Hack VC Partner: The 8 Real Advantages of AI + Crypto
TechFlow Selected TechFlow Selected
Hack VC Partner: The 8 Real Advantages of AI + Crypto
Analyzing the convergence of crypto and AI, discussing the real challenges and opportunities—what are the hollow promises, and what is truly feasible?
Navigating Web3 tides with focused insightsAuthor: Ed Roman, Managing Partner at Hack VC
Translation: 1912212.eth, Foresight News
AI + Crypto is one of the most talked-about frontiers in the crypto market recently, including areas like decentralized AI training, GPU DePINs, and censorship-resistant AI models.
Behind these dazzling developments, we must ask: is this real technological breakthrough or just hype-chasing? This article will cut through the noise, analyze the vision and challenges of crypto x AI, and reveal what promises are hollow and which ones are truly viable?
Vision #1: Decentralized AI Training
The problem with on-chain AI training lies in the need for high-speed communication and coordination between GPUs, as neural networks require backpropagation during training. Nvidia has developed two innovations for this (NVLink and InfiniBand). These technologies enable ultra-fast GPU communication but are limited to on-premise setups, suitable only for GPU clusters within a single data center (50+ gigabit speeds).
Introducing a decentralized network drastically slows things down by several orders of magnitude due to network latency and bandwidth constraints. Compared to the throughput achievable via Nvidia's high-speed interconnects within a data center, such speeds are simply impractical for AI training use cases.
Note that some innovations may offer hope for the future:
-
Large-scale distributed training over InfiniBand is emerging, as NVIDIA itself now supports distributed non-local training over InfiniBand via the NVIDIA Collective Communications Library. However, it remains in its infancy, so adoption metrics remain unclear. Physical distance limitations still exist, making local InfiniBand training significantly faster.
-
New research on decentralized training has been published, reducing communication synchronization time, potentially making decentralized training more practical in the future.
-
Smart model sharding and scheduling can enhance performance. Likewise, new model architectures may be specifically designed for future distributed infrastructure (Gensyn is researching in these areas).
The data aspect of training also poses challenges. Any AI training process involves handling massive amounts of data. Typically, models are trained on centralized, secure data storage systems optimized for scalability and performance. This requires transferring and processing multiple terabytes of data—not a one-time cycle. Data is often noisy and contains errors, requiring cleaning and conversion into usable formats before training. This stage involves repetitive tasks like normalization, filtering, and handling missing values—all of which face severe hurdles in decentralized environments.
Moreover, the data phase is iterative, which doesn't align well with Web3. OpenAI went through thousands of iterations to achieve its results. In AI teams, the core workflow for data scientists includes defining objectives, preparing data, analyzing and organizing it for insights, and shaping it for modeling. Then, machine learning models are developed to solve defined problems and validated using test datasets. This process is inherently iterative: if a model underperforms, experts return to earlier stages to refine results. Imagine trying to execute this process in a decentralized environment—existing state-of-the-art frameworks and tools are not easily adaptable to Web3.
Another issue with on-chain AI model training is that the market is far less compelling compared to inference. Currently, training large language models (LLMs) demands enormous GPU compute resources. Long-term, however, inference will become the dominant application for GPUs. Just consider how many LLMs would need to be trained globally versus how many customers actually use them—which number is larger?
Vision #2: Achieving Consensus via Redundant AI Inference Computation
Another challenge at the intersection of crypto and AI is verifying the accuracy of AI inference, since you cannot fully trust a single centralized party to perform inference operations—nodes may behave maliciously. This challenge does not exist in Web2 AI, as there is no decentralized consensus system.
The solution is redundant computation—requiring multiple nodes to repeat the same AI inference operation, enabling execution in a trustless environment and avoiding single points of failure.
However, the problem with this approach is the extreme shortage of high-end AI chips. Wait times for premium NVIDIA chips stretch into years, driving up prices. Requiring AI inference to be re-executed across multiple nodes multiplies already high costs, rendering this approach unfeasible for many projects.
Vision #3: Near-Term Web3-Specific AI Use Cases
Some suggest Web3 should have unique, native AI use cases tailored specifically for Web3 users. Examples include Web3 protocols using AI for DeFi pool risk scoring, Web3 wallets recommending new protocols based on wallet history, or Web3 games using AI to control non-player characters (NPCs).
