Coinbase Ventures: A look at the Crypto x AI stack

Written by Jonathan King, Head of Coinbase Ventures

Compiled by: 0xjs@ Jinse Finance

The future of AI can be built on blockchain technology, as cryptography helps increase accessibility, transparency, and application scenarios in emerging technology areas. The fusion of the efficiency, borderless nature, and programmability of cryptocurrency with AI has the potential to change the way humans and machines interact with the digital economy, including enabling users to have sovereignty over their personal data. This includes the rise of "agent networks" in which AI agents running on cryptographic infrastructure can drive economic activity and growth.

Disclosures and Notes: Coinbase Ventures portfolio companies are noted with an asterisk (*) when first mentioned in this article.

So what does this look like? AI agents transact on crypto infrastructure. Software code created by AI, including smart contracts, leads to a surge in on-chain applications and experiences. Users own, manage, and profit from the AI ​​models they contribute to. AI is used to improve user and developer experience within the crypto ecosystem, enhance smart contract capabilities, and create new application scenarios. And so on.

As we envision this crypto x AI future, today we will reveal our core thesis about the future of this transformative technology convergence. The key points are as follows:

• We believe that cryptocurrency/blockchain technology does not have to be used to enhance capabilities or solve emerging challenges at every layer of the AI ​​technology stack. Instead, cryptocurrency can play an important role in bringing broader distribution, verifiability, censorship resistance, and native payment channels to AI, while benefiting from AI mechanisms to power new user experiences on-chain.
• Crypto x AI can give rise to “agent networks”, a transformative paradigm where AI agents running on crypto infrastructure can become important drivers of economic activity and growth. We foresee a future where agents have their own crypto wallets to autonomously transact and fulfill user intent, access low-cost, decentralized computing and data resources, or use stablecoins to pay humans and other agents for completing tasks required to achieve their overall objective function.
• Initial arguments supporting this thesis include: (1) cryptocurrencies will become the preferred payment channel for commercial activities between agents and humans and between agents; (2) generative AI and natural language interfaces will become the primary interaction methods for users seeking to conduct transactions on-chain; and (3) AI will create the vast majority of software code (including smart contracts), triggering a "Cambrian explosion" of on-chain applications and experiences.
• Crypto x AI consists of two core sub-areas: (1) Decentralized AI (cryptocurrency -> AI), defined as building general AI infrastructure to inherit the characteristics of modern peer-to-peer blockchain networks; and (2) On-chain AI (AI -> cryptocurrency), defined as building infrastructure and applications that leverage AI to power new and existing use cases.
• The Crypto x AI ecosystem can be divided into the following levels: (1) Computing (i.e., networks focused on providing potential graphics processing units (GPUs) for AI developers); (2) Data (i.e., networks that enable decentralized access, compilation, and verifiability of AI data pipelines); (3) Middleware (i.e., networks/platforms that enable the development, deployment, and hosting of AI models/agents); (4) Applications (i.e., user-oriented products (B2B or B2C) that utilize on-chain AI mechanisms)

At Coinbase, we are on a mission to help update the financial system, making it more secure and reliable while increasing accessibility and usability for consumers and builders. We believe Cryptox AI will play a big role in that. In this blog, we will dive into the why, how, and next steps of Cryptox AI.

Crypto x AI Introduction

The AI ​​market has seen significant growth and investment, with VC firms pouring nearly $290 billion into the space over the past five years. The World Economic Forum notes that AI technology could increase annual U.S. GDP growth by 0.5-1.5% over the next decade. AI applications are showing strong traction, with apps like ChatGPT4 setting new records in user growth/adoption. However, as the AI ​​market rapidly evolves, several challenges have emerged, including data privacy concerns, demand for AI talent, ethical considerations, centralization risks, and the rise of deepfakes. These challenges drive current discussions in the Crypto x AI space as stakeholders seek solutions that leverage the strengths of both technologies to address these emerging issues.

