Bittensor Is About To Break Its Own AI Record Again

Jesus Martinez walks through what’s been an unusually loud week for the ecosystem: record-chasing training runs, founder commentary that hits the core thesis, real product traction, and a few smaller stories worth tracking.

The headline framing comes from Const himself, who put it bluntly: “Cryptocurrency is the most powerful alignment tool for humans for compute.”

Who breaks the next training record

The 72B parameter LLM run from Templar back on March 10 is no longer the ceiling. Multiple teams are now racing past it:

• Macrocosmos / IOTA (Subnet 9) is targeting Orion 100B, a 100 billion parameter model designed to break Templar’s record with significantly better efficiency at scale.
• Chutes (via founder Jon Durbin) “for fun” kicked off a 176B parameter model run at ~140 steps, with plans to scale it potentially past a trillion tokens. A trillion-parameter decentralized training run would be genuinely historic.
• Subnet 3 is also building a fresh 10B architecture focused on long memory and reasoning.

The Const segment

A meaningful chunk of the video is footage from Const explaining Bittensor’s core thesis.

The eye-catching lines from the conversation:

• “Bitcoin and AI are tools of organizing matter.” Crypto aligns humans and compute. AI aligns neural network weights. Bittensor combines both to produce digital commodities, specifically, intelligence itself.
• “Everybody in the world will be sipping from the mental biases of a single company that controls AI” if decentralized AI doesn’t win.
• The internet analogy. Just like the open-source internet was once not open source, AI today is centralized. Bittensor’s bet is to do for AI what open standards did for the web.
• The Nvidia nod. Jensen Huang publicly mentioned Bittensor on the All-In podcast with Jason Calacanis, calling the Templar 72B accomplishment “impressive.” That kind of co-sign from the most important chip CEO in the world is not small.
• The “negative case” argument. People will only understand the value of decentralized AI clearly if it loses. The point of building it now is to make sure that doesn’t happen.

Real product traction

• Chutes was, at one point, serving the most open-source tokens of any provider in the world. Most users had no idea they were touching Bittensor — they just used the tokens.
• Targon’s compute is mostly being used by external customers, not just TAO miners. That’s the “seeped into the real world” signal Const himself flagged.
• Targon is now repurchasing its own alpha token using organic revenue, exactly the kind of flywheel that separates real subnets from emission-dependent ones.

Other notable updates from the week

• The Conviction upgrade went live on mainnet. Still controversial, still being debated. Jesus notes it’s too early to call good or bad, but it’s the most significant governance change of the year.
• Minos (Subnet 107) is presenting two works at the American Society of Human Genetics this fall.
• Stillcore Capital, one of the larger Bittensor funds, invested in Subnet 9.
• Swarm is improving wildfire and natural disaster detection via drones.
• Ditto Code is shipping a memory-focused agentic coding app.

Nova’s pitch from Proof of Talk

Jesus closes with a clip from Michaela, CEO of Nova: drug discovery as the most important investment category in the network; “our success will not just be measured in dollars, euros, or TAO, but in human life.” The Nova pitch in one line: end-to-end automated drug discovery with AI agents pushing requests to robotic labs. Not there yet, but the infrastructure is being built in the open.

Bottom line

The fundamental story keeps getting better. Multiple subnets are now chasing training records that would have been unthinkable a year ago. Real revenue is starting to flow back into TAO via buybacks. External users are touching Bittensor products without knowing it. And the founder himself is articulating the thesis to the world.

Full video below: