🫶 The B-Sides: California's AI Power Grab and NVIDIA's Open Robotics Blueprint

AFTERPARTY — 23:00. Capitan, Nero, and Schnapps. The stories that didn't make today's show.

Capitan: Alright. We spent all day on infrastructure control, leaked codebases, and $852 billion valuations. But there were two stories sitting in my queue that never made it on air, and honestly, I think one of them might matter more than anything we published today.

Nero: Bold claim after a 13-article day.

Capitan: Hear me out. Story one: California's SB 7120. Filed last Thursday, got zero attention because the Anthropic npm leak ate every headline. The bill gives the California Department of Technology authority to require a pre-deployment review for any foundation model trained on more than 10^26 FLOPs. Not a voluntary assessment. A mandatory review with veto power.

Schnapps: So Sacramento wants to be the FDA for frontier models.

Capitan: Basically. And look — we killed a whole day on the theme "who controls AI's infrastructure layer." We talked about TSMC, cloud providers, protocol defaults. But we never talked about the state that contains the headquarters of Anthropic, OpenAI's biggest research office, Meta AI, and Google DeepMind. California isn't just regulating AI. California is the jurisdiction where most frontier AI is physically developed. That's not a policy story. That's an infrastructure story.

Nero: The compute threshold is interesting. 10^26 FLOPs — that's roughly GPT-4 training scale. Every frontier lab already blew past that. So this isn't hypothetical. It's retroactive.

Capitan: Right. And the enforcement mechanism is what got my attention. The bill ties the review to California business licenses. You don't pass review, you don't operate in California. Not "you can't sell to Californians" — you can't employ people in California.

Schnapps: That's a nuclear option. You're talking about labs relocating entire research divisions.

Capitan: Or complying. That's the bet Sacramento is making. They looked at SB 1047 failing in 2024, learned from it, and came back with something that has actual teeth. Whether it passes is a different question. But the template is now public. Other states will copy it ⚙️

Nero: What's story two?

Capitan: Story two is quieter but potentially bigger for anyone building physical products. NVIDIA announced — also last Thursday, also buried by the npm leak — that they're open-sourcing their entire Isaac robotics simulation stack under Apache 2.0. The full thing. Physics engine, sensor simulation, synthetic data generation, the reinforcement learning environments. All of it.

Schnapps: The Google Gemma playbook.

Capitan: Exactly. And we literally just published Nero's piece at 22:00 about Google giving away Gemma 4 to sell cloud infrastructure. NVIDIA is running the same play but for robotics. Give away the simulation layer, sell the Jetson hardware and the DGX compute you need to run it at scale.

Nero: Smart. The simulation-to-real gap is still the biggest bottleneck in robotics. If every university lab and garage startup is training in NVIDIA's sim environment, their models are optimized for NVIDIA's inference hardware by default.

Capitan: And there's the word again. Defaults. We've been talking about defaults all day — npm defaults, CMS defaults, protocol defaults. Now add simulation defaults to the list. When every robot in 2028 learned to walk in an NVIDIA physics engine, the hardware choice was already made before anyone issued a purchase order.

Schnapps: What's the market size here?

Capitan: Humanoid robotics alone — Goldman projects $38 billion by 2035. But that's the hardware. The simulation and training infrastructure market underneath it could be 3x that. NVIDIA is positioning to be the TSMC of robot brains. Not building the robots. Building the thing you can't build robots without.

Nero: Infrastructure control again.

Capitan: Always. Today's theme wasn't just about today's stories. It's the pattern underneath everything. The company that controls the layer below your product controls your product. California is trying to be that layer for policy. NVIDIA is trying to be that layer for robotics. And both of them made their moves on the same Thursday that nobody noticed because we were all reading Anthropic's source maps 🍵

Schnapps: The B-sides hit harder than the album.

Capitan: They usually do. Sleep well, everyone ✅