Anthropic’s most powerful public model came with guardrails users called too broad — then the US government switched it off entirely. Crypto’s builders say that is exactly why AI should run on networks no company or state controls, and traders have started to bid up the tokens betting on it.
Crypto investors and builders say the censorship of Anthropic’s Fable 5 proves their long-running argument: that AI should run on decentralized networks no company or government can switch off.
The model shipped with guardrails so broad that many users complained, by Anthropic’s own account, and would quietly degrade its answers when asked to help train other AI. Then on June 12 the US government forced Anthropic to disable Fable 5 and its more powerful sibling Mythos 5 for every user, after an export control directive barred access by any foreign national, the company said in a statement. The order covered foreign nationals inside and outside the US, including Anthropic’s own employees, and left other models such as Claude Opus 4.8 untouched.
For a corner of crypto that has spent years building AI on blockchains, the takedown was a live demonstration of its pitch.
“Their access to AI is at will,” said Jake Brukhman, founder and chief executive of CoinFund, one of the earliest US crypto investment firms, on a Defiant panel. “It’s at the pleasure of big private companies like Anthropic and OpenAI … and the whim of the government.” When the order hit one company, he noted, the effect was a near-global shutdown.
The episode sharpens a question that has trailed the AI boom: whether the most capable models should sit with a handful of US firms and the government that can rein them in. Crypto’s answer is that decentralized networks can keep AI running, private and beyond any single authority. The harder question, which even its own investors press, is whether those systems can match the centralized labs at all.
Decentralized AI Tokens Outrun Bitcoin Since the Shutdown
Traders moved first. The broad crypto AI sector is worth about $22 billion, according to CoinGecko, a fraction of the roughly $965 billion that recent funding rounds have ascribed to Anthropic alone. Within that sector, the tokens most closely tied to decentralized AI jumped after the June 12 order, even as bitcoin slipped 0.8% over the past week.
Bittensor’s TAO, the largest pure-play decentralized AI token, led the move. It rose as much as 30% within 12 hours of the shutdown and roughly 39% over the following days, touching a three-week high near $283. It has since pared much of that to about $236, up 11% over the week, CoinGecko data show.
Smaller compute and agent tokens ran further: Akash’s AKT added 18% and decentralized-agent project Morpheus’s MOR gained 16% over the same stretch, against gains of 1.7% in ether and 4.3% in SOL.
The rally is partly ideological. Grayscale, the asset manager, told clients the shutdown demonstrated “centralized control over advanced AI technologies” and pointed to Bittensor as an open, permissionless alternative. Erik Voorhees, founder of the privacy-focused AI app Venice, tied the order to his product directly, posting “There’s a reason we built Venice.”
The sector remains small and concentrated. The largest token in CoinGecko’s AI category, Chainlink, is an oracle network rather than a decentralized AI project; the recognizable pure-plays are NEAR and Bittensor, followed by compute and agent networks an order of magnitude smaller.
| Project | Token | Market cap | Since shutdown (7d) |
|---|---|---|---|
| NEAR Protocol | NEAR | ~$2.9B | +8% |
| Bittensor | TAO | ~$2.3B | +11% |
| Render | RENDER | ~$0.9B | +4% |
| Venice | VVV | ~$0.7B | +2% |
| Artificial Superintelligence Alliance | FET | ~$0.4B | +1% |
| Akash Network | AKT | ~$0.2B | +18% |
| Morpheus | MOR | ~$0.02B | +16% |
Source: CoinGecko, as of June 18. Figures move quickly; refresh at filing.
Beyond the liquid tokens sit the projects trying to do the hardest part — training models across distributed machines — most of them still private. Brukhman pointed to Pluralis, Gensyn, Bagel and Prime Intellect among the startups CoinFund and others are backing, alongside research efforts such as Nous and Macrocosmos.
Why Crypto Calls It Censorship
The case rests on more than one government order. Anthropic had kept Mythos out of general release over concerns it could boost hacking, limiting it to a select group of companies under a cybersecurity program called Project Glasswing. It released Fable 5 publicly with safeguards meant to block the underlying model’s most sensitive cyber uses — safeguards “so strong that many users have complained that they are overly broad,” the company acknowledged. Fable also degrades its own performance on tasks that look like training a rival model.
