Crypto and AI Are Far From Meaningful Integration, IC3 Survey Concludes 

A 155-page survey published June 8 by the Initiative for CryptoCurrencies and Contracts (IC3) concludes that crypto and AI remain in the very early stages of meaningful integration, finding that the industry has consistently prioritized demonstrating feasibility over demonstrating value across nearly every major area of the field.

IC3 Survey Identifies Feasibility-Value Gap as the Field’s Central Problem

The distinction between feasibility and value runs through almost every major finding in the document. It is the difference between showing that something can be done and showing that doing it this way is better than the alternatives. 

On that second question, the survey finds the crypto-AI field largely silent. The document’s executive summary lists two proof points it says the crypto community still needs to produce. The first concerns decentralized AI infrastructure, where the survey states the industry “has largely focused on demonstrating the feasibility of training large models in a decentralized fashion,” while “opportunities to compete with centralized AI platforms on cost for specific use cases and regimes need to be better quantified.” 

The second concerns agentic payments, where the survey calls on the crypto community to “quantitatively demonstrate the benefits of crypto for agentic payments, rather than only demonstrating feasibility.”

DePIN Networks Show Cheaper Hardware but Lack Benchmarks to Prove Cheaper AI Jobs

The survey’s treatment of decentralized physical infrastructure networks, known as DePINs, illustrates the gap in concrete terms. The authors acknowledge that renting a DePIN GPU can be substantially cheaper than renting a comparable cloud service, but note that cheaper hardware does not automatically produce cheaper AI jobs. 

The survey states that “the throughput and latency requirements of an AI job can significantly impact the overall cost of a job” when nodes communicate over the public internet, and that “very large AI jobs, such as training frontier models, are typically throughput-bound.” 

Its conclusion is direct: “Today, direct cost comparison is challenging because we lack systematic benchmarks that profile AI jobs on DePIN networks and compare them to traditional cloud infrastructure.”

Industry Proofs-of-Concept Advance Model Sizes but Skip Total Cost Reporting

Industry efforts have produced measurable results. By August 2024, Macrocosmos AI had trained 700M and 7B parameter models on the Bittensor network. Prime Intellect trained a 10B parameter model asynchronously in November 2024. 

The largest model being pre-trained on DePIN infrastructure at the time of writing, according to the survey, was a 40B parameter model on the Psyche Network. The survey’s assessment of these efforts is limited, however: “a broader point about existing industry proofs-of-concept is that they do not typically report total cost metrics.” 

Without standardized cost models comparing end-to-end training costs on centralized versus decentralized infrastructure, the survey states, the competitive case for decentralized AI remains unquantified.

Five Widely Circulated Industry Claims Examined and Found Unsupported

The survey’s closing chapter, titled “Misconceptions and Half-Truths,” examines five broadly circulated claims in the crypto-AI space and finds each either false in its strong form or unsupported by current evidence. 

The first is that blockchains can help distinguish AI-generated content from human-generated content. The survey finds a blockchain can guarantee the integrity of a provenance record after submission but “cannot guarantee that the information was true at the moment it was recorded.” 

The authors note that a user can photograph an AI-generated image with a C2PA-compliant camera, producing valid cryptographic credentials identifying the output as an authentic human photograph, and conclude that blockchain’s role “is limited to preserving claims about content, not resolving the broader challenge of distinguishing human from AI-generated material.”

The second claim is that decentralization can solve bias and fairness problems in AI. The survey finds algorithmic bias “arises inherently in the training process and is typically mitigated by revised training or inference techniques,” meaning “decentralization does not address the source of the problem.” 

The third is that giving AI agents a crypto wallet makes them autonomous. The survey distinguishes between intelligence autonomy and execution autonomy and argues that agentic wallets deliver neither, stating that “AI systems do not become more intelligent by possessing a wallet. Nor do they become more resistant to human manipulation or shutdown.” It further notes that “blockchains are not uniquely required for such automation. Centralized financial infrastructure can be, and has been, accessed programmatically by AI agents.”

The fourth claim is that recording model data provenance on a blockchain produces trustworthy AI. The survey identifies three reasons this does not follow, including that “the non-determinism inherent in most stochastic training renders it infeasible to verify model weights against a training pipeline, even in principle.” 

The fifth claim is that decentralization inherently makes AI more cost-effective, which the survey finds unsubstantiated, concluding that “we do not have enough data to predict when a job will be cheaper on existing DePIN or DeAI platforms vs. a centralized cloud service provider.”

Survey Identifies Smart Contract Fraud Detection and Agentic Payments as Defensible Use Cases

The survey is not uniformly skeptical. It identifies AI-based approaches for detecting fraudulent or buggy smart contracts as having produced results, though it notes these techniques are “most effective in controlled settings with ample training data.”

Zero-knowledge proofs and trusted computing environments repurposed to make AI inference less susceptible to tampering are described as among the more defensible use cases at the current moment. 

The authors also identify agentic payments as a genuine opportunity, specifically because crypto infrastructure can offer properties including neutrality and censorship resistance that centralized payment systems cannot. The survey frames this not as an established advantage but as an opportunity the community needs to quantify.

Survey Editors Describe Combining Crypto and AI as Soldering Jell-O

The survey’s editors, Giulia Fanti of Carnegie Mellon and Ari Juels of Cornell Tech, frame the field’s core structural challenge in direct terms. Juels is quoted saying in related materials:  

“Crypto is a ‘hard’ technology, built on cryptographic primitives with rigorous security properties and programs that enforce unambiguous results. AI is a ‘soft’ technology: No one fully understands or can fully trust the models on which it depends. Combining the two naively can be like soldering Jell-O.” 

The survey, with contributions from researchers at Cornell Tech, Carnegie Mellon, Princeton, Yale, Technion, and other institutions, maps where the research supports commercial claims and where the gap between the two has yet to be closed, concluding that the field remains in its early stages and that the work of closing that gap remains largely ahead of it.