DeAI compute jumps from <1B to 100B parameters as regulators echo Bitcoin's 2014 playbook
Export bans, identity-gated access and a 100x rise in distributed training capacity frame open-source AI as crypto's 2014 rerun.
Distributed AI training capacity has scaled from under 1 billion parameters to 100 billion over the past two years, according to a note from Brownstone Research’s Chain of Thought newsletter authored by analyst Ben Lilly. The growth curve, he argues, mirrors the early infrastructure buildout that preceded Bitcoin’s regulatory clearing in the 2020s — and the policy response to open AI is following the same script regulators once used on crypto.
Access restrictions tightening in real time
Two access-control moves anchor the current cycle: a U.S. export ban on Anthropic’s latest model release, which Lilly expects to push the firm toward identity-verified, permissioned distribution, and OpenAI’s decision to cap its GPT-5.6 rollout to trusted partners only. Lilly frames the direction of travel bluntly: “It’s for your protection, you see. It always is.”
The security rationale traces back to Anthropic CEO Dario Amodei’s July 2023 congressional testimony, where he called open source “a good thing” across most scientific fields and described the risk from open models released to that point as “relatively limited” — while warning that continued scaling of open-source AI was heading “down a very dangerous path.” Lilly reads this as the opening move in a familiar sequence: label the open alternative as dangerous, then position the closed, commercially controlled product as the default safe option.
The threat case gets sharper detail from NSA chief Joshua Rudd, relayed by Sen. Mark Warner, who described Anthropic’s “Mythos” model breaching “almost all of our classified system, not in weeks, but in hours.”
The 2014 parallel
Lilly’s comparison point is Bitcoin’s own early hostility from Washington: Rep. Jared Polis buying the first Bitcoin on Capitol Hill in 2014, Sen. Joe Manchin calling for a ban on a “dangerous currency,” and the 2023 allegations known as “Operation Choke Point 2.0,” in which regulators were accused of pressuring banks to sever ties with crypto firms. The industry survived that pressure, and policy has since moved toward defined rules — the GENIUS Act has passed and the CLARITY Act remains pending, both cited as evidence that early crackdowns eventually yield to regulatory frameworks rather than eliminating the asset class.
Where open models sit on the performance curve
On raw capability, the gap is narrowing rather than widening. GLM-5.2 scored on par with Anthropic’s Sonnet 4.6, released in February, putting the leading open model roughly three to four months behind the closed frontier — and Lilly expects an open competitor to Mythos and GPT-5.6 to surface by fall.
The structural shift he highlights is decentralized training on peer-to-peer networks resembling Bitcoin and Ethereum, where contributors supply compute in exchange for a stake in the resulting model rather than a security reward. Three early-stage projects named in the note: Dark Bloom, offering low-cost private inference on idle Mac hardware; c0mpute, a decentralized inference network; and Pluralis, which trains models across distributed consumer GPUs. Lilly expects additional token-incentivized compute networks to launch as the model builds out.
His conclusion: attempts to restrict open AI models will ultimately fail, and positioning in decentralized AI infrastructure now “will be like buying Bitcoin in 2014, back when it was still ‘dangerous.'”
Read more: Jeff Booth: Bitcoin’s Future Depends on Seeing It as a Protocol, Not an Asset
Leave a Reply