Written by: Hack VC
Compiled by: TinTinLand
AI is one of the hottest and most promising tracks in the crypto market recently:
Decentralized AI Training
GPU DePINs
Uncensored AI Models
Are these breakthroughs or just hype? Hack VC dissects the top Crypto✖️AI integration ideas to explore the real challenges and opportunities!
Ideas with potential and challenges
First, let’s start with the promise of Web3 AI — there’s a fair amount of hype around these ideas, but the reality may not be that great.
Idea #1: Decentralized AI Training
The problem with doing AI training on-chain is that training requires high-speed communication and coordination between GPUs, as neural networks need backpropagation when training. Nvidia has two innovations for this (NVLink and InfiniBand). These technologies effectively make GPU communication faster, but they are limited to GPU clusters within a single data center (50+ Gbps speeds).
If you introduce a decentralized network, speeds suddenly drop by orders of magnitude due to increased network latency and bandwidth. This is a non-viable option for AI training use cases compared to Nvidia’s high-speed interconnects within data centers.
Of note, here are some innovations that may offer new possibilities in the future:
Distributed training over InfiniBand is happening at scale, as NVIDIA itself supports distributed, non-local training via the NVIDIA Collective Communications Library. However, this is still early days, so adoption metrics are yet to be determined. Bottlenecks due to physical distance still exist, so local training over InfiniBand is still significantly faster.
There has been some novel research published recently that shows that it may be possible to make decentralized training more practical in the future by relying on less communication synchronization.
Smart sharding and model training scheduling can also help improve performance. Likewise, new model architectures may be designed specifically for decentralized infrastructure in the future (Gensyn is researching these areas).
The data part of training is also a challenge. Any AI training process involves processing large amounts of data. Typically, models are trained on centralized and secure data storage systems with high scalability and performance. This requires transferring and processing terabytes of data, and it is not a one-time cycle. Data is often noisy and contains errors, so it must be cleaned and converted into a usable format before training the model. This stage involves repetitive tasks of normalization, filtering, and handling missing values. These all pose serious challenges in a decentralized environment.
The data part of training is also iterative, which is not very suitable for Web3. It took OpenAI thousands of iterations to achieve its results. The most basic task scenario for a data science expert in an AI team includes defining the goal, preparing data, analyzing and structuring the data to extract important insights and making it suitable for modeling. Then, a machine learning model is developed to solve the defined problem and its performance is verified using a test dataset. This process is iterative: if the current model does not perform as expected, the expert will return to the data collection or model training stage to improve the results. Now, imagine this process in a decentralized environment, where the best existing frameworks and tools in Web3 are not easily accessible.
Another problem with training AI models on-chain is that it is a less interesting market compared to inference. Currently, a lot of GPU computing resources are used for AI LLM training. But in the long run, inference will become a more common use case for GPUs. Think about it: how many AI LLMs need to be trained to satisfy users around the world, and how many customers are using these models?
One solution that is making progress on all fronts is 0g.ai, which provides on-chain data storage and data availability infrastructure. Their ultra-fast architecture and ability to store large amounts of data on-chain allows for fast, iterative on-chain AI model training.
Idea #2: Use excessively redundant AI reasoning calculations to reach consensus
One of the challenges of CryptoxAI is verifying the accuracy of AI reasoning, because you cannot fully trust a single centralized party to make reasoning, and there may be a risk of nodes misbehaving. In Web2 AI, this challenge does not exist because there is no decentralized consensus system.
One proposal to address this problem is redundant computing, where multiple nodes repeat the same AI inference operations so that they can operate in a trustless manner with no single point of failure.
The problem with this approach is that we live in a world where there is a severe shortage of high-end AI chips. There are multi-year waiting periods for high-end NVIDIA chips, which drives up prices. If you also need to have your AI inference re-executed multiple times on multiple nodes, you are now multiplying these expensive costs. This is not feasible for many projects.
Idea #3: Web3-specific AI use cases in the short term
Some have suggested that Web3 should have its own unique AI use cases, specifically for Web3 users. This could be, for example, a Web3 protocol that uses AI to risk-score DeFi pools, a Web3 wallet that suggests new protocols based on your wallet history, or a Web3 game that uses AI to control non-player characters (NPCs).
Currently, this is a nascent market with many use cases still to be discovered. Some challenges include:
Web3 native use cases require less AI transactions because market demand is still in its infancy.
