0:00 Supabase just open-sourced a benchmark that runs AI coding agents against its own real-world tasks, and publishes the scores for everyone to see.0:09 That move connects directly to what we talked about in episode 143.0:13 Last week, it was the EU AI Act forcing an age of consequence with new laws.0:18 This week, you're seeing the consequence play out in the market—a demand for PROOF, not just promises.0:25 The age of hype is over.0:27 The age of execution has begun.0:29 Here’s the sweep of what’s moving today.0:31 First, the money is following the metal.0:34 Three separate funding rounds for robotics and physical AI just crossed a combined one-point-nine BILLION dollars.0:42 The key phrase you're hearing isn't "valuation." It's "deployment." The question is no longer how much a company is worth on paper, but how many robots it can actually get into the field.0:54 Companies like Unitree are reportedly making real progress with industrial customers, turning prototypes into products.1:02 This isn't science fiction.1:04 This is stackable, deployable, physical AI hitting the factory floor.1:08 Next, the entire conversation around AI cost is changing.1:12 For the last few years, it's been all about the GPU.1:15 Who has the most H100s.1:17 Who has the biggest training cluster.1:19 That's over.1:20 The new metric, the phrase you need to know, is "tokens per watt per dollar." It's a measure of INFERENCE efficiency.1:28 How much useful work can you get out of your hardware for every dollar you spend and every watt you burn?1:35 This isn't just about Nvidia anymore.1:37 It's about the entire physical infrastructure chain.1:41 Behind every chatbot response sits a stack of computing power, high-speed networking, advanced cooling systems, and data-center equipment that all need to work in perfect concert.1:52 The focus is shifting from the star performer—the GPU—to the entire orchestra.1:57 And that brings us to the sleeper conversation gaining momentum.2:02 It’s about chip design.2:03 Specifically, AMD.2:04 There's a growing, detailed argument making the rounds that AMD's chiplet architecture—their modular design—is perfectly suited for where AI is going next.2:15 The argument is that this approach gives them a fundamental advantage in cost, manufacturing yield, and flexibility.2:22 It’s a technical argument, but the implication is huge.2:26 As AI moves from a few massive training runs to trillions of daily inference requests, the economics of producing chips at scale becomes the dominant factor.2:36 The discussion is shifting from raw power to economic reality.2:40 And finally, there's the benchmark that started us off.2:44 Supabase Evals.2:45 It’s not just another leaderboard.2:47 By open-sourcing a test that runs agents like Claude Code and Codex against REAL, messy, everyday programming tasks, Supabase is sending a clear message.2:57 The lab is closed.2:58 The real world is the only test that matters now.3:02 So let's dive deeper into the two big shifts this week: the new economics of AI, and the relentless push into the physical world.3:10 They are deeply connected.3:12 Let's start with the money.3:14 That one-point-nine billion dollars flowing into physical AI isn't just another venture capital headline.3:21 It's a signal.3:22 For years, AI progress was measured by parameters and benchmarks on abstract tasks.3:27 Now, the value is shifting to deployment.3:30 To things that move, that build, that interact with the messy, unpredictable physical world.3:36 This is a much harder problem.3:38 And it changes where the innovation needs to happen.3:41 It's not just about a better model, it's about better sensors, more durable hardware, more efficient motors, and software that can handle the chaos of reality.3:52 When you see a company like Unitree, which makes quadruped robots, getting traction with industrial clients… that’s not a demo.4:00 That’s a manufacturing line.4:02 That’s a logistics warehouse.4:04 That’s AI with a job to do.4:06 And this push into the real world is what’s forcing the second, even bigger shift.4:11 The radical change in how we think about the cost of AI.4:15 For the past few years, the entire industry has been obsessed with training.4:20 The massive, power-hungry, eye-wateringly expensive process of creating a foundational model.4:26 Inference—the act of actually USING the model to answer a question or generate code—was seen as the cheap part.4:33 The after-thought.4:34 That assumption is now being completely dismantled.4:38 The new reality is something called Agentic AI.4:41 Autonomous agents.4:42 Think of fleets of AI workers that you give a task to, and they work on it 24/7.4:47 They don't just answer one query.4:49 They orchestrate tools, they reason, they move data, they run multiple processes in parallel to achieve a complex goal.4:57 And here's the thing.4:59 This kind of work… it doesn't just need