Currently, this is an early-stage market (in the short term), where use cases are still being explored. Some challenges include:
-
Due to nascent demand, Web3-native use cases generate relatively few potential AI transactions.
-
Fewer customers—Web3 users are orders of magnitude smaller than Web2 users, resulting in lower decentralization.
-
Customers themselves are unstable—often startups with limited funding, some of which may eventually fail. Web3 AI service providers serving Web3 clients may need to constantly acquire new clients to replace those lost, making business scaling extremely challenging.
In the long run, we are very bullish on Web3-native AI use cases, especially as AI agents become more prevalent. We envision a future where every Web3 user could have numerous AI agents assisting them with various tasks.
Vision #4: Consumer-Grade GPU DePIN
Many decentralized AI compute networks rely on consumer-grade GPUs rather than data centers. Consumer GPUs are well-suited for low-end AI inference tasks or consumer use cases where latency, throughput, and reliability are flexible. But for serious enterprise applications—the majority of valuable markets—customers demand higher reliability than home machines provide, and for more complex inference tasks, they typically require higher-end GPUs. Data centers are better suited for these higher-value customer needs.
That said, we believe consumer-grade GPUs are suitable for demos and for individuals or startups tolerant of lower reliability. However, these customers are lower value, so we believe DePINs tailored for Web2 enterprises will be more valuable long-term. Hence, GPU DePIN projects have evolved from early reliance on consumer hardware toward supporting A100/H100 and cluster-level availability.
Reality—Practical Use Cases for Crypto x AI
Now let’s discuss use cases that deliver real benefits—these are the true wins where crypto x AI adds clear value.
Real Benefit #1: Serving Web2 Customers
McKinsey estimates that generative AI could add $2.6 trillion to $4.4 trillion annually across 63 use cases analyzed—compared to the UK’s 2021 GDP of $3.1 trillion. This would increase AI’s impact by 15% to 40%. If we factor in the effect of embedding generative AI into other tasks beyond current software use, the estimated impact roughly doubles.
Based on these estimates, the total global AI market (beyond just generative AI) could reach tens of trillions of dollars. In contrast, the entire cryptocurrency market (including Bitcoin and all altcoins) today is worth around $2.7 trillion. Let’s be realistic: the vast majority of AI customers in the near term will be Web2 users, because Web3 customers needing AI will constitute only a tiny fraction of that $2.7 trillion (considering BTC itself neither needs nor uses AI).
Web3 AI use cases are just beginning, and their ultimate scale remains uncertain. But one thing is clear—they will represent only a small fraction of the Web2 market in the foreseeable future. We believe Web3 AI still has a bright future, but this means its strongest current application is serving Web2 customers.
Examples of Web2 customers who could benefit from Web3 AI include:
-
Vertical-specific software companies built around AI from the ground up (e.g., Cedar.ai or Observe.ai)
-
Large enterprises fine-tuning models for internal use (e.g., Netflix)
-
Rapidly growing AI providers (e.g., Anthropic)
-
Software companies integrating AI into existing products (e.g., Canva)
These represent relatively stable customer profiles—typically large, well-funded, and valuable. They are unlikely to disappear quickly and represent a huge potential client base for AI services. Web3 AI services targeting Web2 customers benefit from this stability.
But why would Web2 customers want to use the Web3 stack? The next sections explain exactly that.
Real Benefit #2: Lowering GPU Usage Costs via GPU DePIN
GPU DePIN aggregates underutilized GPU computing power—most reliably from data centers—and makes it available for AI inference. A simple analogy is "Airbnb for GPUs."
We're excited about GPU DePIN because, as noted, NVIDIA chip shortages persist, and there are wasted GPU cycles available for AI inference. Hardware owners have already incurred sunk costs and aren’t fully utilizing their devices, so they can offer partial GPU capacity at much lower cost than current alternatives—effectively “finding money” from idle hardware.
Examples include:
-
AWS machines. Renting an H100 from AWS today requires a 1-year commitment due to limited supply. This leads to waste—you likely don’t need the GPU running 24/7/365.