From Vitalik Buterin's blog about Crypto x AI

Crypto x AI combines the decentralized infrastructure of blockchain and the ability of AI to mimic human cognitive functions and learn from data, creating a synergy that could revolutionize industries. Blockchain redefines system architecture, data/transaction verification, and distribution. AI enhances data computing, analysis, and provides new content generation capabilities. This intersection has sparked both excitement and skepticism among developers in both technology communities, driving the exploration of new application scenarios and potentially accelerating adoption in both fields in the long run. While both cryptocurrency and AI are general terms covering a variety of different technologies and topics, we believe the intersection of the two fields can be broken down into two core sub-fields:

• Decentralized AI (Crypto -> AI) enhances AI capabilities through a permissionless and composable infrastructure of cryptocurrency. This unlocks some application scenarios, such as democratized access to AI resources (such as computing, storage, bandwidth, training data, etc.), collaborative, open source model development, verifiable reasoning, or immutable ledgers and cryptographic signatures for content origin and authenticity.
• On-chain AI (AI -> Crypto) brings the benefits of AI to the crypto ecosystem, improving user and developer experience through large language models and natural language interfaces, or enhancing smart contract capabilities. On-chain AI uses two approaches: (1) developers integrate AI models or agents into their smart contracts and on-chain applications; (2) AI agents use crypto channels (such as self-hosted wallets, stablecoins, etc.) to make payments and delegate decentralized infrastructure resources.

While both subfields are still in their infancy, the potential of “cryptocurrency-infused AI” or “AI-infused cryptocurrencies” is enormous and is expected to unlock a range of yet-to-be-imagined application scenarios, especially as computing infrastructure and intelligence speeds continue to improve.

Crypto x AI: The key unlocking factor of the “network of intelligent agents”

One aspect of the intersection of cryptocurrency and AI that we find particularly exciting is the concept of AI agents running on crypto infrastructure. This integration aims to create “agent networks,” a transformative paradigm that enables enhanced security, efficiency, and collaboration in AI-driven economies, underpinned by strong incentive structures and cryptographic primitives.

We believe AI agents can become important drivers of economic activity/growth and primary “users” of applications (both on-chain and off-chain), gradually replacing human users in the medium to long term. This paradigm shift will force many Internet-native companies to rethink their core assumptions about the future and provide the necessary products, services, and business models to better serve an agent-dominated economy. That being said, we do not believe that cryptocurrency/blockchain technology must be used to enhance capabilities or solve emerging challenges at every layer of the AI ​​technology stack. Instead, cryptocurrency can play an important role in bringing broader distribution, verifiability, censorship resistance, and native publishing channels to AI, while benefiting from AI mechanisms to power new user experiences on-chain.

The initial arguments supporting this argument are as follows:

• Cryptocurrency will become the preferred payment channel for commercial activities between agents and humans and between agents: Cryptocurrency is the programmable currency native to the Internet and has multiple advantages for driving agent-based economies. As AI agents become more autonomous and conduct micro-transactions at scale (such as paying for inference, data, API access, decentralized computing or data resources, etc.), the efficiency, borderless nature and programmability of cryptocurrency will make it a better medium of exchange than traditional fiat currency channels. In addition, agents will need unique and verifiable identities (i.e. "know your agent") to ensure compliance with regulatory rules and compliance requirements when transacting with businesses and end users. Low-fee blockchains, smart contracts, self-hosted wallets (such as Coinbase AI Wallets), and stablecoins can help simplify and reduce the cost of complex financial agreements between agents, while the verifiability and immutability of decentralized networks will ensure the importance and auditability of AI agent transactions.
• Generative AI and natural language interfaces will become the primary interaction method for users seeking to transact on-chain: As natural language processing speed and AI's contextual understanding of cryptocurrencies improve, interacting on-chain through conversational interfaces will become the default norm and expectation for users, which is consistent with current Web2 trends such as ChatGPT. Users only need to describe their desired transaction intentions in natural language (such as "exchange X for Y"), and artificial intelligence agents will translate these intentions into verifiable smart contract code, providing the most efficient and cost-effective transaction execution path.
• AI will create the vast majority of software code (including smart contracts), triggering a "Cambrian explosion" of on-chain applications and experiences: AI's code generation capabilities are rapidly developing in Web2 (such as Devin, Replit) and are fundamentally changing the software development paradigm. We believe that this shift will soon take center stage in the crypto space, with the near-term focus on significantly lowering the barrier to entry for new and existing developers. However, the future state is that AI "software agents" generate smart contracts and highly personalized applications from scratch in real time based on user preferences, and store and verify them on-chain.

