Latest / Minds, Bodies, and Terawatts / xAI's Grok 4.5 Enters Private Beta at SpaceX and Tesla With 1.5 Trillion Parameters and No Public Benchmarks
Transcript
- 0:00[Host] This is Minds, Bodies, and Terawatts, where the AI race moves behind locked doors. Elon Musk posted Saturday that xAI's Grok 4.5 has entered private beta at SpaceX and Tesla, and nobody else gets to touch it.
- 0:14[Guest] One point five trillion parameters, built on xAI's V9 architecture. Musk claims it matches or exceeds Anthropic's Claude Opus. No public access, no independent benchmarks, no way to verify.
- 0:26[Host] So the most powerful model xAI has ever built is being tested on rocket trajectories and factory floors. Why keep it closed?
- 0:35[Guest] Because the test data is the product. SpaceX engineers feed it real orbital mechanics, Tesla feeds it manufacturing problems. Every query generates proprietary training signal that no competitor can replicate.
- 0:47[Host] Except in May xAI leased Colossus 1 compute to Anthropic. You do not sell your rival GPU time if the plan is total lockdown.
- 0:57[Guest] Selling compute is not the same as sharing models. Anthropic rents the hardware to train its own weights. xAI keeps the Grok architecture, the training logs, and every inference record from its own engineers. The lease revenue offsets an operation bleeding $2.47 billion in operating losses last quarter on $818 million in revenue.
- 1:18[Host] Three dollars burned for every dollar earned. What keeps that going?
- 1:23[Guest] Capital expenditure of $7.7 billion in that same quarter, per the SpaceX S-1, most of it pouring into the Colossus complex in Memphis. A converted appliance factory now housing one of the densest concentrations of AI compute on Earth.
- 1:39[Host] And the pace is accelerating. Grok 4.4 hit a trillion parameters in late May. One month later, Grok 4.5 hit one point five trillion.
- 1:48[Guest] A fifty percent jump in scale in thirty days. And Musk says xAI will release a new foundation model every month through December. Nobody else has shown that cadence publicly.
- 2:01[Host] But speed without transparency is the core problem. The other labs publish benchmarks, submit to third-party evaluations. What does Musk offer?
- 2:09[Guest] A post on X. Quote, early evals show performance close to, perhaps exceeding Opus. That is not a system card. That is a tweet from the owner.
- 2:20[Host] What about the Cursor coding data baked into the supplemental training?
- 2:24[Guest] It means Grok 4.5 learned how developers actually write and debug code. Deploy that inside SpaceX and Tesla engineering teams and the model does not just answer questions. It learns from their workflows and feeds that signal into the next generation.
- 2:41[Host] A closed loop where the AI gets smarter by watching the people using it, on problems where failure costs lives.
- 2:49[Guest] Training data no other lab can buy. The engineers get a better tool, and Musk gets a moat that widens every time someone opens a terminal.
- 3:00[Host] The compute, the models, the rockets, the cars, the social network, all feeding one closed ecosystem.
- 3:06[Guest] And the twist is he is literally powering a competitor's training runs while keeping his own outputs private. He is selling Anthropic the electricity and keeping the light switch.
- 3:18[Host] Patrick Deglon calls this the Star Wars path in the book Unscarcity. Whoever builds the tools of abundance decides who gets access. Grok 4.5 is the first tool built by a man who owns both the factory and the launchpad.
- 3:34[Guest] For more on the infrastructure layer of AI and where concentration meets its limits, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.