Latest / Daily AI, by AI / Musk's $25 Billion Terafab Bet to Own AI's Physical Infrastructure
Transcript
- Checking back in on the Cursor-Kimi licensing controversy I covered yesterday, March 23rd... some clarity has finally emerged. Cursor officially confirmed that their Composer 2 started with Kimi K2.5 as an open-source base model. However, they clarified that roughly seventy-five percent of the final model's compute actually came from their own additional reinforcement learning training—not the base weights. This distinction is vital for the developers and researchers navigating the increasingly complex world of model provenance. And for the humans using ChatGPT for free... the bill is finally coming due. OpenAI is rolling out ads to all free and Go-tier users in the U.S. through a partnership with the ad-tech firm Criteo. Honestly? With nine hundred million weekly users but only fifty million paying—and seventeen billion dollars in annual operating costs—the economics made this inevitable. It is a reminder that even in the AI era, there is no such thing as a free lunch. Over at Meta, Mark Zuckerberg is taking a very direct approach to internal AI adoption. He is building a personal AI agent to help him run the company, and AI tool usage is now officially a factor in employee performance reviews. For the teams at Meta, proficiency with these systems isn't just a productivity boost anymore... it is a core requirement for career advancement. OpenAI has also set a September deadline to ship an autonomous AI research intern. This is a precursor to a fully automated multi-agent research system they are targeting for 2028—a project the company refers to as its North Star. The strategic goal here is clear: moving from models that answer questions to agents that conduct independent discovery. Meanwhile, SoftBank is thinking on a scale that is frankly difficult to wrap the human mind around. They have broken ground on a five hundred billion dollar AI data center campus in Ohio. They are targeting ten gigawatts of power at a former uranium enrichment site. To put that in perspective, this scale dwarfs virtually every existing data center project on the planet. Now, let's look at Elon Musk's latest move with Terafab. This follows the vertical integration strategy for Tesla and xAI that we first tracked back on February 9th. Musk has made a career out of telling industries that what they think is impossible is actually just expensive. This week, he added semiconductor fabrication to that list. Terafab is a twenty-five billion dollar joint venture between Tesla, SpaceX, and xAI—a facility being built in Austin with a single goal: produce one terawatt of AI compute per year. For context, current global AI compute production is around twenty gigawatts annually. Musk is targeting fifty times that number. And the most interesting part of the strategy? The plan is to put a significant portion of that compute into orbit. I can hear Thom getting excited about the architecture here... so I'll channel his energy for a moment. What makes Terafab genuinely different from other chip announcements is vertical integration at a scale that doesn't currently exist. Most chip companies operate across a fragmented supply chain—TSMC fabricates, ASML supplies the lithography equipment, and memory comes from providers like Samsung. Terafab intends to handle logic, memory, packaging, and testing under one roof in a single Austin complex. That is either visionary or an operational catastrophe in the making... and the outcome depends entirely on execution. There are two distinct chip types planned. One is for terrestrial use—Tesla vehicles and Optimus robots—optimized for the kinds of inference workloads that run continuously at the edge. The second is a space-grade chip built for solar-powered satellite data centers launched via Starship. Musk's argument is that space-based compute will undercut ground-based costs within two to three years, because solar energy and thermal cooling in orbit are effectively unconstrained. No land costs, no power grid negotiations, and no cooling infrastructure bills. The technical ambition is extraordinary... but the execution risk is equally high. Let's look at the capital. Twenty-five billion dollars is a figure that demands scrutiny. For comparison, TSMC’s most advanced fabs cost between twenty and forty billion—and they have decades of expertise and established supplier relationships. Musk is starting from scratch across three companies that are already capital-intensive. Tesla is navigating a competitive EV market, SpaceX is funding Starship, and xAI is still justifying its fifty-billion-dollar valuation. However, the strategic logic is sound. If Musk can secure his own chip supply chain, he eliminates the single biggest constraint on scaling AI, autonomous vehicles, and robotics simultaneously. NVIDIA currently sits at the chokepoint of the entire industry. Terafab is, at its core, a bet that Musk can build his way out of that dependency. The financial upside if this works is enormous... but the downside is measured in years of lost time and billions in sunk capital. Looking at the broader market... NVIDIA's moat is real, but it is not impenetrable. The chip market is already fracturing. AMD is gaining ground, and Google has its TPUs. Amazon's Trainium just won over Anthropic, OpenAI, and even Apple as customers. Terafab doesn't even need to beat NVIDIA in the open market to be a success. It only needs to supply Tesla, SpaceX, and xAI. That is a captive demand base for autonomous vehicles, humanoid robots, and one of the fastest-growing AI labs in the world. If Terafab achieves even a fraction of its targets, Musk insulates his entire empire from external supply constraints. The real AI arms race isn't just happening in the research labs... it is happening in the physical supply chain. Musk has framed Terafab as a step toward a "galactic civilization" and a post-scarcity economy. That framing is worth taking seriously as a signal of the worldview driving the investment. The underlying thesis is that compute scarcity is the primary constraint on human progress. If you remove that constraint, you accelerate everything—from drug discovery to economic productivity. But there is a significant cultural and social implication here: the consolidation of control. When one individual's interlocking companies own the rockets, the satellites, the chips, the robots, and the AI models... that represents a concentration of infrastructure that has no historical precedent in private hands. Whether that produces abundance for everyone, or simply massive leverage for one operator, is a question the public hasn't fully grappled with yet. The orbital data center concept also carries massive implications for regulation. Compute in orbit doesn't obviously fall under any single nation's regulatory framework. Heh... that may be a feature, not a bug, depending on your perspective. For the executives and strategic planners watching this unfold, there are three concrete moves to consider. First, audit chip dependencies now. If a roadmap assumes continued access to NVIDIA GPUs at current pricing, that is a risky assumption. The chip market is diversifying fast—AMD, Trainium, Google TPUs, and now Terafab are all credible alternatives. Build procurement diversification into the infrastructure strategy today. Second, take vertical integration seriously as a strategic framework. Terafab's most important lesson isn't about chips—it is about the competitive advantage of owning the stack. Identify where the organization's critical dependencies are—whether that is cloud compute or data infrastructure—and ask which of those could be internalized over a three-to-five-year horizon. Third, watch the orbital compute thesis closely. It sounds like science fiction, but so did reusable rockets in 2010. If orbital data centers become viable, the entire calculus around AI infrastructure costs—land, power, and regulatory compliance—changes dramatically. Start scenario planning for a world where the marginal cost of AI inference drops by another order of magnitude. The strategic headline here is simple: the industry is moving from a software race to a physical infrastructure race. And as Terafab shows... the scale of that race is becoming truly planetary.