$TAOMy positions aren’t heavy, but I’ve been thinking about this project recently. Let me share how I see it today.
This past March, something happened that many people didn’t pay close attention to. More than 70 people, strangers to each other, with the game GPUs at home—no companies, no data centers—just gathered online and trained a 72-billion-parameter language model. It scored 67.1, roughly comparable to Meta’s Llama 2 70B, which was produced with a few hundred million dollars. The paper was published on arXiv, in black and white—go look it up yourself.
What is the point or meaning of this? Companies like OpenAI, Google, and Anthropic are valuable largely because they’ve effectively monopolized compute power. If you want to use it, you have to pay them—prices are set by them. What Bittensor wants to do is to open this up: organize idle GPUs around the world. Whoever contributes gets TAO. Those who perform poorly get kicked out—the protocol decides. This logic isn’t new, but distributed training that produced a 72-billion-parameter language model is genuinely impressive.
At All-In, Jensen Huang was asked repeatedly whether decentralized AI training really works. He pulled this up and said something like “a modern version of folding@home”—back then people used idle CPUs to help scientists calculate proteins; now people use idle GPUs to help the network run AI. He has a whole stack of AI businesses to manage and doesn’t need to stand on anyone’s platform, but he said it. TAO rose 15% that day, and trading volume was $677 million.
The institutions are moving too. Grayscale increased the weight of the $TAO in its AI fund from 31% to 43%, the largest single position, and it’s still running the spot ETF application. Intel’s subnet Targon signed a corporate contract with PwC France—the first subnet case where an external enterprise truly pays money, not that vague wording like “strategic partnership.”
One more detail that not many people mention: a Bittensor subnet recently beat Claude and Cursor on a SWE programming benchmark—spending less than $1 million. Meanwhile, Anthropic spent tens of billions.
Then in May, the network upgraded: subnet capacity increased from 128 to 256, and the incentive mechanism was adjusted too. TAO rewards started concentrating toward subnets with real usage; those nobody uses slowly just wait to die. In Q1, total AI services revenue was $43 million. Most of the money for Web3 AI projects comes from the tokens themselves, not from AI services. In this space, this number is definitely something you can point to.
But on April 10, the Covenant AI team announced it was exiting, and in the process cleared about $10 million $TAO—and the same day it dropped more than 20%. When one subnet team leaves, the entire network reacts like that; it doesn’t match the word “decentralization.” Later, co-founder Const said it himself: the incentive layer is still controlled by the core team, decentralization is something on the roadmap, and it hasn’t been reached yet.
There’s another number that has to be thought through: Q1 external revenue was $43 million, but the network’s total annualized subsidies are $52 million+. Most subnets keep themselves alive by getting the network to pay them; there are still very few customers that truly pay from outside. If this bottleneck can’t be plugged, a $3.5 billion market cap will be hard to sustain.
June’s new proposal Root Reborn wants to turn validators from passively collecting fees into actively placing bets: they choose which subnets to bet on, and returns compound in the form of subnet tokens—not everything is cashed out to create sell pressure. The logic is sound, and it’s still in the testnet, but whether it can be implemented is another matter. THORChain also announced it will natively integrate TAO, so that once it’s ready, cross-chain transfer can be done without trust.
In the end, it’s a $3.5 billion market cap buying a future that hasn’t happened yet. Endorsement from Jensen Huang, Grayscale adding to the position, and a PwC contract.
I’m mainly watching whether this external-payments side really takes off. If it does, it’ll absolutely soar; if it doesn’t, then all those things before are just stories.