TAO has run roughly 28% in a week, trading in the mid-$250s with a market cap near $2.8 billion, and the broader basket of Bittensor subnet tokens has swelled to a combined $1.5 billion as nearly every token in the ecosystem posts double or triple-digit gains over thirty days. The AI crypto sector as a whole pushed past $25 billion in market cap. The easy explanation, the one most outlets reached for, is that a U.S. export order restricting access to certain centralized AI models sent traders rotating into decentralized alternatives.
That catalyst is real. It is also the least interesting part of the story.
The Headline Catalyst Will Fade
Export-control headlines move crypto for days, sometimes a week or two. Anyone who traded the various chip-restriction cycles knows the rhythm: a policy shock creates a rotation, the rotation creates a narrative, the narrative creates a chart that looks unstoppable right up until the next headline reprices it. Treating a regulatory shock as the foundation of a thesis is how traders end up holding the top of a momentum move with no idea why they are still in it.
If the export order were the whole story, this would be a trade with a short shelf life. The reason Bittensor is worth a second look has almost nothing to do with what Washington did and everything to do with what a few dozen anonymous contributors did with commodity hardware.
The Milestone Buried Under the Price Action
Before the rally, Bittensor's Subnet 3 produced something the coverage mostly skipped: Covenant-72B, a 72-billion-parameter language model trained permissionlessly across the network by more than seventy contributors using ordinary internet-connected hardware. No single data center. No central coordinator renting out a cluster the size of a warehouse. A model in the same parameter class as serious commercial systems, assembled by strangers pooling spare compute.
Understanding why that matters requires sitting with the assumption it challenges.
Why centralized training was the moat
The entire investment case for the dominant AI companies rests on a physical fact: training a frontier model requires tens of thousands of high-end chips wired together in one place, drawing the power of a small city. That requirement is the moat. It is why a handful of firms with billions in capital and privileged chip access were supposed to own the field indefinitely. Distributed training over consumer-grade bandwidth was widely considered too slow and too lossy to produce anything competitive.
What Covenant-72B actually demonstrates
Covenant-72B is a proof, at meaningful scale, that the moat has a gap in it. If a 72-billion-parameter model can emerge from a permissionless swarm, the claim that frontier-scale AI must be centralized starts to look less like physics and more like a head start. That reframing is what a decentralized compute network is ultimately a bet on, and it is the thing actually being repriced underneath the export-order noise.
A regulatory shock gave the market a reason to look. A working 72-billion-parameter model is the reason the look might stick.
The LeveX Take
Here is the position, stated plainly: this rally has a durable core and a speculative shell, and most participants are buying the shell while telling themselves they own the core.
The durable core is the training-credibility milestone, which could re-rate decentralized compute over quarters if the results hold and replicate. The speculative shell is the export-driven momentum and the $1.5 billion of subnet tokens, many of which are illiquid, unproven, and priced as though Covenant-72B already guarantees the entire ecosystem's future. One impressive model is a proof of concept. It is not yet a moat of its own, and the gap between "this can work" and "this wins" is where a lot of capital tends to evaporate.
That two-sided reality is exactly what makes TAO a candidate for Multi-Trade. A trader who believes in the long-term decentralized-compute thesis can hold a core long on TAO, and on the same contract open a separate, smaller short sized to hedge the obvious near-term froth, each leg with its own leverage and stop. When conviction and caution point in opposite directions on a single asset, running both as distinct positions on the same pair lets you stay exposed to the structural bet while actively trading the volatility around it, instead of being forced to pick one and abandon the other.
The broader point for anyone trading the AI-token complex: separate the catalyst that got you to look from the development that justifies staying. Export controls are a headline. A model trained by a swarm is a thesis. Confusing the two is how this sector mints and destroys fortunes in the same quarter.
What Would Confirm the Thesis
The next leg of this story will not come from a policy announcement. It will come from whether Covenant-72B's results hold up under independent scrutiny, whether Subnet 3 can do it again with a larger or more capable model, and whether real usage flows to the network rather than pure speculation. Those are the signals that separate a structural re-rating from a rotation that reverses the moment the news cycle moves on. Hold them to dates and benchmarks rather than vibes.
For now, the most useful thing a trader can do is hold two ideas at once: the decentralized-compute bet is more credible than it was a month ago, and the price is moving faster than the proof. You can trade TAO on LeveX futures and the rest of your book on spot, and the Crypto in a Minute series unpacks how narrative catalysts and real fundamentals pull token prices in different directions.