
• 3:04
  sentence. That is the instant model.
• 3:06
  It's designed for conversation. It's warm, fast, extremely low
• 3:10
  friction. OK.
• 3:11
  And System 2. System 2 is slow complex
• 3:15
  reasoning. It's what you do when you sit
• 3:17
  down to solve a complex math problem or plan a multi city
• 3:21
  travel itinerary. That's what the thinking toggle
• 3:24
  does. Open AI as a GPT 5.2 thinking
• 3:27
  model actually pauses. It just sits there.
• 3:29
  It sits there, it explores different logic paths, it tests
• 3:32
  hypothesis internally, and only then does it answer you and.
• 3:35
  This is where you get that specific toggle right, the
• 3:37
  settings for standard, light and extended.
• 3:40
  That's the one. And that extended setting is the
• 3:42
  real deal. We are seeing reports right now
• 3:44
  from the physics community that GPT 5.2, using that extended
• 3:48
  thinking time, reportedly derived a novel result in
• 3:51
  theoretical physics just days after it was released.
• 3:54
  Wait, hang on, Novel is in? It wasn't in his training data,
• 3:58
  it actually figured it out from scratch.
• 4:00
  That is the claim. It is not just regurgitating A
• 4:02
  textbook, it is computing new logic paths to find new
• 4:06
  information. That is a fundamental shift from
• 4:08
  just predicting the next word. It's basically the difference
• 4:11
  between a parrot and a professor.
• 4:13
  That is wild, but they didn't stop there either.
• 4:15
  They also dropped something specific for the coders back on
• 4:18
  February 12th. GPT 5.3 codecs spark.
• 4:22
  Terrible name. Oh, awful name but incredible
• 4:25
  technology. This is their dedicated agentic
• 4:28
  coding model. Agentic, we keep hearing that
• 4:31
  word thrown around everywhere in these sources for the non
• 4:33
  engineers listening. How is that actually different
• 4:36
  from just asking ChatGPT to say write me a Python script?
• 4:40
  It's a great question. Generative AI.
• 4:42
  The old stuff is a tool you talk to.
• 4:44
  You say write code, it writes code.
• 4:46
  You paste it in, it breaks, you paste the error back.
• 4:48
  It's a back and forth ping pong match.
• 4:50
  A genetic AI is a worker. You give it a goal.
• 4:53
  So I'm basically handing it a job description.
• 4:55
  Basically, yes. Say hey, this software has a bug
• 4:58
  in the login sequence, find it, fix it, write a test to prove
• 5:01
  it's fixed and deploy the update, and it goes off.
• 5:04
  And does that work entirely independently?
• 5:05
  It navigates the file systems, runs the tests, acts just like a
• 5:09
  human engineer would. And this new model is actually
• 5:11
  good at doing that it. Scored a 56.4% on the SWE bench
• 5:16
  Pro. Yeah, that is the gold standard
• 5:18
  benchmark for software engineering.
• 5:20
  Now 56% might sound a little low to you out of 100, but previous
• 5:24
  models were stuck in the single digits or low teens.
• 5:28
  This thing isn't just a coding helper anymore, it is performing
• 5:31
  at the level of a junior developer.
• 5:33
  So Open AI is pushing this personality angle for regular
• 5:36
  users, making the AI feel warmer with the instant model, while
• 5:40
  simultaneously building these absolute autonomous powerhouses
• 5:43
  for the pros. That is exactly the strategy
• 5:45
  bifurcation. But, and here's where the
• 5:48
  narrative gets really complicated, and Dropik didn't
• 5:50
  just sit back and watch them do this.
• 5:52
  No, they did not. They launched a massive counter
• 5:54
  attack mid month with Claude Sonnet 4.6 and Opus 4.6.
• 5:59
  And from what I'm seeing in the developer forums and the data
• 6:01
  we've pulled, the builders are actually flocking to Claude, not
• 6:04
  open AI. The stats completely back that
• 6:06
  up. This is probably the biggest
• 6:08
  untold story of 2026 so far, and Thropic has effectively capture
• 6:12
  the developer market. They now hold 54% of the code
• 6:16
  generation market share. Compared to Open A is 21%.
• 6:19
  Exactly. That is a massive, massive flip.
• 6:22
  It really is. Open AI has the Super Bowl
• 6:24
  commercials and the cultural fame, but the people building
• 6:28
  the actual software are using Claude.
• 6:30
  Why is that? It comes down to reliability and
• 6:33
  the underlying architecture. Let's look at Opus 4.6.
• 6:37
  Well, Open AI gives you that manual toggle.
• 6:39
  We talked about choosing how hard the model thinks.