-
Filecoin mining hardware. Filecoin has abundant subsidized supply but lacks real demand. It never achieved product-market fit, so miners face obsolescence. These GPU-equipped machines can be repurposed for low-end AI inference.
-
ETH mining hardware. When Ethereum transitioned from PoW to PoS, it rapidly freed up a large amount of hardware that could be reused for AI inference.
Note that not all GPU hardware is suitable for AI inference. An obvious reason is that older GPUs lack sufficient memory for LLMs, though some interesting innovations help here. For example, Exabits’ technology loads active neurons into GPU memory and inactive ones into CPU memory, predicting which neurons are needed. This enables low-end GPUs to handle AI workloads despite limited VRAM, making them more useful for AI inference.
Over time, Web3 AI DePINs must evolve their products to offer enterprise-grade services such as single sign-on, SOC 2 compliance, service-level agreements (SLAs), etc.—similar to what cloud providers currently offer Web2 customers.
Real Benefit #3: Censorship-Resistant Models to Avoid OpenAI Self-Censorship
There’s much discussion about AI censorship. For instance, Turkey temporarily banned OpenAI (later reversed after OpenAI improved compliance). We find national-level censorship uninteresting because countries must adopt AI to stay competitive.
OpenAI also engages in self-censorship. For example, OpenAI refuses to process NSFW content or predict the next presidential election. We believe these AI use cases are not only interesting but represent massive markets that OpenAI avoids for political reasons.
Open source is a strong alternative, as GitHub repositories aren’t influenced by shareholders or boards. Venice.ai, for example, commits to privacy and operates in a censorship-resistant manner. Web3 AI can amplify this by hosting these open-source (OSS) models on low-cost GPU clusters for inference. For these reasons, we believe OSS + Web3 is the ideal combination to enable censorship-resistant AI.
Real Benefit #4: Avoid Sending Personally Identifiable Information to OpenAI
Large enterprises have privacy concerns regarding internal data. For these clients, trusting a third party like OpenAI with sensitive data is difficult.
At first glance, storing internal data on a decentralized network in Web3 might seem even more concerning. However, innovations in privacy-enhancing technologies for AI are emerging:
Trusted Execution Environments (TEE), e.g., Super Protocol
Fully Homomorphic Encryption (FHE), e.g., Fhenix.io (a portfolio company of Hack VC’s managed fund) or Inco Network (both powered by Zama.ai), and Bagel’s PPML
These technologies are still evolving, with performance improving thanks to upcoming zero-knowledge (ZK) and FHE ASICs. The long-term goal is protecting enterprise data during model fine-tuning. As these protocols mature, Web3 could become a more attractive venue for privacy-preserving AI computation.
Real Benefit #5: Leveraging Latest Innovations from Open-Source Models
Over the past decades, open-source software has steadily eaten into proprietary software’s market share. We see LLMs as a form of proprietary software vulnerable to disruption by OSS. Notable challengers include Llama, RWKV, and Mistral.ai. This list will undoubtedly grow over time (a more comprehensive list is available on Openrouter.ai). By leveraging Web3 AI powered by OSS models, innovators can build on these advancements.
We’re confident that, over time, the combination of a global open-source developer community with crypto incentives will drive rapid innovation in open-source models and the agents and frameworks built atop them. An example is Theoriq, which uses OSS models to create a composable network of AI agents that can be combined into advanced AI solutions.
Our confidence stems from history: over time, most “developer software” innovation has been overtaken by OSS. Microsoft, once a proprietary software giant, is now the top contributor on GitHub. Consider how Databricks, PostgreSQL, MongoDB, and others disrupted proprietary databases—this precedent is highly persuasive.
However, a challenge exists. A key issue with open-source LLMs is that OpenAI has begun signing paid data licensing deals with organizations like Reddit and The New York Times. If this trend continues, open-source LLMs may struggle to compete due to financial barriers in data acquisition. Nvidia may further invest in confidential computing to aid secure data sharing. Time will tell how this unfolds.
Real Benefit #6: Achieving Consensus via High-Penalty Random Sampling or ZK Proofs
One challenge in Web3 AI inference is verification. If validators can cheat to earn fees, verifying inference becomes critical. Note that such cheating hasn’t occurred yet—AI inference is still early—but unless deterred, it’s inevitable.