These insights suggest that the boundaries between AI and cryptocurrency will become increasingly blurred in the future, creating a new paradigm of intelligent, autonomous, and decentralized systems. Based on this framework, let’s take a deeper look at the technology stack that enables the fusion of cryptocurrency and AI.

Opportunities in the Crypto x AI Stack (Current)

The quest to “infuse crypto with AI” or “infuse AI with crypto” has spawned a new and complex field that is developing rapidly, with many builders eager to take advantage of market momentum. Today, we believe that the Crypto x AI space can be divided into the following tiers: (1) Compute (i.e., networks focused on providing potential graphics processing units (GPUs) for AI developers); (2) Data (i.e., networks that enable decentralized access, orchestration, and verifiability of AI data pipelines); (3) Middleware (i.e., networks/platforms that enable the development, deployment, and hosting of AI models/agents); (4) Applications (i.e., user-facing products (B2B or B2C) that leverage on-chain AI mechanisms)

calculate

Both AI model training and reasoning execution require a lot of computing GPU resources. As AI models become more complex and the demand for computing increases, advanced GPUs like NVIDIA are in short supply, resulting in long wait times and increased costs. Decentralized computing networks are emerging as a potential solution to these challenges, in the following ways:

• Building a permissionless marketplace for buying, renting, and hosting physical GPUs
• Build a GPU aggregator that enables anyone (e.g. Bitcoin miners) to contribute their excess GPU computing power to perform on-demand AI tasks and receive token incentives in return
• Financialize physical GPUs by tokenizing them as digital assets on the chain
• Develop distributed GPU networks for compute-intensive workloads (e.g., training, inference)
• Creating infrastructure that enables AI models to run on personal devices (similar to decentralized Apple smartphones)

These proposed solutions all aim to increase GPU computing supply and accessibility while offering very competitive pricing. However, given that most players in this space have varying levels of support for advanced AI workloads, face challenges related to the lack of co-location of GPUs, and in some cases lack developer tooling and uptime guarantees comparable to centralized alternatives, we believe these products are unlikely to see mainstream adoption in the near to medium term. Emerging areas and example projects building at this layer include the following:

• General computing: decentralized computing marketplaces that provide GPU computing resources that can be used for a variety of applications (e.g. Akash, Aethir)
• AI/ML Computing: Decentralized computing networks (e.g., io.net, Gensyn, Prime Intellect, Hyperbolic, Hyperspace) that provide GPU computing resources for specific services (e.g., GPU aggregators, distributed training and inference, GPU tokenization, etc.)
• Edge computing: Compute and storage networks that power large language models on-device for personal, contextualized reasoning (e.g. PIN AI, Exo, Crynux.ai, Edge Matrix)

data

Scaling AI models requires ever-growing training datasets, and large language models are trained from trillions of words of human-generated text. However, currently public, human-generated data is limited (Epoch AI estimates that high-quality language/data sources may be exhausted by 2024), which raises the question of whether the lack of training data will become a major bottleneck in the performance of AI models, potentially causing their performance to stagnate. Therefore, we believe that data-focused crypto x AI companies have the following opportunities to address these challenges:

• Incentivize users to share their private/proprietary data (e.g. Data DAOs — on-chain entities where data contributors can see economic benefits from contributing their private data to social platforms and manage how the data is used and monetized)
• Create tools for generating synthetic data assets from natural language prompts, or provide user incentives to scrape data from public websites
• Incentivize users to help preprocess datasets for training models and maintain data quality (e.g. data annotation/reinforcement learning from human feedback)
• Build a multi-sided, permissionless data marketplace where anyone can be compensated for their contribution

These opportunities have given rise to many of the emerging players we see in the data layer today. However, it is worth noting that centralized incumbents have existing network effects and proven data compliance regimes that traditional businesses value in the AI ​​model lifecycle, which may leave little room for decentralized alternatives. Nonetheless, we believe that the data layer of decentralized AI presents a significant long-term opportunity to address the "data wall" challenge. Emerging areas and example projects building on this layer include the following:

• Data Marketplace: A decentralized data exchange protocol designed for data providers and consumers to share and trade data assets (e.g. Ocean Protocol, Masa, Sahara AI)
• User-owned/Private Data (including DataDAOs): Networks that incentivize the collection of proprietary datasets (including private user-owned data) (e.g. Vana*, NVG8)
• Public & Synthetic Data: Networks/platforms for scraping data from public websites or generating new datasets via natural language prompts (e.g. Dria, Mizu, Grass, Synesis One)
• Data Intelligence Tools: Platforms and applications for querying, analyzing, visualizing, and providing actionable insights about on-chain data (e.g. Nansen*, Dune*, Arkham, Messari*)
• Data storage: File storage networks for long-term data storage/archiving and relational database networks for managing frequently accessed and updated structured data (e.g. Filecoin, Arweave*, Ceramic*, Tableland*)
• Data compilation/provenance: Networks and platforms that optimize data ingestion pipelines and processing for AI and data-intensive applications, and ensure correct provenance tracking and verifiable authenticity of AI-generated content (e.g., Space and Time, The Graph*, Story Protocol)
• Data annotation: Networks and platforms that improve reinforcement learning and fine-tuning mechanisms for AI models by incentivizing distributed human contributors to create high-quality training datasets (e.g. Sapien, Kiva AI, Fraction.AI)
• Oracles: Networks that use AI to provide verifiable off-chain data for on-chain smart contracts (e.g. Ora, OpenLayer, Chainlink)

middleware

To realize the full potential of an open, decentralized AI model or agent-based ecosystem, new infrastructure needs to be built. Some high-potential areas that builders are exploring include the following:

• Leverage large language models with public weights to power on-chain AI applications, while building foundational models that can quickly understand, process, and act on on-chain data
• Distributed training solutions for large base models (e.g. 100B+ parameters); this is often seen as a distant dream due to various technical complexities, but recent breakthroughs from Nous Research, Bittensor, and Prime Intellect are trying to change this
• Leverage zero-knowledge or optimistic machine learning (i.e. zkML, opML), trusted execution environments (TEEs), or fully homomorphic encryption (FHE) to enable private, verifiable reasoning
• Enable open, collaborative AI model development through resource coordination networks, or build agent networks/platforms that leverage crypto infrastructure channels to enhance the potential of AI agents in on-chain/off-chain application scenarios

While some progress has been made in building these foundational infrastructure primitives, production-ready, on-chain large-scale language models and AI agents are still in their infancy, and we do not expect this to change until the compute, data, and model infrastructure matures. Nonetheless, we believe this category is very promising and is a core focus of Coinbase Ventures’ investment strategy in this space, driven by the long-term implied growth and demand for AI services. Emerging areas and example projects building at this layer include the following:

• Open source large language models: open source publicly accessible AI models that allow anyone to use, modify, and freely distribute them (e.g. LLama3, Mistral, Stability AI)
• On-chain model creators: networks and platforms that can create basic large language models for on-chain application scenarios (e.g. Pond*, Nous, RPS)
• Training and fine-tuning: Networks and platforms that enable incentivized, verifiable training or fine-tuning mechanisms on-chain (e.g. Gensyn, Prime Intellect, Macrocosmos, Flock.io)
• Privacy: Networks and platforms that use privacy-preserving mechanisms for the development, training, and inference of AI models (e.g. Bagel Network, Arcium*, ZAMA)
• Inference networks: Networks and platforms that use cryptography/proofs to verify the correctness of AI model outputs (e.g. OpenGradient*, Modulus Labs, Giza, Ritual)
• Resource coordination networks: Networks designed to facilitate resource sharing, collaboration, and coordination for AI model development (e.g. Bittensor, Near*, Allora, Sentient)
• Agent Networks and Platforms: Networks and platforms used to facilitate the creation, deployment, and monetization of AI agents in on-chain/off-chain environments (e.g. Morpheus, Olas, Wayfinder, Payman*, Skyfire*)

application

AI agents are beginning to make their mark in the crypto space, with early examples like Dawn Wallet (a crypto wallet that leverages AI agents to send transactions and interact with protocols on behalf of users), Parallel Colony* (an on-chain game where players collaborate with AI agents that own their own wallets and can create their own paths in the game), or Venice.ai (a generative AI application with verifiable reasoning and privacy-preserving mechanisms/natural language prompts). However, application development is still largely experimental and opportunistic, with various application ideas emerging in large numbers amidst the hype in the space. Nonetheless, we believe that advances in AI agent infrastructure and frameworks will enable the crypto industry to move from primarily reactive smart contract applications to more complex proactive applications in the medium to long term. Emerging areas and example projects building on this layer include the following:

• AI Companion: Applications for creating, sharing, and monetizing user-owned AI models and agents with personalization and contextual awareness (e.g. MagnetAI, MyShell, Deva, Virtuals Protocol)
• Natural Language Processing-based interfaces: Applications that use natural language prompts as the primary interface/entry point for interacting with and executing on-chain transactions (e.g. Venice.AI, Veldt)
• Development/Security Tools: Applications/tools for developers that use AI models/agents to enhance on-chain developer experience and security mechanisms (e.g. ChainGPT, Guardrail*)
• Risk Agents: Services that use machine learning models or AI agents to help protocols dynamically adjust and respond to on-chain risk parameters in real time (such as Chaos Labs*, Gauntlet*, Minerva*)
• Identity (Proof of Person): Applications that use cryptographic proofs and machine learning models to verify user proof of personhood (e.g. Worldcoin*)
• Governance: Applications that use AI agents to execute transactions based on human-driven governance decisions/feedback (e.g. Botto, Hats)
• Trading/Decentralized Finance: AI-driven trading infrastructure and decentralized finance protocols that use AI agents to automatically execute on-chain transactions (e.g. Taoshi, Intent.Trade)
• Games: On-chain games that utilize intelligent non-player characters or AI mechanics to drive core gameplay mechanics (e.g. Parallel*, PlayAI)
• Social: Applications that use AI mechanisms to drive social experiences on the chain (such as KaiKai, NFPrompt)

in conclusion

While the Crypto x AI stack is still in its infancy, we believe there will be significant progress in decentralized AI infrastructure, on-chain AI applications, and the emergence of a “network of agents” where AI agents will become the primary driver of economic activity. While challenges remain in areas such as computing infrastructure and data availability, the synergy between cryptocurrencies and AI is likely to accelerate innovation in both fields, leading to more transparent, decentralized, and autonomous systems. As the field continues to evolve rapidly, driven by new teams receiving funding and more mature teams working to find product/market fit, it will be critical for Internet-native companies and developers to adapt to the changing paradigm and embrace the potential of Crypto x AI to create new applications and experiences that were previously unimaginable.

Overall, Coinbase Ventures is excited about the future potential and opportunities in Crypto x AI, and we are actively investing in every layer of the architecture.