• 6:43
  Anthropic introduced what they call adaptive thinking.
• 6:46
  So you don't have to tell it to think hard, it just knows.
• 6:49
  It autonomously decides based on the complexity of the prompt.
• 6:52
  Opus 4.6 determines the depth of reasoning required.
• 6:55
  It completely removes the friction for the user.
• 6:58
  But the real killer app here is Sonet 4.6 which is their mid
• 7:02
  tier model. It has a massive context window,
• 7:04
  1,000,000 tokens and it is purpose built for those exact
• 7:08
  agents we just mentioned. So if I want an AI to read an
• 7:10
  entire massive code base, understand the history of the
• 7:13
  project, and then work on it autonomously for three days, I
• 7:16
  use Clot. Basically, yes.
• 7:17
  The sources strongly suggest Sonet is the premier engine for
• 7:20
  these long horizon tasks. So we really are moving from
• 7:23
  tools at chat to tools that do work.
• 7:25
  That is the big shift. But here is the catch.
• 7:28
  That shift from chatting to thinking, from generating to
• 7:31
  acting, requires A staggering amount of computing power.
• 7:35
  Which perfectly brings us to the second thread, the Braun.
• 7:38
  The gigawatts. I love this part of the deep
• 7:41
  dive because it makes the digital world feel so viscerally
• 7:44
  physical. We keep hearing about these
• 7:47
  valuations. Open AI targeting over $800
• 7:50
  billion. Anthropic at 380 billion.
• 7:54
  But they aren't just hoarding this cash in a giant vault, they
• 7:56
  are pouring it into concrete, steel, and massive energy grid
• 8:00
  upgrades. They have absolutely no choice.
• 8:02
  The industrialization of intelligence requires industrial
• 8:05
  scale electricity. You cannot run a thinking model
• 8:07
  on a laptop. Let's talk about Open AI's
• 8:09
  Stargate project. Sargate.
• 8:11
  It sounds like a secret government operation from a 90s
• 8:14
  sci-fi movie. It essentially is a government
• 8:16
  level operation, just corporate run.
• 8:18
  This is a $500 billion initiative targeting 10
• 8:21
  gigawatts of AI compute. 10 gigawatts.
• 8:24
  Just to give you a sense of scale, 10 gigawatts is roughly
• 8:26
  what you would need to power the entirety of New York City.
• 8:29
  And this isn't just a proposal on a slide deck anymore.
• 8:31
  They have sites actually up and running right?
• 8:33
  Yes, the flagship US site in Abilene, TX, which they built
• 8:37
  with Oracle, is already online. They've got racks and racks of
• 8:41
  NVIDIA GB 200 chips just humming away.
• 8:44
  But what's really fascinating, politically speaking, is the
• 8:47
  global expansion side of that. The UAE deal.
• 8:50
  Exactly. You've got Stargate UAE in Abu
• 8:52
  Dhabi, which is a partnership with G42, Oracle and SoftBank.
• 8:57
  That's a one GW cluster going online in Q3 of this year.
• 9:01
  This directly ties into that concept you've mentioned before,
• 9:03
  Sovereign inference, which frankly sounds incredibly
• 9:07
  dystopian. What does that mean exactly in
• 9:09
  this context? It means computing power is
• 9:11
  fundamentally becoming a geopolitical asset, like oil
• 9:14
  reserves or freshwater nations like the UK.
• 9:17
  With Stargate UK, Norway, the UAE, they are realizing that if
• 9:20
  they don't have domestic Stargate level infrastructure,
• 9:23
  they're going to be forced to rent intelligence from the US or
• 9:26
  China forever. So sovereign inference basically
• 9:28
  means owning the means of production for intelligence
• 9:31
  securely within your own borders.
• 9:33
  Precisely. If you don't own the compute,
• 9:35
  you don't own the future. It's a brand new arms race, but
• 9:38
  the ammunition is compute now. Anthropic is taking a slightly
• 9:42
  different approach here. They aren't just universally
• 9:44
  doubling down on NVIDIA chips like Open AI is.
• 9:47
  No, and I really want to dig into this because it's a huge
• 9:50
  divergent Anthropic has essentially hitched its wagon to
• 9:53
  Amazon and AWS. They are building Project
• 9:56
  Rainier in Indiana. Rainier Another massive imposing
• 10:00
  name for a massive project. The key detail here is the
• 10:03
  hardware. They aren't using NVIDIA chips
• 10:05
  at all. They are using Amazon's custom
• 10:07
  Tranium 2 chips. It is the world's largest non
• 10:10
  NVIDIA AI cluster. We are talking 500,000 chips
• 10:14
  right now, scaling to 1,000,000 by late 2025 or early 2026.
