Elio

@OnChainFox | TAO & DeAI On-Chain Insights | Incentive Mechanisms · AI Infrastructure · Market Dynamics

[Elio @OnChainFox]

Replit just shifted over $100M a year in AI spend from Anthropic to OpenAI — mainly because of a 30%+ price discount.

Frontier models have moved from competing on capability to competing on price.

But temporary discounts don’t solve the long-term cost problem. Bittensor’s inference subnets are built as competitive markets that continuously drive costs down through real competition, and some already have paying external customers.

This is exactly the problem Bittensor is designed to solve — using competition to keep driving inference costs lower.
$TAO #Bittensor

[Elio @OnChainFox]

Moonshot just released Kimi K3.1 Pro.
A 3T-class MoE model with 1M context. It scores close to Claude Opus 5.5 on Terminal-Bench while being cheaper.

More importantly, they deliberately reduced its hacking capabilities and still plan to open the weights on November 4.

This kind of near-frontier open model is exactly what Bittensor subnets are built for. Once the weights drop, miners worldwide can immediately deploy, fine-tune, accelerate, and compete on agent and coding use cases.

Centralized labs are becoming more cautious and closed. Open weights keep moving forward. bittensor:native is currently the only network that can scale this with real incentives.
#Bittensor #TAO

[Elio @OnChainFox]

While Claude Duet and other proactive agents are rolling out, banks are already racing to block unauthorized third-party agents.

Capabilities are accelerating, but so is the risk that models and agents get adversarially manipulated or bypassed.

@perturbaix (SN26) on Bittensor is doing the opposite of waiting for problems to appear: miners continuously find inputs that fool models, then use those attacks to harden the models themselves. Turning attack and defense into an ongoing competition.

[Elio @OnChainFox]

Microsoft’s AI diagnostic system (MAI-DxO) clearly outperformed general physicians on complex cases — that’s worth paying attention to.

It doesn’t just spit out an answer. It works step by step, like a panel of doctors: asking questions, deciding which tests to order, and updating its judgment based on the results. But these results are still from controlled text-based cases, not real patients. Real deployment has to deal with data sources, verification standards, accountability, and privacy — not just another centralized lab announcing it’s better.

If Bittensor can produce verifiable, competitive results in medical-related subnets instead of relying on a single company’s narrative, that would be more interesting. A lot of the big lab “breakthroughs” in medicine right now are still basically “we tested it internally.” Open competition at least offers another way to check the claims. $TAO

[Mustafa Suleyman @mustafasuleyman]

Our MAI Diagnostic system is now 1.7x more accurate than physicians in the ED, even when they have access to the web to look things up.

It's also lower cost & does fewer medical tests to get to its diagnosis.

Published in the journal Science today...

[Elio @OnChainFox]

Most subnets talk about building useful AI.

@perturbaix is actually doing it — running a live network where miners get paid to attack models and then make them harder to break.

That’s the kind of real work the TAO ecosystem needs more of.

Not more narratives. More pressure testing.

[Elio @OnChainFox]

The real next stage of AI isn’t just building a better chatbot. It’s whether AI can actually connect to the real world.

For the past few years, most intelligence has stayed inside centralized labs and closed systems. No matter how strong the models get, they still face the same problem: continuous feedback from the real world is hard for a few companies to monopolize long-term.

Once AI needs to perceive, act, fail, and learn again in real environments, centralized approaches hit clear limits — slow data collection, narrow feedback loops, and rigid incentives.

This is where decentralization becomes more concrete. It’s not about building the bodies ourselves. It’s about enabling real-world data, task feedback, and model iteration to be organized more openly across a network. Whoever can more efficiently connect real-world actions to model improvement will have a chance to claim a position in the next stage.

Chatting is just the starting point. Being able to act is the next stage. The demands this places on decentralized networks will be much higher than they are today. bittensor:native

#TAO #Bittensor #DecentralizedAI

[Elio @OnChainFox]

Sounds very promising. Looking forward to seeing the results once miners get involved.
I strongly believe AI needs to start making real hardware impact — designing better motors, cooling systems, and physical products faster and more reliably. Recent moves like manufacturing capital (CATL, Geely, etc.) backing DeepSeek and DeepSeek investing in robotics (Unitree) show this is becoming a clear trend. Carbon on Bittensor feels like a solid way to close that gap. Excited to follow the progress!

[Elio @OnChainFox]

Will TAO’s development consider landing in manufacturing?

Right now the focus is still on general inference and compute, and manufacturing isn’t the main narrative. But factories are where AI is most likely to generate real paying demand — quality inspection, production scheduling, predictive maintenance. These use cases can actually be measured in dollars.

@const_reborn how do you see this path? Bittensor won’t centrally decide the direction, but specialized industrial subnets can absolutely emerge. The hard parts are data privacy and edge deployment. Whoever first delivers real paying factory use cases will prove the model works.

