Excerpts From Our Q4'25 Letter
The following is Unsupervised Capital's Q4'25 Letter to its Partners, with sensitive information redacted.
Q4 was difficult for crypto. U.S. equities posted modest gains, gold continued its remarkable run, yet Bitcoin struggled, and with it, the broader crypto market followed it down.
Sentiment within crypto turned decisively negative. A lack of compelling narratives and weak price action, while other asset classes performed well, led many investors to lose conviction. Interest concentrated around the few projects generating meaningful revenue: Hyperliquid, Pump, and select DeFi protocols. Because these became “the only bright spots”, consensus shifted to "blockchains are for finance." In my view, that represents peak despair.
This perspective loses sight of what Bitcoin actually demonstrated, which is that blockchains can transcend finance. Bitcoin proved that open competition combined with aligned incentives is a superior form of organization for large-scale coordination problems. That principle remains as valid as ever. As long as one believes it can be generalized beyond money, there is no reason to despair. Bittensor continues to demonstrate that it can.
Bittensor’s environment of open competition, open source contribution, and incentive alignment, where contributors are rewarded against explicit objective functions, is producing results that AI research institutions like EpochAI, and prominent AI lab co-founders like Jack Clark, are beginning to recognize. We believe we're approaching the inflection point of the S-curve, where attention, builders, and capital deployment begin to accelerate.
It is also worth remembering that Bitcoin’s timing mattered. Bitcoin did not emerge in a vacuum. It gained traction in the aftermath of the global financial crisis, as monetary expansion made the need for a credible counterweight to fiat obvious. We think decentralized AI will follow a similar path. As control over frontier AI systems concentrates among a small number of closed labs, the need for open alternatives also becomes obvious. From that perspective, the timing for decentralized AI feels increasingly favorable given current dynamics.
In December, we published what has become the go-to institutional report on Bittensor outlining our investment thesis. While it has been circulated widely, the most valuable outcome was internal. We spent months returning to first principles and pressure-testing the structural advantages that emerge from combining crypto and AI. Beyond serving as a counterweight to centralized AI, this architecture creates a more competitive environment that meaningfully accelerates AI development.
The intelligence frontier is ultimately a race defined by sustained innovation velocity. Subnets are designed for exactly that. This clarity has mattered in a quarter where confidence was scarce, and it continues to shape how we deploy capital going forward.
Broader AI Themes
Machine intelligence continued its exponential progress through Q4. Dario Amodei, CEO of Anthropic, has described a Moore's Law-like dynamic for cognition, where model capabilities are roughly doubling every four to twelve months. Importantly, this applies not just to closed frontier models, but to open-source models as well.
The most visible manifestation of this progress this quarter was in AI coding tools. Claude Code crossed a threshold where software creation no longer requires a traditional programming background. GPT-5.2 Codex quickly followed, with comparable or superior performance depending on the task. Coding tools are having a ChatGPT-level moment, and the downstream implications extend far beyond developer productivity.
The AI bubble debate has persisted alongside this progress. Skeptics continue to point to AI capex growing faster than near-term revenue. But the variable that actually matters is the rate of improvement. Bubbles form when investment outruns progress, when spending continues despite technological stagnation. Today, we are seeing the opposite. Each quarter delivers meaningfully more capable systems than the last. If that changes, we will reassess. For now, the data points in one direction.
Implications to Bittensor
The implications for Bittensor follow directly. At the highest level, improving sentiment around AI is a clear tailwind for the decentralized AI sector. It pulls incremental capital toward AI assets beyond traditional equities and private markets, while simultaneously increasing attention on the need to decentralize AI development. Bittensor remains the most credible crypto-native AI network and is best positioned to absorb these flows, particularly as access broadens through vehicles like Grayscale’s TAO trust and its proposed ETP conversion. Strength in TAO will likely function as a funnel, pulling attention, liquidity, and experimentation downstream into Subnets.
More concretely, better open-source models directly benefit the Bittensor ecosystem. Miners now have more capable foundations to build on. This helps across the board, whether it's improving deepfake detection on BitMind, 3D asset generation on 404-GEN, or serving inference for increasingly capable open models on Chutes, Targon, and Lium.
But the development we've been thinking about most is what “vibe coding” implies for mining specifically. Historically, mining has been gated by technical complexity. Contributors needed deep domain expertise, familiarity with the protocol, and the ability to manage data pipelines and infrastructure. That set a high bar for participation.
Tools like Claude Code and Codex are changing this. Mining increasingly resembles a prompt engineering game rather than a pure software engineering problem. Technical competence still matters, but the emphasis shifts toward system design, intuition, and iteration speed. This opens the door to a broader class of contributors who can compete on ideas rather than implementation alone.
We think this change to the mining landscape will have a material impact on Bittensor. The network gets stronger as more miners participate and as miner sophistication increases. These coding tools would be doing both at once.
Finally, there is a second-order effect we think is underappreciated. AI systems are accelerating the rate of development across nearly all domains, including AI itself. As the pace of building exceeds human capacity to coordinate manually, automated systems that define objectives and measure outcomes become essential. Subnets do exactly this. They encode objectives, attract distributed contributors, and reward performance. As the world continues to speed up, this architecture becomes more necessary, not less.
Market Thoughts
TAO was among the stronger performers in digital assets this quarter, finishing Q4 down 17% compared to BTC (-23%), ETH (-28%), and NEAR (-43%). In a difficult quarter for crypto broadly, TAO continued to demonstrate relative resilience.