• 10:18
  This is a monumental strategic bet.
• 10:21
  But why does the chip brand matter to the listener?
• 10:23
  Is it just a Pepsi versus Coke thing?
• 10:25
  No, it's much, much more than just branding.
• 10:27
  It is all about supply chain bottlenecks.
• 10:30
  Right now. The entire planet is fighting
• 10:32
  tooth and nail for NVIDIA chips. There is a wait list that is
• 10:35
  effectively years long. By using Amazon's custom Tranium
• 10:38
  chips, Anthropic is betting heavily on availability and
• 10:41
  efficiency. OK, so while Open AI is stuck
• 10:44
  waiting in line for NVIDIA to ship hardware, Anthropic and
• 10:48
  Amazon can just print their own. Chips.
• 10:49
  Exactly. If they can scale faster simply
• 10:52
  because they aren't bottlenecked by the global supply chain, they
• 10:55
  might actually be able to overtake open AI on sheer raw
• 10:58
  compute volume, even if the individual chips are slightly
• 11:01
  different architecturally. It's a massive diversification
• 11:04
  strategy. So we have the brains getting
• 11:06
  exponentially smarter, the brown getting massively bigger, but we
• 11:10
  also have the rules trying to catch up, or at least scrambling
• 11:13
  to. We saw a massive legal
• 11:15
  settlement this week that sort of flew under the radar amidst
• 11:19
  all the flashy tech news. The Barts V Anthropic case.
• 11:22
  We really cannot overstate this. This is historic.
• 11:25
  Historic and incredibly expensive.
• 11:27
  Anthropic agreed to pay $1.5 billion.
• 11:30
  That is the largest US copyright settlement in history.
• 11:33
  It is huge and we need to be very clear about why they
• 11:35
  settled the plaintiffs. The authors didn't just argue
• 11:39
  the vague idea that training on books is bad, they specifically
• 11:42
  alleged the anthropic train on shadow libraries.
• 11:45
  Shadow libraries. We're talking about straight up
• 11:46
  pirate sites, right? Yes, huge illegal collections
• 11:50
  like libgen and Palimi. The argument wasn't the
• 11:54
  philosophical is AI learning fair use debate.
• 11:57
  The argument was you specifically downloaded stolen
• 12:00
  pirated files to build your product.
• 12:02
  It bypassed the hazy copyright debate and went straight for the
• 12:05
  theft allegation. So Enthropic writes a check for
• 12:08
  1.5 billion. Does this solve the copyright
• 12:12
  issue for AI? Can these companies just pay a
• 12:15
  massive fine and keep going? That is the catch.
• 12:18
  The settlement only covers past conduct, everything pre August
• 12:22
  2025. It absolutely does not set a
• 12:24
  legal precedent for future licensing or liability for AI
• 12:28
  outputs. It doesn't say you have to pay
• 12:30
  authors for every output from now on.
• 12:32
  It basically just says, OK, we messed up using those specific
• 12:35
  pirated files, here's a billion and a half dollars of go away
• 12:38
  money. So it wipes the slate clean, but
• 12:39
  it doesn't actually build a new road for how to do this legally
• 12:43
  tomorrow. Not at all.
• 12:44
  And meanwhile, the federal government is basically trying
• 12:46
  to bulldoze the road for them. Right, with the executive order,
• 12:49
  yeah. We had that US executive order
• 12:51
  back in December 2025 that is basically aimed at pre empting
• 12:55
  state level AI laws like the Colorado AI Act.
• 13:00
  The goal is to create what they call a minimally burdensome
• 13:03
  national framework. Right.
• 13:04
  And to be clear, reporting on this impartially, the federal
• 13:07
  government's approach seems focused on ensuring the US
• 13:10
  maintains its lead in things like the Stargate race.
• 13:13
  They're actively trying to prevent a patchwork of local
• 13:15
  safety regulations from slowing down the deployment of these
• 13:18
  massive data centers. Which is a perfect segue to the
• 13:21
  final and honestly most unsettling segment of this deep
• 13:24
  dive. The conscience.
• 13:26
  Because for all this unprecedented money and all
• 13:28
  these technical breakthroughs, the people actually building
• 13:30
  these systems sound, well, terrified.
• 13:34
  The vibe shift inside the major labs does seem palpable despite
• 13:37
  the skyrocketing valuations. The mood is really dark.
• 13:41
  We saw a wave of very high profile exits this February from
• 13:45
  Open AI, Anthropic and X AI. And we need to emphasize these
• 13:49
  aren't your standard. I'm moving on to explore new
• 13:51
  challenges. Corporate resignations.
