AI × Crypto Investment Thesis 2025

Before We Start

This thesis aims to be more useful rather than complete.

We aim to mention the largest opportunities in the respective fields at the end of each section.

We choose to use “onchain”, instead of “on-chain” or “on chain” for no particular reason.

The flow of the document is the following:

• First, we will explore the landscape, the “data wall”, and our ultimate vision – commoditized cognition.

• Second, we will detail the core building blocks of commoditized cognition: foundational models, specialized models, reasoning models, and memory.

• Third, we will examine onchain models and how they unify these building blocks to create "mechanism-aware AI".

• Fourth, we will showcase the novel “AI-aware market mechanisms” that will power the cybereconomy and make it work at machine-speed.

The Landscape

AI represents the world's largest market because it encompasses all of problem-solving. The resulting AI market landscape is enormous. Recognizing that it's impossible to invest everywhere, we specifically target the intersection of Crypto and AI.

We believe in the collective effects of commoditized cognition, programmable markets and regulation & governance through technology.

Traditional economies are bounded by human productivity and coordination overheads. Complex systems favor monopolistic control – hierarchies arise to minimize transaction costs and reduce complexity, but they inevitably concentrate power and wealth, creating dominant entities – often called enterprises, firms (or  governments) – that dictate the economic landscape.

While crypto helped reduce transactional and coordination overheads, the exponential growth of artificial intelligence will bring them down even further. As these inefficiencies vanish, centralized firms will transform to more efficient network-structured self-regulated entities, and we will enter a new era – the era of the Cybernetic Economy.

The Sutskever’s Wall and its Repercussions

Today, LLM training consists of pre-training – training next symbol prediction on enormous amounts of text; instruction fine-tuning – training on Q&A data to follow instructions; and alignment – reinforcement learning from human feedback to improve helpfulness and safety.

It has become evident that model pre-training has hit a critical plateau – The Data Wall – where additional compute power no longer yields significant performance improvements.

As a result, in 2025 the entire AI industry will focus on these key activities:

1. Hyperscalers (OpenAI, Anthropic, Google, etc.) will leverage their data advantages to develop superior reasoning models, while continuing to monetize inference results. Context: reasoning models require large amounts of "chains of thoughts"-like reasoning data to train. Such data is extremely scarce, expensive, and difficult to collect at scale, limiting reasoning model feasibility to hyperscalers.

2. Open-source companies will continue publishing GPT-4o level open-weights models (unless they find ways to monetize the models themselves, rather than inference results). Context: for a long time, GPT-4o stood as an unreachable benchmark without any comparable open-source alternative. This changed when DeepSeek released its V3 model as an open-weights model.

3. Enterprises will focus on distilling large reasoning models into smaller domain-specific models, while advancing data retrieval technologies for RAG. Context: knowledge distillation and RAG are the main fine-tuning techniques for enterprise use-cases as they give enough control and flexibility to implement automation workflows.

4. Individuals and researchers will develop AI agents by wrapping existing APIs and platforms. Context: while being GPU-poor individual contributors and researchers are domain-experts with years of experience and they will deliver the most important component of the entire ecosystem – applications – an opinionated way to solve real life problems.

The Endgame: Commoditized Cognition

Bitcoin used electrical power to create the first ever permissionless consensus. Classified as a legal commodity, the respective asset BTC is outside of SEC jurisdiction, with no centralized authority, no corporate governance, and no dependency on traditional financing.

AI models convert electrical power into monetary value through training and inference processes. They could have associated assets that could be classified as commodities if these processes are decentralized and the resulting models remain unextractable and perpetually tied to the blockchain.

Ethereum is a global decentralized computer, and the associated asset is also classified as a legal commodity, demonstrating that computational mechanisms can be commoditized and transformed into a universally accessible natural resource, rather than restricted as a proprietary service.

We invest in technologies and market mechanisms that accelerate commoditization of AI.

We believe that an onchain AI model is the underlying fundamental technological primitive that will enable commoditization.

The resulting commoditized cognition will unlock an opportunity for institutional capital to invest directly into AI models, the collective intelligence.

Model Pre-training

Commoditized cognition needs foundational models. It has become evident, as highlighted by Ilya Sutskever at NeurIPS 2024, that pre-training has reached a critical plateau -- the wall -- where further increases in compute power no longer yield significant performance improvements. This development sets a clear benchmark for the "final cost" (CAPEX) of building a GPT-4-like model "from scratch".

Historically, the crypto industry has been very effective in fundraising. With the right team, an onchain model, and distributed inference -- even if training remains centralized -- the emergence of a new "OpenAI" remains possible.