GPUs.5:02 It needs a TON of CPU power for coordination, for running what are being called "agent sandboxes." It's a completely different kind of workload.5:11 One analyst on Twitter, MikeLongTerm, put out a detailed thread on this, projecting that agentic AI will push inference from a small fraction of AI compute to over ninety percent within two years.5:24 Ninety.5:24 Percent.5:25 So what does this all add up to?5:27 It means the entire economic model of AI is flipping on its head.5:31 The focus is no longer on the one-time cost of training a model.5:36 It's on the continuous, operational cost of running millions of agents, performing trillions of tasks, every single second of every single day.5:45 And when you're operating at that scale, tiny efficiencies matter.5:49 A one percent improvement in cost isn't a small win.5:53 It's a game-changing competitive advantage.5:56 This is why you're suddenly hearing about "tokens per watt per dollar." It's a ruthless measure of efficiency.6:03 It's why the conversation has expanded beyond the GPU to the entire data center stack—the networking, the cooling, the power delivery.6:12 Every single component in that chain is now a potential bottleneck, or a potential source of competitive edge.6:19 It’s a return to first-principles engineering.6:22 Now.6:22 This is where the argument around AMD gets really specific.6:26 For years, AMD’s CEO Lisa Su has been talking about this shift.6:30 She predicted that the opportunity would expand far beyond training and into inference, into enterprise use, into these agentic fleets.6:39 And she bet the company's architecture on it.6:42 The argument you're seeing online now is that AMD's chiplet strategy is the payoff to that bet.6:48 Instead of designing one giant, monolithic processor—which is expensive and difficult to manufacture—AMD builds its processors from smaller, modular "chiplets" connected together.7:00 Here’s why that matters NOW.7:02 First, cost and yield.7:04 Smaller dies have dramatically better manufacturing yields.7:07 When a defect happens on a silicon wafer, it only kills one small, relatively cheap chiplet, not a massive, expensive processor.7:16 You get less wasted silicon.7:18 You can turn partially defective chiplets into lower-cost products instead of throwing them away.7:24 AMD has said publicly that for high core-count chips, this design can cost roughly HALF of an equivalent monolithic chip.7:32 In a world where you need to deploy millions of CPUs to manage your AI agents, a fifty percent cost reduction is not a feature.7:40 It's a revolution.7:42 Second, scalability and flexibility.7:44 AMD uses the same core chiplet designs across its entire product stack, from consumer desktops to the most powerful server processors.7:53 This dramatically cuts down on design and manufacturing costs.7:57 It allows them to scale up core counts for massive server chips almost linearly, while competitors using older designs see costs explode exponentially.8:07 This is the key.8:08 The agentic AI future requires a different balance of compute.8:12 The old ratio of GPUs to CPUs in a data center is breaking.8:16 The forecast is moving toward a one-to-one, or even a one-to-two ratio of GPUs to CPUs.8:22 That means the demand for CPUs isn't being replaced by GPUs—it's being ADDED to by this new layer of agentic orchestration.8:30 You're looking at a future where monthly token consumption is measured in the quadrillions.8:36 One estimate I saw pegged it at 35 quadrillion tokens a month.8:40 That's a 160-times growth in just two years.8:43 That level of demand breaks old models of production.8:46 It requires a new way of thinking about building the engines of intelligence.8:51 The argument is that AMD saw this coming, and built the architecture for THIS world, not the last one.8:58 This isn't about fanboy-ism.9:00 This is about economics.9:01 It’s a structural argument about manufacturing, cost, and scale.9:06 And it connects everything we're seeing this week.9:09 The push for real-world deployment in robotics, the need for verifiable performance shown by Supabase Evals, and the relentless focus on the cost-per-token...9:19 they all point to the same conclusion.9:22 AI is industrializing.9:23 The period of blue-sky research and dazzling demos is giving way to a new era defined by engineering discipline, supply chain management, and brutal economic competition.9:35 The winners won't be the ones with the flashiest PowerPoints.9:39 They will be the ones who can execute at scale, reliably, and more efficiently than anyone else.9:45 This week wasn't about a new model launch.9:48 It was about the sound of the factory floor getting louder.9:52 It was about the shift from what AI could do, to what it actually costs to do it, a billion times a day.9:58 The questions have changed.10:00 And the companies that have the right answers are the only ones that will matter.
