A team of AI engineers recently fed a full soccer match into one of the world's most celebrated AI models and asked it to annotate the footage. The machine charged them about $5,000 in comput
A team of AI engineers recently fed a full soccer match into one of the world's most celebrated AI models and asked it to annotate the footage. The machine charged them about $5,000 in computing fees and returned data so useless it scored zero on their benchmark. A specialized model built by a small independent team did the same job for $10, in two minutes, and got it mostly right.
That experiment, described to TheStreet Roundtable by the founder who ran it, is a small data point in a much larger fight: whether artificial intelligence will be owned by a handful of trillion-dollar companies that rent it back to everyone else, or built in the open by thousands of independent teams.
The open side of that fight has a headquarters of sorts in Bittensor, a blockchain network powered by the TAO token that runs like a marketplace for AI: independent teams operate specialized projects called subnets, and contributors earn token rewards for producing the best results.
Three of its leading builders and advisers laid out their case to TheStreet Roundtable, and it rests on three claims.
The first claim is commercial. The famous general-purpose models, for all their benchmark scores, fail in the places businesses actually need them.
The soccer experiment above belongs to Max Sebti, founder of Score, who has worked in AI data since 2018, first collecting data for AI labs and then helping build one of the largest quant communities in finance.
"We paid 5 grand to get data that no one can use in the sport analytics world," he said of the frontier model's effort.
The specialized model his team trains on Bittensor scored roughly 70% on the same benchmark, processed the match in two minutes and cost $10. "There's a big gap between what you would see online and what you would be able to use in production conditions," Sebti said. His company started with a sports hedge fund and soccer clubs and now finds its biggest traction in fuel retail, onboarding names such as Shell and Eni, because the next leap in AI, he argues, is vision rather than text. "Cameras are blind and dumb, and we just make them intelligent," he said.
A 500-to-1 cost advantage on a task a frontier model failed outright is not a rounding error. It is a crack in the premise that bigger and more centralized means better.
The second claim is physical. Even if the giants' models worked everywhere, the buildings required to train them are becoming absurd.
Will Squires, CEO and cofounder of the AI research lab Macrocosmos, has spent two and a half years on the problem of training large models without a giant data center. Frontier models now require facilities that "increasingly look like city blocks," he said, citing research group Epoch AI's projection that frontier compute will demand 4 to 16 gigawatts by 2030. A single gigawatt is roughly the output of a nuclear reactor.
"We believe that's unsustainable," Squires said.
His counterexample is already humming. "In aggregate, Bitcoin is the largest supercomputer," he said, putting its network at 18 gigawatts, many times the largest single AI campus, assembled by no one in particular and owned by no one at all. Macrocosmos built IOTA to apply the same logic to AI training, linking scattered, mismatched GPUs into one working cluster. Squires said its latest results are comparable to centralized alternatives, and a newly released SDK now lets teams move centralized training onto distributed compute.
If distributed training holds up, the industry's scarcest resource stops being the city-block data center and becomes coordination itself. Which raises the third claim.
Toufic Adlouni, managing partner of Renno & Co., the crypto-focused law firm he started in 2017 that has grown to about 40 people including 23 lawyers, advises builders across the Bittensor community. His diagnosis of crypto at large is blunt: much of it is stuck on financial speculation, and the projects that matter are the ones solving "a real coordination problem."
The stakes, as he sees them, extend beyond crypto. Web2 platforms concentrated power and are "attention fracking us," he warns, a trend he expects AI to intensify as value flows to a few companies.
But Adlouni reserves his sharpest warning for his own side. Too many builders, he said, believe their software settles everything: "people think that code is law." It doesn't. Code does not replace "real laws, regulations, courts," and trouble arrives "when things are not papered up," when partners skip written agreements and count on the code to decide the outcome. Decentralization, in other words, does not exempt anyone from the institutions it hopes to improve on.