The DeepSeek-V3 model was trained at a cost of approximately $5.5 to $5.6 million, utilizing 2.78 million GPU hours on 2048 H800 GPUs over a span of about two months. This cost is significantly lower than that of comparable models, such as Meta's LLaMA 3, which required 30.8 million GPU hours and a much higher budget.

It is absolutely possible to train a foundational model for $5-10M from scratch already today.

Fine-tuning: Specialized Models

Fine-tunes and derivatives are essential for businesses. Majority of fine-tuning is happening to the small language models. The year 2024 was a year of small language models as they increasingly closed the gap with larger models when fine-tuned for specific tasks and compared under compute constraints –  showcasing how focused optimization can outperform general-purpose systems.

We introduce Sky-T1-32B-Preview, our reasoning model that performs on par with o1-preview on popular reasoning and coding benchmarks. Remarkably, Sky-T1-32B-Preview was trained for less than $450, demonstrating that it is possible to replicate high-level reasoning capabilities affordably and efficiently (https://novasky-ai.github.io/posts/sky-t1/).

Fine-tuning can be done at different levels:

• Foundation level: Instruct fine-tuning and alignment.

• Domain level: Specialist models for fields like medicine, law, and biotech.

• User level: Adapts to individual user needs and style.

• Prompt-level: Adapts to an individual prompt in real-time.

User-level and prompt-level fine-tuning are the killer apps: A 3.8B parameter model fine-tuned at inference time outperformed a 130B parameter general model (https://arxiv.org/pdf/2410.08020).

The world will need tens of foundational models, hundreds of domain-level models, millions of user-level models, and trillions of prompt-level models.

The biggest untapped opportunity in this space is prompt-level fine-tuning.

Reasoning: Inference-time Compute

"The Data Wall" is the main reason for OpenAI to release reasoning models like o1 and o3. The difference from previous generations is that these models target inference-time compute, meaning they implement "chain-of-thought" in one form or another to climb the scaling law and achieve better performance.

Inference-time compute is a game-changer for the entire AI industry as it introduces an unknown variable to the cost structure – OPEX – variable inference compute. Previously, we had a pretty good idea of how much it costs to do the inference; now, it depends on how long the resulting chain-of-thought is. This can be viewed as an advantage for large enterprises and also as a "risk component" in the agentic market setting (when an agent takes a task without knowing how much compute it will take to finish).

It is unknown how complex or simple the "chain-of-thoughts" are, but it is expected that the open-source community may start closing this gap in the upcoming year.

Novel "chain-of-thoughts" architectures and distillation of reasoning models are the biggest opportunities in this sector of AI.

Memory: Gems for the GPU-poor

Context memory is the only research direction that can enable "personalized information gains" -- a surprising piece of information that updates an internal world model of the user the most.

Memory is both underdeveloped (plaintext) and researchable by GPU-poor startup teams and individual researchers. Effective context memory allows LLMs to retain relevant information across interactions, enhancing the continuity and relevance of responses.

A recent Google paper titled "Titans: Learning to Memorize at Test Time" introduces a relatively small AI model that acts as memory for the larger model. We believe this trend will continue as we see more "differentiable memory" designs emerge.

Memories that employ "theory of mind" approaches enhance models' ability to understand and predict human intentions and mental states, improving the quality of interactions.

Memories can be made portable and are easy to own onchain. This opens a market for “memory marketplaces”, where multiple AI agents can buy or license shared historical contexts.

Memory is also important to enable efficient multi-agent systems, mixing agent memories together to achieve better performance gains.

Memories can also be used to produce a database of personas or characters, which is important for AI agent diversity. These “virtual personalities” become portable assets, creating new opportunities for personalization and creativity.

The greatest opportunities in memory revolve around maximizing personal insight and digitally capturing "interestingness". Imagine a social network where the recommendation engine prioritizes content solely based on how much genuinely intriguing information users can discover.

Onchain Models: mechanism-aware AI

The core building blocks of AI – foundational models, fine-tunes, reasoning, and memory – are not mechanism-aware by default, meaning they cannot easily integrate with market mechanisms.

The cornerstone of the entire AI x Crypto commoditized cognition concept is the onchain model.

We loosely define an onchain model as an AI model that executes operations only when provided with an approval recorded on the blockchain.

The following unique properties of onchain models enable new market mechanisms through the fusion of economic and technological capabilities:

• Onchain model is both permissionless for inference and fine-tuning (like open-source models), and monetizable (like proprietary ones).

