Blockchain investment firm Pantera Capital is backing a new generation of networks designed to unlock computing power, data, and energy.
AI may feel intangible: lines of code creating content and making decisions in the cloud, seemingly out of nowhere. But behind every prompt sits an increasingly strained physical infrastructure. As AI systems become more agentic (operating continuously, making decisions, and executing tasks on behalf of users) the demands on compute, energy, and storage grow even more intense. Much of this infrastructure is controlled by a handful of large technology platforms. As AI adoption accelerates, concerns about cost, access, and control are intensifying, raising questions about who gets to build and operate the most powerful systems and on what terms.
That pressure is forcing a reevaluation. A growing number of companies and developers are experimenting with alternatives to centralization, tapping spare computing capacity, aggregating data from across the internet, and drawing on large pools of independent contributors. This shift is particularly relevant as agentic AI systems require persistent, scalable, and verifiable resources to operate reliably. Some of the most ambitious efforts are coming from the world of blockchain, where token-based incentives can coordinate decentralized infrastructure, enabling networks where autonomous AI agents can access compute, transact, and even compensate contributors without relying on a central intermediary.
For investors, attention is turning to how this more distributed layer of AI infrastructure is built and how access to it is governed. Fortune Brand Studio spoke with Cosmo Jiang, board director at Solana Company and general partner at Pantera Capital, an investment firm specializing in blockchain and digital assets, about why data, computing power, and energy are key constraints in AI development and how new blockchain networks are attempting to unlock and coordinate these resources in fundamentally different ways.
Why is AI infrastructure such an important area of focus for Pantera?
A lot of investors still talk about AI mainly through the lens of models and applications, but the real bottlenecks are underneath that layer. If you really boil it down, training and running AI models requires three main resources: data, compute, and energy. Each is finite and becoming more valuable as demand rises. That makes infrastructure absolutely vital.
One of the superpowers of blockchain is the use of token incentives to coordinate a disparate group of people toward a single goal. In this context, blockchain can be used to open capacity that traditional centralized systems leave underused by coordinating large numbers of independent resource providers. The industry calls these blockchain protocols decentralized physical infrastructure networks, or DePIN.
We’re spending a growing share of our research time and capital on this because we see it as a major area of innovation in the years ahead.
Why are data, compute, and energy becoming such critical constraints in AI?
Humans have an innate demand for progress, and so, unsurprisingly, the demand for intelligence is unbounded. If we can create AI, then the limiting factor is more so the supply of inputs than the demand for output. When I think about what the inputs are, the base ingredients that produce AI inference, it boils down to data, compute, and energy. Without data, you can’t train or improve models. Without compute, you can’t process the workloads required to train or run them. And without energy, none of this happens at scale. These inputs are not infinite, and as more companies build and deploy AI products, the competition for them intensifies.
That creates cost pressure, access issues, and a concentration problem, because a relatively small number of large platforms control a great deal of the underlying systems. That dynamic creates an opening for new frameworks that can source and manage AI resources creatively.
What is it about blockchain that makes it useful in solving those kinds of infrastructure problems?
For me, the real innovation of blockchain is using tokens as coordination technology. It’s a way to align incentives across thousands of globally distributed contributors without any single company owning or managing the system. That matters enormously for AI. Today, the data, compute, and energy that power AI are controlled by a handful of tech giants. Blockchain offers a different model in which there’s no single point of control. Just shared incentives, at global scale.
Blockchains give you a way to organize contributions from many different participants in an open and transparent system, without requiring a single central operator to own or manage everything directly. In the context of AI, that matters because much of the capacity needed is already out in the world.
Blockchain-based systems can create economic incentives for individuals or organizations anywhere in the world to provide that capacity while also tracking and verifying that useful work has been done. That lowers the barriers for contribution, which, in aggregate, can be a meaningful resource. Blockchain also supports granular, pay-as-you-go usage where developers and AI providers pay small amounts for compute, data, and bandwidth as they use it, with those payments going directly to the entities providing it. The x402 standard, developed by Coinbase and Cloudflare and now maintained by the Linux Foundation, is beginning to enable this kind of usage by allowing services to be paid for on a per-request basis.
How does that idea translate in practice?
One of the clearest examples is Grass. You can think of Grass as a data aggregation network serving large AI labs. Today’s leading models need multimodal and live internet data, but collecting that data at scale through traditional web-scraping infrastructure is difficult even when it is public internet data. Large platforms are very good at detecting single-source scraping operations and shutting them down.
Grass, which has received investment from Pantera, approaches the problem differently by drawing on a large base of personal devices around the world. Individuals can install software that contributes bandwidth and data collection capacity in the background, with blockchain-based reward systems used to manage and compensate those contributions. These individuals are otherwise customers of large internet platforms and, individually, don’t take up a lot of bandwidth, changing the structure of the network and impractical to prevent. Grass protocol’s ability to aggregate fresh public internet data is a particularly attractive resource that AI companies rely on to train their models.
Hivemapper is another example of a protocol aggregating images from a decentralized set of dashcams, including on fleets of delivery or rideshare drivers, and using that to build a real-time mapping database. This data is then useful for use in training or real-time inference for self-driving vehicles.
And what about compute? Where are you seeing the same model play out there?
AI demand is rising quickly, but access to high-performance compute is still relatively constrained and expensive. Blockchain networks can help create open marketplaces for that capacity. This model is compelling because it can make AI infrastructure more flexible and potentially more capital efficient. Instead of everything flowing through a handful of hyperscalers, you start to create a market where compute can be discovered and allocated across a much broader base of providers. There are several protocols doing this, such as Akash and Bittensor, and they’ve been somewhat successful in aggregating supply. There is a technical challenge of coordinating decentralized training, which when figured out will unlock more demand.
Are these models being used in the real world?
It’s still early days, but we’re seeing signs of real activity. In Grass’s case, the network has built a large multimodal dataset, with demand from major AI companies stronger than expected and tens of millions of dollars already generated from paid data collection.
We’re also seeing wider evidence that blockchain-based micropayments is gaining traction. In the AI agent payments market, for example, activity reached roughly $1.6 million over a 30-day period after adjusting for inflated or duplicate transactions. Much of this is happening over emerging standards such as Coinbase’s x402 protocol, over high-throughput networks such as Solana and Base. All this suggests that things are beginning to move from concept toward real-world use.
Looking ahead, what would success look like for this new model of AI infrastructure, and what should investors focus on?
Success would mean these networks are no longer seen as experimental, but as reliable, widely used infrastructure. In practice, it would mean developers gaining easier access to compute, AI companies working with more diverse and up-to-date datasets, and individuals able to contribute resources and be compensated for them.
For investors, the key question is whether blockchains are actually consistently providing either a better service, for example Grass’ data that is otherwise unavailable or a similar service at a cheaper cost or unlocking compute that would otherwise go unused. The most promising networks will be those that attract sustained participation, demonstrate real demand, and become embedded in the AI economy.
References to any companies contained herein are provided for illustrative purposes only and do not constitute an investment recommendation.