Artificial intelligence represents the largest technological revolution in history, igniting a global race unlike anything we’ve seen. Current AI models already score in the top percentiles of standardized university exams and outperform humans in numerous tasks—including AI research itself. Even at its current stage, AI is reshaping industries like search, customer service, content creation, programming, and education.
We expect AI’s capabilities, funding, and societal influence to accelerate from here. Every major tech company recognizes AI’s critical importance and is investing accordingly. NVIDIA’s revenue—a key proxy for AI capital expenditure—is projected to exceed $100 billion in 2024, more than doubling 2023’s results and quadrupling figures from the year before.
As Sundar Pichai, CEO of Google, noted:
“The risk of under-investing here is far greater than the risk of over-investing.”
Startups, too, see AI as a disruptive force capable of challenging long-standing incumbents. An estimated $83 billion has been invested in AI startups over the past 18 months alone.
Given that AI capabilities often grow exponentially with increased computing power, it’s plausible that we could achieve something resembling Artificial General Intelligence (AGI) within the next decade.
In this article, we argue that competitive dynamics will lead to a world of millions of models—and that crypto is the ideal foundation for this multi-model future. We’ll explore why a multi-model ecosystem is the logical endgame for AI, highlight the unique advantages crypto brings to AI, and outline the emerging crypto x AI tech stack with real-world examples.
Why a Multi-Model Future Is Inevitable
Today, we’re moving toward a world dominated by a handful of large, vertically integrated tech companies producing all-powerful “god models.” However, we believe this is not the endgame—and for several compelling reasons:
- Single-Point Risk: Organizations and developers building AI-powered experiences don’t want to depend on a single closed-source provider that can alter models, change terms of service, or discontinue access entirely.
- Cost vs. Capability Trade-Offs: The extremely large, general-purpose models favored by big tech are inherently expensive to train and run. This makes them overkill—and uneconomical—for many applications. As AI scales, users will optimize for cost efficiency. In many cases, smaller specialized models will outperform larger general ones. Research supports this across domains like medical imaging, fraud detection, speech recognition, and more.
- Vertical Integration: As Apple has repeatedly demonstrated, the best products often come from vertically integrated systems. Entrepreneurs will seek competitive advantages by building on their own specialized models, allowing them to capture more value and attract greater investment.
- Privacy Concerns: AI will become deeply embedded in organizational workflows, handling sensitive data. Many organizations will hesitate to entrust that data to external models.
For these reasons, we believe the future will consist of many smaller, specialized models tailored to specific use cases. Developers will fine-tune open-source models like LLaMA or Mistral AI using proprietary data. Some models will run on servers; others will operate locally on user devices for privacy or on decentralized networks for censorship resistance.
This is a world of modular AI “Legos,” where developers compete to deliver value and users mix and match services to meet their needs. New infrastructure—routing, coordination, synthesis, payment systems—will be needed to deconstruct the “god model” stack and serve this new AI economy.
It’s also a world where crypto thrives.
How Crypto Unlocks a Better AI Future
Crypto intuitively feels like a natural fit for a multi-model AI world—but hype has led to significant funding for projects that may lack substance. This makes it difficult to separate real value from speculation. However, we believe crypto is not a meme; it offers unique advantages that enable better products—some of which would be impossible without it.
Here are the key properties crypto brings to AI:
- Coordination Layer: Crypto excels at facilitating collective coordination without central control. It solves chicken-and-egg problems through cryptonative incentives, rapidly bootstrapping new networks of users and resources.
- Open, Permissionless APIs: Crypto networks serve as open, permissionless APIs accessible to anyone, anywhere, without KYC, credit cards, or third-party approval. This is essential for autonomous AI agents that need to access services, deploy code, and transfer value without human intervention.
- Trust Minimization: Crypto offers cryptographic guarantees that rules won’t change, access won’t be revoked, and execution can be verified. This is critical in a modular AI stack where users must trust many external services.
- Censorship Resistance: Applications deployed on crypto networks are unstoppable. As AI grows more powerful, governments will likely seek to control it. Crypto x AI offers a form of unstoppable intelligence, independent of existing power structures.
The Crypto x AI Tech Stack
With these benefits in mind, let’s explore the most promising areas where crypto and AI intersect.
Decentralized Compute and Data Centers
Compute utility is generally divided into training and inference. Decentralized compute offers advantages in both areas.
Decentralized Training: Training across distributed nodes is technically challenging due to communication and latency requirements. However, several teams are making progress. Solving this—while ensuring quality and responsibility on a permissionless network—could unlock global marginal cost advantages. Token incentives may be essential for bootstrapping supply, potentially even giving compute providers ownership in the models they help train.
Verifiable Inference: This involves using AI to extend trust-minimized systems. Models can run off-chain, publishing proofs that they operated as expected. This lets projects move governance decisions (e.g., risk parameters in a money market) to an off-chain model trustlessly. It also provides users assurance that outputs came from the intended model—a critical feature as AI is applied to more important tasks.
