Key takeaways
- Compute hardware demand is growing at approximately 30% annually, with projections reaching $420 billion by 2030.
- Centralized providers like AWS and CoreWeave serve medium to large businesses, while DePIN projects better serve small to medium-sized businesses with specific use cases.
- Each new generation of large language models requires significantly more computational power than previous versions.
- AI-Fi refers to infrastructure enabling investors to gain exposure to AI growth through underlying compute resources and investments.
- DePIN networks using consumer-grade hardware may be more economically viable for developing markets than first-world economies.
Why compute demand keeps climbing
Ashton Addison: How fast is demand for compute growing, and what is driving it?▶ 00:45
Nick: The need for compute is growing exponentially. Based on several sources, aggregate growth in compute needs is about 30% a year. By 2030, only five years away, demand for compute hardware could reach about 420 billion. You might think there's a limited number of use cases and we'll hit the curve, but it doesn't work that way. Compute demand only grows more and more, because innovation and engineering require more and more hardware. xAI is building out data center infrastructure, and AWS and CoreWeave are scaling their GPU data centers. We're even getting requests from gigantic companies asking us to help them get hundreds of millions of dollars' worth of hardware. We don't talk about it as much as we did six months or a year ago, but it's still growing.
Ashton: Where does all that compute actually come from?▶ 02:33
Nick: There's the centralized side and the decentralized side. On the centralized side you have AWS, Google Cloud and CoreWeave, plus others like Lambda. Most of them own their hardware and provide it to customers. Others, like RunPod, aggregate from different sources, similar to a DePIN but Web2-based. Then the Web3 DePIN projects aggregate from as many sources as they can, sometimes even consumer-grade hardware from retail people like you and me. What matters is which provider has the specific hardware you need. For a really heavy training model, even AWS or GCP may not have everything. A DePIN pulling from varied sources may not sustain the speed and efficiency you need. So customers building heavy AI products often end up with custom-built servers. They go to an operator with a very specific set of requirements, and the operator builds it.
Compute demand only grows more and more, because innovation and engineering require more and more hardware.
Where DePIN fits
Ashton: Could DePIN networks, including phones and home computers, grow dramatically in 2025 as the compute problem becomes more prominent?▶ 05:38
Nick: I think of different compute providers as serving different customer profiles. DePINs are excellent for small to medium-sized businesses, while the centralized players serve medium to large ones. The difference is where they pull their resources from. Each DePIN solves the use cases it was built for. Hyperbolic, for example, is really good at AI inferencing and training. As for aggregated resources from phones or laptops, I don't know exactly what problems they solve. I'd gather they bring economies of scale and make certain resources cheaper. You can already get consumer-grade RTX 4090s pretty affordably by the hour. My question is how much cheaper it needs to get. Maybe it isn't built for first-world economies like the US or Canada. It might be more effective in Latin America or the sub-Saharan African market.
How Compute Labs turns GPUs into investments
Ashton: Compute Labs talks about "AI-Fi." What does the term mean?▶ 09:33
Nick: We started using the term AI-Fi back in March. I won't lie, a little of it is just marketing. What we're really doing is building infrastructure that lets investors get into the growth of AI and AI investments. It's also about making sure people know that by investing in compute resources, they're enabling the growth of the AI industry. It's a commodity approach: invest in the core asset that builds, powers and enables AI.
Ashton: How does the business work, from the money coming in to the hardware going out?▶ 11:28
Nick: We raise capital through institutional, traditional finance and also through retail investment, from the markets where we can take capital. We then buy compute hardware, mostly enterprise-grade, like the B200. We also dabble in consumer-grade products like 4090s. Then we lease that hardware to operators. You can call them data centers, data center operators or cloud service providers. We lease it to them so they can keep scaling and growing their businesses. Most of the time they already have a customer or demand lined up, which is why they need more hardware to serve those customers.
Ashton: Where does blockchain come in? Do you have a token?▶ 16:12
Nick: Yes. We are tokenizing our hardware. We raise capital from institutional and retail investors by issuing a structured product, an investable asset. It gives them the rights to the hardware's revenue, profits and liquidation proceeds, and that's what the token means. If you invest in an H200 product, you get a GPU NFT of an H200, a one-to-one representation of the hardware. You get all of its revenue, and when we sell or liquidate it, you get those proceeds too. That's what blockchain is doing here: tokenizing the actual hardware.
It's a commodity approach: invest in the core asset that builds, powers and enables AI.
Bottlenecks and what's next
Ashton: Is the problem a shortage of supply, or inefficiencies in getting compute hardware running?▶ 12:48
Nick: The short answer is that there are inefficiencies in lead time. The provider that can deliver most efficiently or fastest is often the one that wins the contract. When you order hardware for a custom-built cluster, it takes about six weeks to arrive. Then you need about four weeks to run it, test it and make it operational. That's a lot of time for an AI company that needs its thing done tomorrow. Another inefficiency is energy. A lot of data centers simply don't have the power to keep scaling. Sometimes it's legislative, and governments may prevent a data center from using more resources. Other times there just aren't enough data centers in certain areas to put hardware in.
Ashton: What's on the roadmap for Compute Labs in 2025?▶ 18:22
Nick: Our priority is helping operators get more resources, whether that's the capital to buy hardware or the hardware itself. We've set up a pilot with the B200 with our partner in Japan, going live in late February or early March. We're excited about that asset because access to the B200 is very limited. Compute Labs is one of eight of the largest suppliers in the world for it, because that hardware won't be available to the mass market until around Q2 or Q3 2025. We also have one operator lined up that we'll soon be fundraising for, and we'll need a ton of capital to serve that customer. We're also looking at small clusters in the range of 1, 3 or 5 million, and we're happy to help those operators too.
This interview has been edited for length and clarity. Recorded February 2025.