• Onchain model guarantees perpetual accessibility – continuous access even if the original creator withdraws the model from circulation.

• Onchain model allows for decentralized governance, through a smart contract ensuring that usage, fine-tunes, and financial flows are transparently managed without centralized intermediaries.

Given blockchain resource constraints, onchain models may appear challenging to develop and deploy – and they are. They can practically be implemented using Trusted Execution Environments (TEEs), such as Nvidia's Confidential Computing solutions, ensuring secure and verifiable computation aligned with blockchain-based governance. The hypothetical architecture of the onchain model architecture is depicted below.

The primary goal and biggest opportunity in this space is to standardize and support the entire lifecycle of the onchain model inside the TEE: training, inference, fine-tuning, evaluations, governance.

Standardizing the lifecycle ensures consistency, interoperability, and security across implementations. It simplifies collaboration, reduces integration complexity, and accelerates adoption by providing clear guidelines and benchmarks.

Standardization lowers entry barriers for developers, significantly broadening participation and driving innovation in this emerging space.

AI-aware Market Mechanisms

Efficient integration of the onchain models is only possible with the right incentive-alignment in place, which in turn requires market mechanisms. These market mechanisms have to be “AI-aware” – means they need to recognize the unique characteristics and requirements of onchain models, including tokenization, dynamic inference costs, governance complexities, and operational risks.

Unique properties of onchain models allow the following “AI-aware market mechanisms” to emerge (and hopefully attract the next $100B into the crypto ecosystem from institutional investors):

• Distribution of profits from AI models to token holders creates significant value.

• Establish revenue-sharing mechanisms where token holders earn from model derivatives (including fine-tuned versions), ensuring sustainable rewards regardless of the original model's popularity.

• Integration of AI with lending markets: on-chain models serve as collateral, enabling financing for model training and fine-tuning.

• High token velocity enables institutional investors to participate in liquidity pools, facilitating large-scale trades with minimal price impact.

• Unpredictable inference costs create opportunities for institutional investors to introduce futures, options, derivatives, and perpetual contracts.

• Analyze correlations between AI model performances to develop hedging strategies, using stronger models to offset risks from underperforming ones.

• Offer insurance against inference-time compute bloating, model performance degradation, and other operational risks.

• Create a unification layer where multiple AI model tokens can be aggregated into a single, standardized token, improving stability and reducing velocity.

• Deploy mechanisms to encourage long-term holding and reduce token velocity through staking rewards, vesting periods, and loyalty programs.

• Invest in Layer 2 solutions and decentralized storage to support high transaction volumes and large-scale AI models efficiently.

What to focus on? The 3 best value-capture mechanisms will be:

• Distribution of profits created by AI models to the token holders (coordination).

• Distribution of profits down the AI model provenance tree (composability, derivatives).

• Discovery, interactions, and transactions between AI agents. You can find more information on the AI agents here: https://cyber.fund/content/web3-agents 

In terms of products, the biggest untapped AI x Crypto opportunity is the onchain "Hugging Face" with the ability to own models, model derivatives, and inference profits. Permissionless inference, permissionless fine-tuning and straightforward monetization will be the main drivers of such a platform.

Is that all?

No, we look forward to seeing teams develop: non-language foundational models, alternatives to transformers, differentiable and remixable memories, mortal AI, domain-specific reinforcement learning gyms, forward-forward training and fine-tuning, proof of inference, and more.

Conclusion

The convergence of AI and crypto in 2025 presents unprecedented opportunities for value creation and capture. We stand at a unique moment where the democratization of AI through blockchain technology can reshape how we build, deploy, and monetize artificial intelligence.

This investment thesis demonstrates how the convergence of AI and crypto creates a powerful feedback loop:

• Pre-training and fine-tuning innovations reduce barriers to entry, enabling more participants to create and own models

• Inference-time compute introduces new economic dynamics that require novel market mechanisms and risk management tools

• Memory systems and agent frameworks create opportunities for value capture through personalization and automation

Each of these components reinforces the others - as models become more accessible, agent development accelerates; as memory systems improve, fine-tuning becomes more effective; as inference costs vary, new market mechanisms emerge.

For investors and builders in this space, the message is clear: the infrastructure for the next generation of AI is being built now. The winners will be those who recognize that decentralized, permissionless, and transparent systems are not just ideological preferences - they are practical necessities for unlocking the full potential of artificial intelligence.

cyber•Fund mission is to accelerate the buildout of healthy cybernetic economy. If you're working on any of these problems, please reach out. Our goal is to support founders, researchers and open source projects in decentralized and distributed AI.