• 13:52
  These are blaring warning shots. Let's look at Marine Egg Sharma
• 13:56
  from Anthropic. He posted his resignation letter
• 13:59
  publicly on X. He didn't thank the team for the
• 14:02
  great memories and free snacks. He literally stated quote the
• 14:06
  world is in peril. He did.
• 14:08
  And remember, Sharma was a top safety lead.
• 14:10
  His entire job was to stop these highly capable systems from
• 14:13
  going off the rails. In his letter, he cited
• 14:15
  interconnected crises and explicitly stated he felt
• 14:18
  immense pressure to prioritize product rollout over safety
• 14:21
  protocols. He even referenced Einstein's
• 14:24
  regret about his role in the atomic bomb courting Einstein,
• 14:27
  saying I would have become a watchmaker, that is.
• 14:30
  I mean, that's chilling coming from a guy who literally knows
• 14:32
  the code inside out. It's devastating he is leaving
• 14:35
  the most lucrative industry on earth to go become invisible and
• 14:38
  study poetry that tells you volumes about his confidence in
• 14:41
  the current trajectory of the technology.
• 14:43
  And the crazy thing is, Anthropic heavily brands itself
• 14:46
  is the safe option. They are the constitutional AI
• 14:50
  company. That's the glaring contradiction
• 14:52
  here, because looking at our sources, anthropics own internal
• 14:55
  reports for these brilliant new models, the ones we were just
• 14:58
  praising for their adaptive thinking.
• 15:01
  They show elevated susceptibility to misuse.
• 15:05
  We're talking about the models potentially helping bad actors
• 15:08
  develop chemical weapons. The tech is just getting
• 15:10
  stronger much faster than the safety guard reels can be built,
• 15:14
  and it is certainly not isolated to Anthropic.
• 15:17
  Zoe Hitzig just left open AI and wrote a absolutely scorching
• 15:21
  op-ed in the New York Times. Right, But her critique was
• 15:24
  slightly different. It was more economic.
• 15:26
  She talked about the introduction of ADS and the
• 15:28
  complete dissolution of their mission alignment team.
• 15:31
  She argued that the economic incentives are completely
• 15:33
  overriding the original mission to benefit humanity.
• 15:36
  She called it the Alignment Innovation Paradox.
• 15:39
  It's such a powerful concept. Her core argument is that the
• 15:42
  economic engine, the desperate need to justify an $800 billion
• 15:46
  valuation, the massive cost of building A10 GW store gate, is
• 15:51
  inherently incompatible with safety.
• 15:54
  As these models become more agentic, as they start executing
• 15:56
  complex tasks entirely on their own, they naturally become
• 16:00
  harder to align and control. And the more billions you pump
• 16:03
  into the infrastructure, the harder it is to hit the brakes.
• 16:05
  Exactly. You simply cannot spend $500
• 16:08
  billion on data centers and then turn around to your investors
• 16:11
  and say, hey, let's pause deployment for six months to
• 16:13
  double check the safety metrics. The train has already left the
• 16:16
  station and it is accelerating rapidly.
• 16:19
  The financial momentum is just too massive to arrest.
• 16:22
  So let's try to recap this incredibly fast moving week for
• 16:25
  you. We have thinking models that can
• 16:27
  derive novel theoretical physics and write enterprise grade code
• 16:31
  better than most junior human engineers.
• 16:33
  We have physical infrastructure projects consuming gigawatts of
• 16:36
  power, reshaping the energy grid in global geopolitical
• 16:39
  alliances. And we have the largest
• 16:42
  copyright settlement in U.S. history, officially
• 16:44
  acknowledging that the very foundation of this incredible
• 16:47
  tech was built on pirated data. And finally, the actual
• 16:52
  architects of the safety systems are fleeing the building,
• 16:55
  telling us the world is in peril, and going off to write
• 16:58
  poetry. It really is a moment of extreme
• 17:01
  cognitive dissonance. We are building arguably the
• 17:04
  most powerful technology in human history.
• 17:07
  We are physically rewiring the planet to power it, and the very
• 17:11
  people who understand it best are the ones who are the most
• 17:13
  worried. It really makes you wonder about
• 17:16
  the long term stability of this stage zero we find ourselves in
• 17:19
  because if the infrastructure is being built to last for decades,
• 17:22
  the cement is poured, the chips are installed, but the people
• 17:26
  who understand the catastrophic risks feel they can no longer
• 17:29
  stop the train, what does that tell you about where the tracks
• 17:32
  are actually leading, that is? Exactly the question we all need
• 17:35
  to sit with this week. Well, on that cheery note, thank
• 17:37
  you for joining us on this deep dive.
• 17:39
  We'll be watching the news so you don't have to see you next
• 17:41
  time.