In March, Google published research bringing quantum computers meaningfully closer to breaking ECDSA, the cryptography securing Bitcoin among many other things. Citing conversations with the government, Google withheld the key innovation, releasing only a zero-knowledge proof that the result was real, not the method behind it. Earlier this month, a distributed crowd of researchers and AI agents, organized around ecdsa.fail, a project from our portfolio company Eigen Labs, rediscovered it in the open. Within days they had pushed past Google’s result. The open version now runs roughly 40% ahead of Google’s benchmark.
Agent swarms are one example of decentralized networks starting to operate at and even beyond the level of the largest centralized AI companies. The more valuable one is playing out in the economics of AI itself. Inference, the process of running a trained model to produce an answer, is where practically all AI revenue sits. Today that is roughly $60-65bln annualized, the vast majority accruing to Anthropic and OpenAI; some analysts expect Anthropic alone to reach $100bln of ARR by year-end. Five years out, projections run to $500bln-$1trn. It is for this reason that the giants have poured some $600bln into AI infrastructure.
What hardly anyone is talking about is that decentralized networks training models on consumer hardware are starting to compete with this state-of-the-art infrastructure, something AI experts dismissed as impossible only a few years ago. Our manager Jake Brukhman, founder and CEO of CoinFund and one of the earliest investors at this intersection, set it out in a recent interview on The Rollup. The latest milestone he points to: last week, Macrocosmos trained a 100 billion parameter model on the Bittensor network. That is commercially viable scale. Frontier models are larger still, but the gap is narrowing fast.
Making the training work, however, is only half the story; the other half is monetization. A trained model is ultimately just a large file of numbers, its so-called weights. In decentralized training runs so far, every participant ended up with a full copy of those weights, leaving the model free for anyone to use and therefore hard to charge for. CoinFund portfolio company Pluralis solves this by splitting the weights across the GPU providers that train the model, so no single party ever holds the full model. The model lives in the network, and using it means paying the network for inference. As Brukhman puts it: “if we can win like 10% of training, we also win the inference”. That is a $50-100bln market within five years, accruing to the participants who co-own these models rather than to Sam Altman or Dario Amodei.
This may be one of the biggest investment opportunities that almost nobody is positioned for.
Want to go deeper? Join our live webinar with Jake Brukhman on June 25th:
This week’s market
- BGCI Index 0.95%
- Bitcoin 1.77%
- Ethereum -0.08%
- Solana 0.56%
All figures are week-to-date as of today 13:00 CET
Theta Blockchain Ventures content
- Webinar with Vance Spencer and Michael Anderson of Framework Ventures (request here)
- Register for our webinar with Jake Brukhman (CoinFund), June 25 (register here)
- Watch Jake Brukhman at Legends4Legends 2025 (watch here)
Blockchain news
- Citi opens new route into private markets with tokenized share offering (learn more)
- Japan‘s three megabanks to debut live stablecoin transactions by March 2027 (learn more)
- New York regulator proposes stablecoin rule to align with federal GENIUS Act, adds reserve limits (learn more)
- Figure* to acquire Kiavi for $717 million to expand RWA tokenization network (learn more)
- CFTC unveils sweeping rule proposal for fast-growing prediction markets (learn more)
- America’s largest banks are building a new digital currency network to stop a massive deposit drain (learn more)
- Mastercard unveils Agent Pay for Machines to support autonomous AI transactions, including stablecoins (learn more)
Relevant financial updates
- Kiavi, an AI-powered lending platform for real estate investors, was acquired by blockchain-powered lender Figure* for $717 million
- Neura Robotics, a humanoid robot manufacturer, raised $1.4 billion in a round led by Tether, who plans to integrate its wallet software into Neura’s robotic systems
- Digital Asset*, the company behind Canton Network, has raised $355mln in a funding round led by a16z crypto with investments from Coinbase Ventures, Polychain Capital, and 22 others at a $2bln valuation
- Morpho*, a decentralized lending protocol, has raised $175mln in a funding round led by a16z crypto, Paradigm, Ribbit Capital with investments from Hashkey Capital, Mirana Ventures, and 12 others
- EDGE Markets*, a banking and centralized payments provider, has raised $29.2mln in a Series A round led by CoinFund with investments from Mantis VC, Bullpen Capital, StepStone Group, Indicator Ventures
- Vinyl Equity, a tokenization-focused transfer agent, raised $20 million from Jump Capital, Mitsubishi, Index Ventures, and others
- MNX*, a decentralized exchange, has raised $6.4mln in a Pre Seed round with investments from North Island Ventures, Scott Moore, Finality Capital Partners, Village Global, and 19 others
- TVL Capital*, a derivatives provider, has raised $5mln in a Seed round led by Framework Ventures with investments from Flow Traders
- Ethena*, an algorithmic stablecoin and decentralized issuer, has received a strategic investment from Janus Henderson
*Underlying TBV portfolio position | Prices as per 12/06
Interesting things to read and listen to
- Jake Brukhman’s op-ed on how decentralized AI training creates a new asset class (read here)
- Jake Brukhman on that decentralized networks training models on consumer hardware are starting to compete with state-of-the-art infrastructure on The Rollup (listen here)
- Restoring access to pre-IPO companies by our manager Pantera (read here)