The $60 billion resource hiding in space, and the start trying to mine it (feat. Matt Gialich, Astroforge) | E2268

March 27, 2026·1 hr 39 min·5 clips

Astroforge uses magnets to land on metal asteroids and lasers to extract platinum worth over $100 million per mission.

The show opens fast. Friday framing, sponsor reads, and off-duty banter take up the runway before the conversation turns to how expensive technical systems actually get built. Then the model talk gets sharper. The host presses on a 72 billion parameter training run across a tau subnet, asking how local miners ran optimized steps, compressed updates, shared them, and helped assemble the model. Cost becomes the reality check. The guest first sounds like he says two or three billion, then clarifies that the run cost about two to three million dollars in tau. That correction changes the whole claim. Now the conversation is about incentives, payouts, miner economics, and whether rollback waste can get smaller. The guest does not oversell it. He says the model is not frontier-level and is closer to Llama 2, which he calls the frontier from two years earlier. Still, the jump is the story. Going from 1.2 billion to 72 billion parameters in nine months gives the segment its charge: progress, not victory. The host keeps translating the mechanism as many optimized local steps, compressed, shared, and put back together into one model. That makes the technical stakes easier to hear. Underneath the jargon is a simple question: can a decentralized network buy useful compute in a different way than a company buying hardware directly. The ending wanders back off-duty. Lon points to Jim Cummings and recommends The Last Stop in Yuma County, while the host closes with subscription pushes and show housekeeping. Messy, yes, but useful. The best stretch is the live pressure test of a technical claim until cost, scale, and credibility start to match.

As heard by us

A technical This Week in Startups segment about a 72 billion parameter model, its cost, and how far it still sits from the frontier.

This Week in Startups keeps its focus on the mechanics of a technical breakthrough, following a 72 billion parameter model trained across the Tau subnet and pressing on cost, labor, and how the work was split up.

Read the full review in PlayNext →

Why you'd press play

You want a founder chat about training a 72B-parameter model, translated into startup math.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from This Week in Startups