0:00 NVIDIA and CoreWeave just brought the industry's first Vera Rubin AI supercomputer online.0:06 That's a system that was just a blueprint a few months ago, validated in June, and it changes the entire game.0:13 Last time, we talked about Alibaba dropping a two-point-four TRILLION parameter model.0:19 But the story today isn't about how big the model is.0:22 The story is about the factory you build it in.0:25 The very definition of a computer just changed.0:29 For years, the race was about the fastest chip.0:32 That race is over.0:33 It's obsolete.0:34 The new race is about the smartest SYSTEM.0:37 The Vera Rubin platform isn't a chip.0:39 It's a six-part symphony of silicon designed to work as one body.0:44 You have the Vera CPU, the Rubin GPU, the NVLink 6 Switch, the ConnectX-9 SuperNIC, the BlueField-4 DPU, and a Spectrum Ethernet Switch.0:53 They are all co-designed to crush the bottlenecks that cripple massive AI workloads.0:58 It's a direct assault on the limits of communication, memory, and compute.1:03 The new keyword is "goodput"—not just theoretical peaks, but reliable, sustained performance.1:10 The business model has flipped.1:12 You don't get a box of parts and a manual anymore.1:15 You get a validated, working AI factory, tested BEFORE it gets to you.1:20 And this isn't just one company's vision.1:23 This is an industry-wide earthquake.1:25 SK hynix released a report just this week saying the exact same thing.1:30 Their words: modern AI data centers can deliver "only when compute, memory, storage, networking, and power and cooling are seamlessly integrated into a single system." You want to see what that looks like in the real world?1:45 Microsoft's new Fairwater AI data center, which also went live in June, is the length of FIVE football fields.1:52 The money tells the same story.1:54 The market for AI data center chips is projected to jump from 123 billion dollars in 2024 to 207 billion this year.2:02 That's not growth.2:03 That is a vertical line.2:05 So, why are we building these cathedrals of compute?2:08 Because the AI models themselves have fundamentally changed.2:12 The old strategy was simple: just add more parameters.2:16 Make the model bigger.2:17 That strategy is now officially dead.2:20 A major academic survey from June makes it brutally clear.2:24 "Scaling alone," it says, "without appropriate architectural design...2:28 can actively lead to...2:30 diminishing returns." The new playbook is architecture.2:34 Specifically, Mixture-of-Experts, or MoE.2:36 Instead of one giant, monolithic brain trying to do everything, you have a team of specialized, smaller models.2:44 The system learns which expert is right for the job and routes the problem there.2:49 We saw the first signs of this last year with models like Meta's Llama 4 and DeepSeek-R1.2:55 They delivered top-tier reasoning and performance without needing a nation-state's budget to train.3:02 They reshaped the entire industry's cost assumptions.3:06 Now, that architecture is demanding an entirely new class of hardware to run at planetary scale.3:12 But this is where it gets bigger than just tech or business.3:16 The reason everyone is willing to spend hundreds of billions of dollars and build five-football-field data centers is that the goal of this entire field has shifted.3:27 We are past just building faster tools.3:30 One of the key papers driving this shift says AI has moved science "from automation to autonomy." Think about that.3:38 From machines that assist computation, to systems that learn, generalize, and explore on their OWN.3:44 We used to build tools to help scientists find answers.3:48 Now, we are building systems that can ask their own questions.3:52 Systems that can form their own hypotheses and design their own experiments inside vast, combinatorial spaces we could never navigate.4:01 The enormous energy bill—gigawatt-hours to train a single model—only makes sense when you understand the prize.4:09 The AI factory isn't just building a better chatbot.4:12 It's building a new kind of scientist.4:15 And that changes everything.