Data
Training LLMs is a multi-step process requiring diverse data and human intervention. Crypto networks can help teams acquire the data and resources they need at every stage.
Data Collection: Crypto incentives can effectively bootstrap the supply side of data collection. For example, Grass incentivizes users to share idle internet bandwidth, helping scrape and structure web data for AI training. Similarly, Hivemapper has mapped 25% of the world’s roads by rewarding contributors with tokens. This model could be applied to other multimodal data types.
Data Marketplaces: Many companies possess valuable data but lack the scale or connections to monetize it. A well-designed data marketplace can connect these suppliers with AI labs in a standardized way. The main challenges include ensuring data quality and managing API compatibility.
Data Preparation: Tasks like labeling, cleaning, and enrichment are often outsourced. Centralized companies like Scale AI dominate this space today, but crypto-based markets and workflow systems could compete effectively by offering better incentives and transparency.
Models
As noted in Delphi Digital’s “The Tower & The Square,” AI model production is controlled almost entirely by big tech and governments. This is a potentially dystopian outcome—one that crypto can help prevent.
Decentralized model projects aim to create neutral, censorship-resistant, and accessible models by decentralizing every part of the model creation process: sourcing data, training, inference, and governance. The goal is a community-owned model outside the control of any “tower.”
Building a decentralized model that competes with state-of-the-art centralized ones is extremely difficult. These projects are moon shots—but if they succeed, they could be immensely valuable.
It’s also worth noting that centralized AI labs embracing crypto principles may tokenize or leverage crypto rails. Examples include NousResearch and Bittensor, which operates as a model creation infrastructure platform.
Applications
As Eric Schmidt recently stated:
“If TikTok is banned…tell your LLM: ‘Copy TikTok, steal all the users, steal all the music, put my preferences in, make this program in 30 seconds, launch it, and if it’s not viral in an hour, do something else.’”
This vision points toward the incredible power of AI agents. But for these agents to operate autonomously, they must access services, transfer value, and form economic relationships without human intervention.
Traditional banking apps, KYC, and registration processes are ill-suited for them. They will inevitably encounter systems designed for humans.
Crypto rails provide the perfect platform. They offer a permissionless, trust-minimized, and censorship-resistant foundation for agents to operate. If an agent needs to deploy an app, it can do so on-chain. If it needs to make a payment, it can send tokens. The code and data of on-chain services are public and uniform, so agents can understand and interact with them without APIs or documentation.
Agents can also act as catalysts for on-chain activity. Shifting the user experience from clicking buttons to interacting through AI assistants could abstract away crypto’s notorious onboarding complexity—removing a major barrier to new users.
Projects like Wayfinder, Autonolas, and Dain are already building toward this future.
Frequently Asked Questions
What is a multi-model AI ecosystem?
A multi-model ecosystem consists of many smaller, specialized AI models tailored to specific tasks, rather than a few giant “god models.” This approach offers better cost efficiency, privacy, and customization compared to one-size-fits-all alternatives.
Why is crypto important for AI?
Crypto provides unique advantages for AI, including permissionless access, trust minimization, censorship resistance, and innovative incentive models. These features enable new types of AI applications and infrastructure that are harder to build using traditional systems.
Can decentralized AI models compete with centralized ones?
It’s challenging, but possible. Decentralized models may initially lag behind centralized counterparts in performance, but they offer advantages in transparency, customization, and alignment with user interests. Over time, as the technology matures, they could become competitive in specific domains.
What are AI agents, and how do they use crypto?
AI agents are autonomous programs that perform tasks without human intervention. They use crypto to access services, make payments, and deploy code on permissionless networks, enabling fully autonomous operation in a trust-minimized environment.
How do data marketplaces work in crypto AI?
Crypto-based data marketplaces use token incentives to reward data contributors and create liquid markets for AI training data. This allows smaller data owners to monetize their assets and provides AI developers with access to diverse, high-quality datasets.
What is verifiable inference?
Verifiable inference involves running AI models off-chain while generating cryptographic proofs that the computation was performed correctly. These proofs are then verified on-chain, enabling trust-minimized use of AI in blockchain applications.
Conclusion
AI is poised to become the most powerful and socially significant resource of the 21st century. Allowing it to be controlled solely by big tech and governments is a dystopian outcome we should avoid. Crypto offers a path to prevent this monopoly—not by relying on philosophical arguments, but by providing genuinely better solutions for developers and users.
We are still in the early stages of both AI and decentralized AI. Much work remains to realize the vision outlined here. At Delphi Labs, we are excited about the future of crypto and AI and are committed to shaping it in collaboration with the best builders in the space.
If you’re an ambitious entrepreneur or founder who believes in the future of AI x Web3, we invite you to explore opportunities like the NEAR x Delphi Labs AI Accelerator. The space is young, and the time to build is now.