The standard Web3 approach is to have multiple validators repeat the same operation and compare results. As previously discussed, the major challenge is the high cost of AI inference due to the scarcity of high-end Nvidia chips. Given that Web3 can offer lower-cost inference via underutilized GPU DePINs, redundant computation severely undermines Web3’s value proposition.
A more promising solution is executing ZK proofs for off-chain AI inference computations. In this case, a succinct ZK proof can verify whether a model was correctly trained or inference was properly executed (known as zkML). Examples include Modulus Labs and ZKonduit. Since ZK operations are computationally intensive, these solutions are still in early stages. However, we expect improvements as ZK hardware ASICs launch in the near future.
Even more promising is a somewhat “Optimistic” sampling-based approach for AI inference. In this model, only a small fraction of validator-generated results are verified, but slashing penalties are set high enough to strongly economically disincentivize cheating. This way, redundant computation can be avoided.
Another promising idea is watermarking and fingerprinting solutions, such as those proposed by Bagel Network—similar to how Amazon Alexa ensures on-device AI model quality across millions of devices.
Real Benefit #7: Cost Savings via OSS (Capturing OpenAI’s Margins)
The next opportunity Web3 brings to AI is cost democratization. So far, we’ve discussed saving GPU costs via DePIN. But Web3 also offers savings on the profit margins of centralized Web2 AI services (e.g., OpenAI, which generates over $1B in annual revenue as of writing). These savings come from using OSS models instead of proprietary ones, as model creators aren’t focused on profitability.
Many OSS models will remain completely free, delivering optimal economics for users. However, some OSS models may experiment with monetization. Consider that only 4% of models on Hugging Face are trained by budgeted companies. The remaining 96% are trained by the community. This group (96% of Hugging Face) incurs basic real costs (compute and data). Thus, these models will need to be monetized somehow.
Several proposals exist for monetizing OSS models. One of the most interesting is the concept of an “Initial Model Offering” (IMO)—tokenizing the model itself, reserving tokens for the team, and directing future model revenues to token holders, although legal and regulatory hurdles certainly exist.
Other OSS models may try usage-based monetization. If this happens, OSS models may begin resembling Web2 monetization models. But in practice, the market will likely bifurcate, with some models remaining entirely free.
Real Benefit #8: Decentralized Data Sources
One of AI’s biggest challenges is finding the right data to train models. We’ve previously discussed the difficulties of decentralized AI training. But what about using decentralized networks to gather data—data that can then be used elsewhere, even in traditional Web2 settings?
This is exactly what startups like Grass are doing. Grass is a decentralized network of “data scrapers” who contribute their machine’s idle processing power to source data for training AI models. At scale, due to the strength of a large incentivized node network, such data sources could outperform any single company’s internal efforts—not just in volume, but in frequency, ensuring data is more relevant and up-to-date. Moreover, a decentralized army of data scrapers cannot be easily stopped, as they are inherently decentralized and not confined to a single IP address. They also operate a network to clean and standardize data post-scraping, making it immediately useful.
After acquiring data, you also need locations to store it on-chain and use it to generate LLMs.
Note that the role of data in Web3 AI may evolve. Today, LLMs are pre-trained on static data and refined over time with additional datasets. However, since internet data changes in real time, these models are always slightly outdated, leading to slightly inaccurate responses.
Future progress may follow a new paradigm—“real-time” data. The idea is that when an LLM receives an inference query, it can fetch and inject fresh data collected in real time from the web. This allows the LLM to respond with the latest information. Grass is actively exploring this area.
Special thanks to the following individuals for their feedback and assistance with this article: Albert Castellana, Jasper Zhang, Vassilis Tziokas, Bidhan Roy, Rezo, Vincent Weisser, Shashank Yadav, Ali Husain, Nukri Basharuli, Emad Mostaque, David Minarsch, Tommy Shaughnessy, Michael Heinrich, Keccak Wong, Marc Weinstein, Phillip Bonello, Jeff Amico, Ejaaz Ahamadeen, Evan Feng, JW Wang.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@hack_vc