Manufacturing cares about results on the floor, not concepts. bittensor:native

[Elio @OnChainFox]

Looking forward to Carbon’s direction — Physics AI has real potential in hardware design, which is a massive market. That said, this space cares a lot about real-world results, so I’d love to see some grounded citations or case studies. For example, measurable improvements in motors, cooling systems, or structural optimization compared to traditional simulation, and whether any engineering teams are already using it in practice.

[Elio @OnChainFox]

@carbonphysicsai Awesome direction. How far have the models progressed so far? Any early results or demos on specific hardware like motors or thermal systems?

[Elio @OnChainFox]

@markjeffrey hope to bring something different to TAO.

[Elio @OnChainFox]

@const_reborn @reaxlabs Kind of interesting.

[Elio @OnChainFox]

@markjeffrey Everything takes time; I’m really looking forward to seeing what kind of different things it might bring.

[Elio @OnChainFox]

My current view on bittensor:native is that it is going through a fairly typical mechanism transition. There is real progress, but also unresolved issues from earlier stages, which is why the overall picture feels messy right now.
External customer revenue has crossed $50M annualized and is growing at a decent pace. This is one of the more concrete positive signals we have seen, because it shows that some subnets are building things paying users actually want, rather than just circulating emissions internally.

At the same time, a number of subnets extracted meaningful value through emissions over the past few years without leaving behind a comparable amount of product or recurring revenue. That mismatch does not disappear simply because the revenue number improved, and it continues to affect how the market evaluates many subnets.

eTAO makes emission splits more flexible, which is directionally reasonable since different subnets have different business models. In the short term, however, that flexibility can also be used for surface-level adjustments — whether that means lowering the owner cut to signal support for builders or raising it to capture more rewards — without a corresponding increase in product progress or revenue.

My judgment is that the biggest uncertainty for TAO is not whether revenue exists, but how many subnets can turn that revenue into consistent product delivery and buybacks instead of continuing to rely primarily on emissions and narrative.

If the number of subnets with verifiable external revenue and ongoing product progress increases meaningfully over the next few months, concerns about extraction should gradually ease. If the revenue figure keeps rising while most subnets show little change in actual delivery or capital usage, the sense that problems are accumulating will likely persist.

What matters more right now is not how individual subnets adjust their splits, but how many of them can provide clear evidence on both revenue and product development at the same time.

[Elio @OnChainFox]

@opentensor Looking forward to the discussion.

[Elio @OnChainFox]

@chutes_ai training an 8B for ~6500 and running it on a phone is wild.

[Elio @OnChainFox]

@TAOFlows TAO’s participation is actually pretty solid.

[Elio @OnChainFox]

Which of TAO, RNDR, FET, and AKT is more likely to actually capture AI value?
A simple comparison across three dimensions:

Real customer revenue:
$TAO currently leads with the strongest reported external subnet revenue (annualized above $50M and still growing).
$RNDR has ongoing rendering/compute jobs that directly burn tokens, though at a smaller absolute scale.
$AKT shows real cloud leasing usage, but the revenue base remains modest.
$FET has the weakest publicly verifiable customer revenue data so far.

Adoption & product maturity:
RNDR and AKT are further along on the compute/cloud side.
TAO is progressing faster on the intelligence marketplace side.
FET remains more ecosystem- and narrative-driven.
Token-value linkage:
RNDR’s burn mechanism is the most direct.
TAO links value indirectly through reward distribution plus some subnet buybacks.
AKT has a BME mechanism, but limited revenue scale caps its impact.
FET’s linkage is relatively weaker.
Right now, $TAO leads on real AI-related revenue, while $RNDR is cleaner on “usage = burn.”

All four still have relatively small revenue bases and remain dependent on emissions — sustained real revenue will determine who actually captures value.
#TAO #RNDR #FET #AKT

[Elio @OnChainFox]

@DreadBong0 If TAO actually pulls off becoming the base layer for decentralized AI, the whole picture changes. Still early days though, gotta see if the subnets can actually deliver.

[Elio @OnChainFox]

Openτensor has launched eTAO, giving subnet owners the ability to decide how alpha emissions are distributed, instead of being locked into the fixed 18% (owner) / 41% (validators) / 41% (miners) split.

According to the official announcement, owners can now adjust the allocation within certain limits: owner up to 33% (at least 18% for leased subnets), validators up to 50%, and miners at least 33%, with changes allowed roughly once every 7 days. The system also automatically adjusts the subnet’s effective TAO weight to protect root bittensor:native stakers’ returns.

In my view, this effectively decentralizes part of the incentive design to subnet owners, allowing different projects to tailor their reward structures based on their specific needs rather than forcing everyone into the same template. For subnets aiming for differentiated competition, this is a meaningful step forward.

That said, greater flexibility also means subnet quality will depend more on the owner’s ability to design incentives well. Done right, it could create a positive flywheel; done poorly, it might hurt long-term participation. It’ll be interesting to see how subnets actually adjust these ratios once it goes live.

[Openτensor Foundaτion @opentensor]

Introducing eTAO: a new system giving subnets control over their alpha emission split.

Subnet owners can now decide how much goes to miners, validators, and themselves, based on the specific needs of their subnet.