Within the Subnet ecosystem, we've observed early signs of stabilization. The Sum of Subnet Tokens (SST), which measures the collective price of Subnet tokens relative to TAO, reached an all-time low of 0.98 on December 8th. It closed Q4 at 1.09 (+11%) and stood at 1.21 (+23%) as of January 23rd. [Redacted]
[Redacted]
Bullish Themes
Bittensor Protocol Improving
1 - Taoflow
Q4 brought several significant upgrades to the Bittensor protocol. The most important was Taoflow, which changed how the chain distributes TAO emissions to Subnets.
Previously, emissions were largely determined by Subnet token price. That mechanism was gameable and allowed top Subnets to remain comfortable without continuously earning their position. Taoflow replaces price-based signaling with net TAO flows into a Subnet, measured through the Subnet-native liquidity pool. In practice, this forces Subnet owners to generate real demand. Maintaining emissions now requires sustained inflows. Extracting value without corresponding buyers results in declining emissions.
We view this as a clear positive for the network and for both TAO and Subnet token holders. Taoflow realigns incentives around value creation rather than extraction. Subnet owners are now structurally encouraged to reinvest in their ecosystem rather than treat emissions as passive income.
However, the Taoflow rollout was contentious. OpenTensor Foundation implemented the change without a formal feedback period, which created discomfort across network participants. That lesson appears to have been learned, and we do not expect future protocol changes of this magnitude to follow the same process.
2 - TAO Halving
The TAO halving occurred in December 2025. Its effects are unlikely to be immediately visible, but the directional impact is straightforward. Fewer TAO emitted by the protocol reduces structural sell pressure over time. As with prior halvings in other networks, the consequences tend to manifest with a lag rather than instantaneously.
3 - Subnet Owner Exits
On the surface, Subnet owners leaving the ecosystem can appear bearish. We view it as healthy when paired with continued entry from new, energetic teams—which we’re seeing.
To date, we have not seen a Subnet owner with a strong idea and strong execution exit the ecosystem. The primary exception is [Redacted]. That context matters.
A small number of other operators have also exited but this does not concern us. Bittensor is best understood as a liquid venture ecosystem. The expectation should be that the majority of Subnets fail. Taoflow has accelerated this natural selection process. Teams that failed to drive speculative or structural demand for their token have been unable to extract emissions and have chosen to shut down. That outcome is correct.
The feedback cycle in this AI-native environment is extremely fast. Teams learn quickly whether product-market fit exists and whether they are equipped to operate as a public market leader. When the answer is no, exiting early is rational. The game getting harder is not a bug, but a feature.
Good Talent is Arriving
[Redacted]
The public investment is Actual Computer (Subnet 95). We recorded a podcast with the team outlining their thesis, which centers on the inevitability of local inference. This fits squarely within a broader theme we are increasingly bullish on, discussed below. Among Subnet operators to join the network over the past year, we believe Tom running Actual Computer stands out as one of the best.
Another team operating in stealth represents a different signal. Their background and approach would be entirely at home in Silicon Valley. Bringing this caliber of founders into the Bittensor ecosystem is still rare, which is exactly why we view it as encouraging.
[Redacted]
Talent follows opportunity. The fact that stronger teams are beginning to engage with Bittensor suggests that the opportunity set is becoming legible beyond the crypto-native core.
Decentralized Compute
Decentralized compute is a thesis nearly as old as crypto itself. Since at least 2017, projects have attempted to coordinate heterogeneous hardware distributed across the globe. None have succeeded at scale. We believe that is changing.
Historically, the sector has faced three core problems:
No attractive product. Decentralized compute networks had nothing compelling to offer. Quality open-source AI models have changed this. We've written about that here.
No privacy. Networks did not keep user data private. This is no longer the case, with Subnets like Targon and Chutes now supporting TEE (trusted execution environments).
Poor infrastructure and high friction. Early networks couldn't handle basic reliability requirements. If a provider disconnected, users experienced downtime. The products were clunky to use. Both have been addressed. Targon has built infrastructure that can migrate workloads to another server within seconds. Chutes has driven significant usage through its OpenRouter integration, demonstrating that the UX gap is closing.
But the bigger story is two macro trends converging that we believe will bring this sector to life.
The first is the buildout of data centers and the emergence of an inference economy. Just as the energy grid has the duck curve (a graph showing power demand spiking at certain times, requiring capacity to be built for peak load), compute will develop a similar dynamic. Utilization will be uneven. This creates structural demand for liquid compute markets: permissionless platforms where anyone can provide idle compute for others to use and withdraw it whenever they want. A secondary marketplace for compute.
The second is the rise of local inference and AI agents/personal assistants. As local servers and agents proliferate, owners will look for ways to monetize idle capacity. Pointing that compute at a Subnet when there are no local jobs becomes an obvious choice.
One category positioned to absorb this idle compute is decentralized training, where models are trained in a distributed manner and contributors are paid for their resources and intelligence. We're increasingly bullish here. Jack Clark has been vocal about this direction, and research coverage is picking up. Today, the two networks running live, permissionless, incentivized training are both on Bittensor: Templar and IOTA. [Redacted]
Outlook
[Redacted]
Best,
Sami Kassab
Managing Partner
------------------------
This commentary is provided by Unsupervised Capital Management LLC for informational and educational purposes only. It does not constitute an offer to sell, a solicitation of an offer to buy, or a recommendation of any security, investment product, or investment strategy. Nothing herein constitutes investment, legal, tax, or other advice, nor should it be relied upon in making any investment decision.
This material contains forward-looking statements that are based on current expectations and assumptions and are subject to risks and uncertainties. Actual results may differ materially. References to specific digital assets, protocols, or projects are not endorsements and should not be construed as recommendations to buy or sell.
Digital assets involve substantial risk, including the potential for complete loss of principal. Past performance is not indicative of future results.