There are fewer customers because there are orders of magnitude fewer Web3 customers compared to Web2 customers, so the market is less fragmented.
The clients themselves are also unstable because they are startups with less funding, so some startups may go out of business over time. Web3 AI service providers that serve Web3 clients may need to reacquire a portion of their customer base over time to replace those that go out of business, making it a more challenging business to scale.
In the long term, Hack VC is very bullish on Web3 native AI use cases, especially as AI agents become more prevalent. We envision a future where every Web3 user has a swarm of AI agents assisting them. An early category leader in this is Theoriq, which enables composable and autonomous on-chain AI agents.
Idea #4: Consumer GPU DePINs
There are many decentralized AI computing networks that rely on consumer-grade GPUs instead of data centers. Consumer-grade GPUs are suitable for low-end AI inference tasks or consumer use cases where latency, throughput, and reliability are flexible. But for professional enterprise use cases (which are most important markets), customers need a higher reliability network than their personal home machines, and if they have more complex inference tasks, they usually need higher-end GPUs. Data centers are better suited for these more valuable customer applications.
It’s important to note that we believe consumer-grade GPUs are fine for demonstration purposes or for individuals and startups that can tolerate lower reliability. But these customers are fundamentally lower value, so we believe DePINs that cater to Web2 enterprises will be more valuable in the long run. As a result, well-known GPU DePIN projects have generally evolved from early stages with primarily consumer-grade hardware to having A100/H100 and cluster-level availability.
Practical and Real-World Use Cases of Crypto x AI
Now, let’s discuss use cases that provide “real benefit,” where Crypto x AI can significantly add value.
Use Case #1: Serving Web2 Clients
McKinsey estimates that generative AI could add the equivalent of $2.6 trillion to $4.4 trillion in value per year across the 63 use cases they analyzed — for comparison, the UK’s total GDP was $3.1 trillion in 2021. This would increase the impact of all AI by 15% to 40%. This estimate would roughly double if we included the impact of embedding generative AI into software currently used for other tasks.
If you calculate the above estimates, this means that the global AI market (excluding generative AI) is likely worth trillions of dollars. In comparison, the total value of all cryptocurrencies, including Bitcoin and all altcoins, is currently only about $2.7 trillion. So let's be realistic: the vast majority of customers who need AI in the short term are Web2 customers, as Web3 customers who actually need AI are only a small fraction of this 2.7 trillion (considering that BTC accounts for half of this market, and BTC itself does not need/use AI).
The use cases for Web3 AI are just getting started, and it is not clear how big this market will be. But one thing is intuitively certain - it will be a small part of the Web2 market for the foreseeable future. We believe that Web3 AI still has a bright future, but it just means that the most powerful application of Web3 AI right now is serving Web2 customers.
Some Web2 clients that could benefit from Web3 AI include:
Vertical software companies built from the ground up with AI in mind (e.g. Cedar.ai or Observe.ai)
Large enterprises that fine-tune models for their own purposes (e.g. Netflix)
Fast-growing AI providers (e.g. Anthropic)
Software companies that add AI to existing products (e.g. Canva)
This is a relatively stable customer base because these customers are generally large and high value. They are unlikely to go out of business anytime soon and represent a large potential customer base for AI services. Web3 AI services that serve Web2 customers will benefit from this stable customer base.
But why would a Web2 client want to use the Web3 stack? The rest of this article will explain this reasoning.
Use Case #2: Reducing GPU Costs with GPU DePIN
GPU DePINs aggregates underutilized GPU computing power (most of which comes from data centers) and uses it for AI reasoning (e.g. io.net). A simple way of thinking is to think of it as "Airbnb for GPUs" (effectively collaborative consumption of underutilized assets).
The reason we are excited about GPU DePINs is as mentioned above, there is a current shortage of NVIDIA chips and there are currently wasted GPU cycles that could be used for AI inference. These hardware owners already have sunk costs and are not fully utilizing their devices today, so these partial GPU cycles can be provided at a cost far lower than the status quo because it is effectively a "windfall" for the hardware owner.
For example:
AWS machines: If you rent an H100 from AWS today, you need to commit to a year’s lease because of limited market supply. This is wasteful because you are unlikely to use your GPU 7 days a week, 365 days a year.
Filecoin Mining Hardware: Filecoin is a network with a large subsidized supply but no significant real demand. Unfortunately, Filecoin never found true product-market fit, so Filecoin miners risked going out of business. These machines have GPUs on them that can be repurposed for low-end AI inference tasks.