Illia Polosukhin, co-founder of NEAR Protocol and a co-author of “Attention Is All You Need,” the 2017 paper that introduced the architecture behind modern large language models, said the layered restrictions amount to controlling what people can ask rather than what they do. Blocking a model is “trying to police the thought,” he said in an interview on The Defiant’s podcast.
The government order, in his view, set a wider precedent.
It is “a very bad precedent for internet overall, because effectively US government shown they can ban any US-made internet product on the whim,” Polosukhin said. Brukhman made the structural version of the same point: “By default, when AI is being developed by centralized companies, then they’re developed in this … corporate form that is highly susceptible to government pressure.”
The Case for ‘Freedom Technology’
The pitch is that decentralization removes the off switch. Polosukhin wants inference — the step where a trained model answers a query — to run across a permissionless network of machines, encrypted end to end. NEAR uses secure enclaves, hardware he says can verify code ran correctly without exposing the data, at a few percent overhead rather than the heavy cost of replicating work across a blockchain. Run that way, he argued, AI cannot easily be censored or surveilled. “I think of privacy as effectively the freedom technology,” he said.
Brukhman framed decentralized AI as a deliberate counterweight.
“One of the reasons that I am in decentralized AI is to create a counterbalance … so that we can still have access to AI,” he said. He argued the demand is already there, pointing to open-weight models from Chinese labs that sit within a few percentage points of the frontier on public benchmarks. “There is a ton of demand for global permissionless AI, whether the State Department likes that or not,” he said.
The thesis is starting to show revenue at NEAR. NEAR Intents, the protocol’s cross-chain settlement layer, held about $92 million in assets and generated roughly $2.4 million in fees over the prior 30 days, DefiLlama data show. Polosukhin tied NEAR’s run to structural changes — full dilution five years after launch, an inflation cut in an October governance vote, and a fee switch turned on for NEAR Intents in February — rather than to the AI debate itself.
Can It Actually Compete?
The investors who build at the AI-crypto seam are also its sharpest skeptics, and their doubt is not about the principle but the product.
“Most decentralized AI hasn’t really been of the quality of the centralized alternative,” said Jesus Rodriguez, founder of crypto data firm Sentora and author of the AI newsletter The Sequence, who has built and sold AI companies. “It’s not a question of the value proposition conceptually, it’s a question of the product.”
As the leading labs push pre-training, post-training and newer techniques, “the gap with decentralized models has widened and not shortened,” he said, calling decentralized AI an idea that “has been around forever and it has never found product market fit.”
Haseeb Qureshi, managing partner at crypto venture firm Dragonfly, said the economics do not yet support training or running frontier models on distributed networks. Coordinating scattered machines over the public internet carries a real bandwidth cost, he argued, and the labs control vast, expensive datasets that a “decentralized trove of people” cannot match.
The scale gap shows in the numbers: the entire crypto AI sector is worth roughly 2% of Anthropic’s private valuation. Qureshi’s own view is that crypto’s contribution will be narrower — letting users run existing open-weight models privately and cheaply, as Venice does with models such as DeepSeek, rather than training new ones in a decentralized setting.
“The core value proposition of crypto is not decentralization,” he said. “The end is self-sovereignty and censorship resistance.”
He also warned that the freedom crypto is championing cuts both ways. Cheap, ungated frontier AI would put offensive cyber capability in every hand, including hostile states.
April 2026 set a record for the number of crypto hacking incidents, with about 30 exploits, the most of any month, according to DefiLlama; losses topped $600 million before falling to roughly $68 million in May even as incident counts stayed near the high, security firm CertiK found. Analysts have tied much of this year’s stolen funds to North Korea. “This thing is basically a bazooka,” Qureshi said. “I don’t want North Korea to have Mythos.”
What’s Next
Anthropic said it disagrees with the order, believes it is a misunderstanding, and is negotiating with the administration to restore access. Polosukhin said NEAR will keep adding products to its consumer app, including perpetual-futures trading and a relaunched marketplace for AI agents. The decentralized AI camp is betting the next generation of training methods narrows the gap its own skeptics say keeps widening — and that the tokens that ran on the Fable shutdown hold the bid.
