The road to a HERO RUN.
The point of Hero Run is to get open models trained and released, as many as possible, and eventually one at the frontier. We don't have to build a lab to do it. We fund the open-training networks already doing the work. That is genuinely hard and getting harder, so here is the honest path.
The engine is live and the treasury grows with every run and trade. Each phase is funded by the volume the one before it generates, up to a permanent ~$100M endowment for open AI.
The frontier isn't that accessible.
Pretraining costs are rising, not falling. That is the whole reason a fundraising engine like this needs to exist.
One HERO RUN: a 5T-A100B model on 50T tokens, GB300-class hardware. Roughly double with reinforcement learning.
Companies on earth that could even justify that spend six months ago.
What a new lab must raise to reach the frontier with low risk. Only a few dozen exist.
Frontier training keeps getting more capable and more expensive. The biggest model you can train for under ~$20M all-in keeps growing, and the ceiling above it keeps rising: more is possible every year, and it costs more every year. Only a few thousand companies could justify a run at that scale, and only a few dozen labs have raised enough to keep doing it.
We fund the labs already doing it.
The treasury's job is not to hoard. It is to deploy fees as compute and research funding to the open-training ecosystem: buying GPU time on markets like Vast.ai, RunPod, and Lambda, backing open training runs, and funding the environments and datasets they need. The goal is simple, put out as many open models as possible.
The archetype of the work these fees are meant to support: open, decentralized training.
- Distributed training of open models: INTELLECT-2 (32B, distributed RL) and INTELLECT-3 (100B+ MoE).
- A global compute marketplace: H100 / H200 / B200 / B300 across 50+ datacenters.
- Open frameworks: Prime-RL, Verifiers, Sandboxes, all public.
- 2,500+ community RL environments and the open SYNTHETIC-2 dataset.
Prime Intellect is independent and named here as an example of the ecosystem, not a partnership or endorsement. Buying their compute, funding their runs, and backing peers like them is exactly how the treasury turns fees into open models.
Where the compute comes from. The treasury sources GPU where it is cheapest and most fungible: marketplaces like Vast.ai, on-demand clouds like RunPod and Lambda, and reserved clusters (CoreWeave-scale) for larger runs. More providers get added as the treasury grows.
Usage today, a frontier run at the end.
The product is live. Every run spends $HERO, every spend is a trade, and 0.665% of it lands in the treasury on-chain. Opt-in runs also log training data. The flywheel is turning from a standing start.
Route the treasury into the first open model built from this loop, partnering with decentralized-training networks for the compute and using the dataset the app collects. Small, useful, and shipped open. Proof the mechanism produces a model, not just a chart.
Deploy fees as compute and research grants across the open-training ecosystem, so many open models get trained and released, not one. The gateway's routing margin and the trading fees both feed this. More volume, more runs we can back.
Help fund a frontier-scale pretraining run. On today's math that is about $10M of compute, and roughly double with RL. Fully open weights. This is the run the whole arcade exists to fund.
A permanent fee stream funding open compute and research, so the frontier always keeps a public option. Reaching it with low risk takes $100M+, and only a few dozen organizations can. We want that capacity to belong to everyone.
Milestones, not dates. Each phase depends on the volume the one before it generates. Ambitious by design, and not a guarantee.
The run is funded by use.
Every model you run and every $HERO you buy pushes the treasury toward Phase 1. That is the whole mechanism.