The race for AI is no longer limited just by chip availability. It's running headfirst into a massive power grid bottleneck while hyperscalers rush to secure gigawatts or electricity. Grid constraints, as well as permitting timelines, are also creating an unyielding ceiling on AI growth and a vast amount of the world's existing GPU capacity sits underutilized, wasting both energy as well as capital.
Joining us live this morning to discuss the AI energy crisis, as well as how decentralized compute marketplaces can unlock idle infrastructure, is Greg Osuri, CEO of Overclock Labs and founder of Akash Network. Greg, great to have you here. Thank you so much for joining us. So I understand that you have testified before Congress before on AI's energy demand as well as climate impact.
So how severe is this energy crisis and why are traditional centralized data centers colliding with climate goals?
Yeah, it's very obvious. I think the U.S. predicts anywhere from 12% to 8% of U.S. energy will be used by AI data centers by 2030. In reality, that's more like 30%. So it is a major crisis that we haven't quite fully understood. The crisis stems all the way from energy production to the supply chain, all the way down to even getting transformers and everything you need to get energy, right.
So, I testified a few years ago before Congress. I've been very vocal about this issue for several years. We've been catching a marketplace for compute. And, you know, our provider base itself. We see this issue brewing up for a few years. And particularly challenging to hyperscalers is because of the political pushback we're seeing with increased energy prices.
Communities are pushing back against building data centers in their local communities. And I think that impacts negatively on the supply chain for the chips itself. So there is a new credit crisis that's brewing because of these cancellations and whatnot. So, I think the crisis is an interesting place.
On one hand, we do have an energy chip shortage. In another hand, we have a pushback against bringing these chips to the market. So it's a very interesting place right now.
Yeah. And, Greg, while I have you here, I do want to get your perspective on access to power. So we all know tech giants have hundreds of billions in capital, but funds can't actually shorten grid build outs or permitting timelines here. So tell us about this access to power and how it's become the real hard ceiling when it comes to AI development.
Yes. So already energized data centers that have long power purchase agreements are in a good place. And that tend to be several crypto miners and several institutions that locked in to enormous, incredible prices a while back. But any new data center build outs are having a very challenging time because you got to understand.
The energy industry does not move as fast as the software industry, right? So while software, while AI is moving incredibly fast, much faster than what software normally moves, energy industry in America is unable to catch up because it's an incredibly complex industry. While there is an effort to bring renewables, unfortunately, renewables are not great at scale.
So I think the landscape you're seeing now is, instead of bringing energy to where the chips are, taking chips to where the energy is. That tends to be a new appealing trend, as my testimony indicated. So we are seeing a lot of a shift from massive hyperscalers to more nimble edge data centers, which could be a potential solution to the energy crisis that we have.
And further, we can even push these data centers to smaller communities that benefit the smaller communities immensely as well. So it's happening right in front of our eyes. And I think in 5 to 6 years, I think the distributed computing and the distributed local AI community will shine over the hyperscaler setups that we have today.
Yeah. And finally, Greg, before I let you go, I do want to get your take on underutilized compute. So what are you doing when it comes to putting idle capacity to work and giving buyers out there certainty on price when it comes to future compute?
It's an interesting point. So right now we do have a shortage of chips, but the shortage is really in the infrastructure more than the actual production, at least from an inference standpoint. There are enormous amounts of GPUs out there, either older models that are sitting unused in data centers or clearinghouses or even consumer grade models that you're seeing in homes behind gaming PCs.
These are powerful GPUs that can be leveraged to deliver intelligence at significantly lower cost. So if you combine the energy crisis where it's incredibly hard to get utility scale energy supply in traditional places, along with the chip crisis, it's very easy to see where the solution is.
As the local AI and local inference technologies are getting better at delivering frontier intelligence right from your home, we're going to see a good portion of intelligence moving from a traditional data center to what I would call home token factories or commercial token factories.
There's enormous underutilized real estate out there in the commercial space that are sitting on great supply of energy that potentially could come from renewables as well, becoming a future data center. So Akash Network has been primarily focused on connecting these small, nimble compute to a grid and also solving the problem of making compute, which is usually locked behind opaque contracts, into more liquid and easily fungible instruments that can be traded and can be speculated as well to build more predictable pricing.
So we're really very excited to where the buck is moving from spot markets to forwards to future markets. And Akash is positioning or working fairly quickly to deliver those predictable prices to consumers.
Well, Greg, we will have to leave it there. Thank you so much for joining us this morning. And thank you so much for sharing all of your insights.
Thank you so much for having me.
