Covenant AI: The Permissionless AI Lab

We view Covenant AI as the first credible demonstration that permissionless, end-to-end AI development can operate continuously at frontier-adjacent scale. The market is underestimating how quickly power, permitting, and interconnect constraints will dominate AI economics and overestimating the durability of purely centralized training models as scale increases.

Within this stack, we remain overweight Templar with high conviction. The timing is increasingly favorable as hyperscalers face interconnection delays and capacity constraints that are shifting AI training to a coordination problem, exactly where Templar's efficiency advantage matters.

Templar has crossed the hardest threshold in decentralized training, proving that bandwidth is no longer the dominant limiter for sustained pretraining at tens of billions of parameters. Covenant72B materially weakens the market’s objection that decentralized pretraining is currently infeasible. The system did not fail at scale, did not fragment under churn, and did not require privileged infrastructure. As decentralized training captures even a modest share of global AI spending, value accrual will concentrate in systems that already work in production. Today, Templar, as the core engine inside Covenant, is the clearest example.

Market setup is still early and mispriced relative to demonstrated execution. In this context, execution distinguishes Covenant as one of the only labs to have successfully completed end-to-end, frontier-scale pretraining runs on purely decentralized infrastructure. While competitors have largely focused on post-training, where lower computational barriers allow for more experimentation and benchmark optimization, Covenant has maintained a structural edge in the far more rigorous pretraining phase.

Bittensor valuations currently hinge on emissions capture as a proxy for utility, and top-10 subnets currently secure 1-10% emissions on ~25-450K TAO market caps. Templar’s SN3 is priced at ~$6.30 (~0.033 TAO, $25.8M FDV), a ~70% discount to sector leaders Chutes (SN64, $85M FDV) and Affine (SN120, $40M FDV), which repriced post-monetization, capturing 10.3% and 8.8% respectively of 3,600 daily TAO emissions (~$70K/~$60K). Templar, at 2.43% of daily emissions (~$17K), trades at ~4.2x FDV-to-Emissions vs. peers' 1.8-3.3x despite proven execution metrics. This valuation gap persists as the market categorizes decentralized pretraining as an experimental hurdle, ignoring the empirical proof provided by the completed Covenant72B model.

If decentralized AI systems meaningfully capture 5% of global training spend, the valuation paradigm for Bittensor subnets must shift from emissions-proxy to cash-flow-multiple. By 2030, global AI infrastructure spend is projected to reach $394B (a 19.4% CAGR). Even assuming a conservative 5% penetration rate for decentralized solutions the DeAI market represents a $20B annual opportunity. If Covenant maintains its lead in permissionless pretraining, capturing a 20% share of these workloads, and Templar extracts a 10% coordination fee, the network’s top line reaches $400M per year. Applying a 20x P/S multiple yields a terminal valuation of $8B.

For Templar, we see a viable path as Crusades completes, emissions resume under the new incentive regime, and the next full-scale training run launches, further de-risking coordination at frontier-adjacent scale. Templar’s market cap sits at ~$25.8M, while other functional subnets can clear meaningfully higher caps. Given Templar’s differentiation, an alternate path exists in licensing its SparseLoCo orchestration software to neoclouds and hyperscalers, enabling them to run efficient distributed training across their own heterogeneous GPU fleets rather than operating a decentralized network themselves. We nonetheless see an obvious path to mean reversion toward higher-quality peers as the market reprices “can it work?” into “how big can it get?” Liquidity has improved materially, with $28.7M of SN3 in the Alpha/TAO pool now supporting institutional entry at current scale. Technical catalysts are building, positioning Templar as one of the clearest mispricings among active Bittensor subnets.