I'd like to open today's edition by reaffirming why the world needs decentralized AI now more than ever. I recently came across a piece by 0xJeff that reaffirms this sentiment and echoes many of the thoughts I've been having:
In a world where closed/frontier AI labs like Anthropic and OpenAI lock in users, aggregate resources (GPUs, talents, capital), and consolidate intelligence (closed models),
+ In a world where US government can choose to ban AI models and choose who can or cannot use them,
+ In a world where people are required to KYC (could be soon) to use frontier AI technology,
Open-source & Decentralized AI is key in navigating such world.
Being able to download a model and run it locally OR run it privately via private AI inference, allow users to avoid getting surveillance, avoid their data getting leaked or trained without permission.
The ability to use, develop, train, own the technology lies in the hands of the users. This is the AI sovereignty thesis.
While Open-source AI & DeAI are 2 separate things, the end goal and the thesis is similar.
Open-source models & tools already give users the privacy & sovereignty benefit. What Onchain/DeAI adds on top is
Economic incentives for people to contribute resources/work
Verifiable/scorable outputs
Permissionless participation and coordination at scale
Potential for collective ownership and value accrual
i.e. anyone can join the network and work from anywhere across the world and be a part of owning the network/the output.
Check out 0xJeff's Substack below and subscribe:
AI on Bittensor vs AI on Base
A comparison between the two largest AI ecosystems in Crypto. We'll compare the bootstrapping mechanism, the ecosystem, the use cases, and the asymmetrical bets within the two eco
In a world first, Chutes AI, subnet 64 of the Bittensor / Opentensor Foundation network, has achieved a fully non-blocking decentralized training of a recurrent AI model, with almost no quality loss as compared to traditional centralized training.
Training an AI model across dozens of mismatched GPUs scattered around the globe, without ever hitting pause, sounds like a problem you solve in a decade, not a Tuesday. Chutes AI says it did it on July 8, 2026.
Back in June, the decentralized compute provider outlined the Parallax framework, enabling distributed GPU resources to keep training without synchronization pauses. Now, it landed at 0.6% quality gap - a commercially meaningful margin of centralized training quality.
Chutes AI achieves non-blocking decentralized training for recurrent models with 0.6% quality gap
Chutes AI claims a world first in decentralized AI training, matching centralized quality within 0.6% using its Parallax framework on Bittensor's SN64.
It was no accident that Chutes experimented on a recurrent AI model - an architecture much less friendly to parallelization than the transformer one (for context, most large language models are transformers):
We ran this on a pure recurrent model, Gated DeltaNet, with no transformer or mixture-of-experts layers to soften the result. That was deliberate. Recurrent models are sequential by nature, every step depends on the one before it, which makes them the hardest architecture to train without tight synchronization. Transformers are far easier to parallelize. If non-blocking decentralized training holds on the hardest case, the easier architectures should follow. Gated DeltaNet is our current target architecture for Parallax, and this is an in-progress research direction rather than a finished system, but the signal is clear.
A recurrent model trained decentralized, with no blocking and almos...
We trained a pure recurrent model across distributed GPUs with fully non-blocking sync, and it landed within 0.6% of the centralized baseline. As far as we know, a first.
Why is this so important? This achievement is a definitive proof that decentralized AI training is commercially viable and that building powerful AI does not necessarily depend on constructing larger and larger data centers.
A recurrent model also drops the key-value cache that grows with every token, which is part of why the architecture is attractive to us in the first place. Getting it to train decentralized, without blocking and without a quality penalty, is a real step toward training strong models on hardware people already have.
Chutes is the number one Bittensor subnet by market capitalization and its token was recently listed by Kraken - the first subnet token to get there. If you're as fascinated with it as I am, you'd check out Mark Jeffrey's Hash Rate Podcast episode with Chutes' John Durbin.
The conversation touches upon the current revenues, profitability, and performance of leading AI labs, comparing organizations that operate under centralized and decentralized models.
If you show me a profitable AI company right now, I'll show you a liar.
It also highlights the loss-generating nature of AI business at the moment, and of AI hyperscalers in particular, and the growing discontent of funders and enterprise clients.
[Most AI companies'] capex is 3 to 5 times more than what they're actually bringing in in revenue.
The simplest and most straightforward business logic suggests that the era of massive spending on AI infrastructure cannot continue indefinitely, and that investors are beginning to lose patience.
OpenAI and Anthropic need to multiply what they bring in as revenue just to keep the lights on and satisfy the compute commitments they've already made. Meanwhile, customers are switching to cheaper alternatives, and all sources of institutional funding have been exhausted. That's why both companies are preparing for trillion-dollar IPOs - the public's money is their last resort.
These hyperscalers used to justify their high prices and the staggering amounts of capital they require with the advanced capabilities of their frontier models. However, open-source models that are much cheaper have quickly become just as powerful.
Chutes proves that there's another way to build AI:
You have two options: one, continue to light money on fire, or two, attack it at the root, which is to actually make the models significantly more efficient.
Efficiency has to go up, just has to. We don't have enough chips, nor enough power.
We don't ever want to take any chances with user data. Period. It violates the whole principle of what we're trying to do.
What Chutes, and the entire Bittensor network have accomplished so far is a clear evidence of the numerous benefits of Decentralized AI. And even leaving ideology and abstract ideas like privacy and fairness aside, DeAI delivers on concrete business metrics like capital efficiency, AI verifiability, and reliability.
In contrast, as Durbin states, the status quo right now is such that the price of AI is variable, its availability is variable, and even what's in the model is variable. No enterprise could function at such a level of uncertainty.
AI is critically important and we need to make sure that it is permissionless, borderless, predictable, and available to everyone, everywhere, with no export controls or restrictions. And the way to achieve that is Decentralized AI.
Thank you for reading! My name is Albena, and every day I share insights into the ground-breaking convergence of blockchain and AI. If you’re enjoying them, hit the subscribe button and never miss a key Crypto × AI update.
The Web3 + AI Newsletter is an independent, ad-free publication that I have been building on my own since 2023. If you find value in my work, please consider supporting it by choosing one of the subscription tiers available at the link below. I greatly appreciate it.
The Web3 + AI Book Club is live on Fable! Join us in exploring our July title - 'Empire of AI' by Karen Hao. Follow the link below to read with us.
I'm looking forward to connecting with fellow Crypto x AI enthusiasts, so don't hesitate to reach out on social media.
Disclaimer: None of this should or could be considered financial advice. You should not take my words for granted; rather, do your own research (DYOR) and share your thoughts to encourage a fruitful discussion.