The compute and training layers of the AI-industrial-complex are currently dominated by centralized Web2 giants that have unparalleled access to capital, state-of-the-art hardware, and vast datasets. While this will likely remain the case for the most powerful general ML models, mid-tier and bespoke models may increasingly source their compute resources from more affordable and accessible Web3 networks. Similarly, for inference needs that exceed the capabilities of personal edge devices, some consumers may turn to Web3 networks for less censored, more diverse outputs. Rather than attempting to overhaul the entire AI stack, Web3 challengers should focus on these niche use cases, and lean more heavily into their unique value proposition around censorship-resistance, transparency, and social verifiability.
The hardware required for training the next generation of foundation models is a scarce and expensive resource, and demand for the most performant chips will continue to outpace supply. This scarcity acts as a centralizing force, as top-line hardware concentrates among a handful of well-resourced incumbents who can leverage it to train and commercialize the most performant and complex foundation models.
But hardware ages quickly, and what happens to outdated, less performant, or mid-tier hardware? It will likely get matched with less sophisticated or more niche models. Allocative efficiency is found in pairing different classes of models with different types of hardware, and this is where a Web3 protocol that coordinates access to a diverse array of low-cost compute resources could play an interesting role. In this scenario, consumers would interact with simple mid-tier models trained on personal datasets and hosted on their edge devices, and resort to high-end models trained and hosted by centralized incumbents only for more demanding tasks, ideally with their user identity obfuscated and prompt data encrypted.
Beyond efficiency, there are growing concerns around bias and potential censorship within centralized models. Given their transparent and verifiable nature, Web3 environments could play a role in training models that are ignored or deemed too controversial by Web2, but are still viewed as valuable by other parts of society. Despite being less competitive in terms of performance and innovation, Web3 protocols could thus carve out a niche by offering model training that is more open, trustworthy, and less susceptible to censorship. At first, the two approaches can coexist side-by-side, each serving a different set of use cases. Over time, however, as Web3 improves its developer experience and platform compatibility, and the network effects of open source AI kick into full gear, it may eventually be positioned to compete on the incumbents’ home turf, especially as consumers become increasingly aware of the downsides of centralized models.
In addition to mid-tier or niche model training, Web3 challengers are well-positioned to provide more transparent and flexible inference solutions. Decentralized inference providers could offer advantages such as zero downtime, seamless model composability, public model evaluation, and more diverse, censorship-free outputs. They also circumvent the vendor lock-in that consumers experience when forced to choose between a select few centralized providers. Similar to training, the competitive differentiator between decentralized inferencing layers and their centralized counterparts isn’t raw compute power, but solving issues related to the closed-source nature of fine-tuned parameters, a general lack of verifiability, and affordability.
Dan Olshansky lays out one potential vision for this with POKT’s router network for AI inference, which could open up new opportunities for AI researchers and engineers to apply their work and generate additional revenue from their bespoke ML/AI models. Importantly, such a network would encourage healthier competition between inference providers by sourcing outputs from a wide variety of sources, both independent/decentralized and large/centralized.
Optimistic outlooks predict that the entire AI stack will move on-chain. While not impossible in the long run, I believe that access to data and compute resources are strong centralizing forces that currently give incumbents a decisive competitive advantage. However, decentralized coordination and compute networks present a unique value proposition around more personalized, affordable, openly competitive, and censorship-resistant AI. By focusing on market niches where these values matter the most, Web3 can develop its own competitive moats and thus help ensure that the most impactful technology of our era evolves in many different directions and empowers a maximally broad range of stakeholders, as opposed to being dominated by a small handful of the usual suspects.
I’d like to thank the entire Placeholder Investment team, Kyle Samani from Multicoin Capital, Anand Iyer from Canonical VC, Keccak Wong from Nectar AI, Alpin Yukseloglu from Osmosis Labs, and Cameron Dennis from the NEAR Foundation for their review and valuable feedback.