ETH Mining Hardware: When ETH transitioned from PoW to PoS, it immediately freed up a large amount of hardware that could be repurposed for AI inference.
It is important to note that not all GPU hardware is suitable for AI inference. One obvious reason is that older GPUs do not have enough GPU memory to handle LLMs, although there are some interesting innovations in this regard. For example, Exabits has a technology that loads active neurons into GPU memory and inactive neurons into CPU memory. They predict which neurons need to be active/inactive. This allows even low-end GPUs with limited GPU memory to handle AI workloads, effectively making low-end GPUs more useful in AI inference.
It is also important to note that Web3 AI DePINs will need to consolidate their services over time and provide enterprise-grade services such as single sign-on, SOC 2 compliance, service level agreements (SLAs), etc. This will be similar to the services enjoyed by Web2 customers in current cloud services.
Use case #3: Avoiding OpenAI self-censorship with uncensorable models
There has been a lot of discussion about censorship of AI. For example, Turkey temporarily banned OpenAI (later lifted the ban after OpenAI improved compliance). We believe that this type of censorship is inherently unhealthy, as countries need to embrace AI to remain competitive.
Even more interesting is OpenAI self-censoring. For example, OpenAI will not process NSFW content or predict the next presidential election. We think there is an interesting and large market for AI use cases that OpenAI doesn't want to touch for some reason.
Open source is a great solution because Github repositories are not beholden to shareholders or boards of directors. For example, Venice.ai promises to protect your privacy and operate in a censorship-resistant manner. The key is of course open source, which drives this. Web3 AI can effectively take these open source software (OSS) models to the next level, running these inferences on low-cost GPU clusters. This is why we think OSS+Web3 is the ideal combination to pioneer censorship-resistant AI.
Use Case #4: Avoid sending personally identifiable information to OpenAI
Many large enterprises are concerned about the privacy of their internal corporate data. For these customers, it is difficult to trust a centralized third party like OpenAI to handle their data.
Web3 may seem more worrisome for these businesses on the surface, as their internal data is suddenly available on a decentralized network. However, there are some innovations in the area of AI privacy-enhancing technologies:
Trusted Execution Environment (TEE), such as Super Protocol
Fully homomorphic encryption (FHE), such as Fhenix.io or Inco Network, and Bagel’s PPML
These technologies are still developing, and performance is improving with upcoming zero-knowledge (ZK) and FHE ASICs. But the long-term goal is to protect enterprise data while fine-tuning models. As these protocols emerge, Web3 may become a more attractive platform for privacy-preserving AI computing.
Use Case #5: Leverage the latest innovations in open source models
OSS (Open Source Software) has been eroding the market share of proprietary software for the past few decades. We view LLM as a sophisticated form of proprietary software that is ripe for disruption by OSS. Some notable challengers include Llama, RWKV, and Mistral.ai. This list will undoubtedly grow over time (a more comprehensive list can be found at Openrouter.ai). By leveraging Web3 AI (driven by the OSS model), it is possible to take advantage of these new innovations.
We believe that over time, a global open source development team combined with cryptographic incentives can drive rapid innovation in the open source model and the agents and frameworks built on top of it. An example is the AI agent protocol Theoriq. Theoriq leverages the OSS model to create a composable network of interconnected AI agents that can be assembled into more advanced AI solutions.
We are confident because of past experience: most "developer software" is slowly surpassed by OSS over time. There is a reason why Microsoft used to be a proprietary software company and now they are the company with the most contributions on Github. If you look at how Databricks, PostGresSQL, MongoDB, etc. have disrupted proprietary databases, this is an example of an entire industry being disrupted by OSS, so there is quite a lot of precedent here.
However, this does come with a catch. One of the sticky aspects of OSS LLMs is that OpenAI has begun signing paid data licensing agreements with organizations such as Reddit and the New York Times. If this trend continues, OSS LLMs may have trouble competing due to the financial barriers to accessing data. Nvidia may double down on its investment in confidential computing as an enabler of secure data sharing. Time will tell how this trend develops.
Use case #6: Consensus via high penalty random sampling or ZK proofs
One challenge with Web3 AI inference is verification. Validators have the opportunity to earn fees by spoofing results, so verifying inference is an important measure. It is important to note that this kind of spoofing has not actually happened yet because AI inference is still in its infancy, but it is inevitable unless measures are taken to prevent this behavior.
The standard Web3 approach is to have multiple validators repeat the same operation and compare the results. The obvious challenge is that, as mentioned above, AI inference is expensive due to the current shortage of high-end Nvidia chips. Given that Web3 can provide lower-cost inference through underutilized GPU DePINs, repeated computation will severely weaken the value proposition of Web3.
A more promising solution is to perform ZK proofs for off-chain AI inference computations. In this case, it is possible to determine whether a model was trained correctly or whether inference was run correctly by verifying a succinct ZK proof (called zkML). Examples include Modulus Labs and ZKonduit. The performance of these solutions is still in its infancy, as ZK operations are very computationally intensive. However, we expect this to improve with the release of ZK hardware ASICs.
What has more potential is an "optimistic" sampling-based approach to AI reasoning. In this model, you only verify a small fraction of the results generated by the validator, but set the penalty economic cost high enough so that if caught, it will create a strong economic disincentive for the validator to cheat, thus preventing them from cheating. In this way, redundant computation can be saved (see Hyperbolic's Proof of Sampling paper for example).
Another promising idea is watermarking and fingerprinting solutions, such as the one proposed by Bagel Network. This is similar to the on-device AI model quality assurance mechanism that Amazon Alexa uses for its millions of devices.
Use Case #7: Saving Money with OSS (OpenAI’s Profit Margin)
The next opportunity Web3 brings to AI is cost democratization. So far, we’ve discussed saving GPU costs through DePINs. But Web3 also offers opportunities to save on profit margins of centralized Web2 AI services such as OpenAI, which currently has over $1 billion in revenue per year. These cost savings come from using open source models instead of proprietary models because the model creators are not trying to profit from them.
Many OSS models will remain completely free, which provides the best economic benefit to customers. However, some OSS models may also try these monetization methods. Consider that of all the models on Hugging Face, only 4% are trained by companies with budgets to help subsidize these models (see here). The remaining 96% of models are trained by the community. These models — 96% of Hugging Face — have underlying real costs (both compute costs and data costs). So these models need some way to be monetized.
There are many proposals to monetize these OSS models. One of the most interesting is the concept of an “Initial Model Offering” (IMO), where the model itself is tokenized, a portion of the tokens are reserved for the team, and a portion of future revenue from the model is distributed to token holders, although there are obviously some legal and regulatory hurdles here.
Other OSS models will attempt to monetize based on usage. Note that if this is achieved, OSS models may become more and more like their Web2 for-profit counterparts. But in reality, the market will bifurcate and some models will remain completely free.
Use Case #8: Decentralized Data Provenance
One of the biggest challenges in AI is getting the right data to train the models. We mentioned earlier that decentralized AI training is challenging. What about using the decentralized web to get data (and then train it somewhere else, even on a traditional Web2 platform)?
This is exactly what startups like Grass are doing. Grass is a decentralized “data scraping” network where individuals contribute their machines’ idle processing power to acquire data in order to inform the training of AI models. Assuming at scale, this source of data can outperform any one company’s internal efforts to acquire data because of the power of a large network of incentivized nodes. This includes not only acquiring more data, but also acquiring it more frequently, making it more relevant and up-to-date. Because data scrapers are decentralized and don’t reside on a single IP address, it’s nearly impossible to stop this decentralized army of scrapers. They also have a network of humans who clean and normalize the data to make it useful after it’s been scraped.
Once you have data, you also need a place to store it on-chain, as well as the LLMs generated with that data. 0g.AI is an early leader in this category. It is an AI-optimized, high-performance Web3 storage solution that is much cheaper than AWS (another Web3 AI economic advantage), while also providing data availability infrastructure for Layer 2s, AI, etc.
It is important to note that the role of data in Web3 AI may change in the future. Today, the status quo of LLMs is to pre-train models with data and refine them with more data over time. However, these models are always slightly out of date because data on the Internet changes in real time. Therefore, the response of LLM reasoning is slightly inaccurate.
A future direction may be a new paradigm - "real-time" data. The concept is that when the LLM is asked reasoning questions, the LLM can use data collected in real time from the Internet for prompt injection. In this way, the LLM can use the most recent data, which is also the direction that Grass is researching.
in conclusion
In summary, while the potential of Web3 AI is huge, current challenges and market demands indicate that the most practical and realistic application is to serve Web2 clients. The future of Web3 AI is bright, but its full application and market expansion will still take time. By identifying and addressing these challenges, we can better lay the foundation for the development and success of Web3 AI.
