Building and Servicing the Next Wave of Data Centers

Joshua Buzzell, Executive Vice President of Engineering at Bitdeer, joined Blake King, Head of Power at Galaxy Digital, and Giovanni Mariani, Technical Manager for North America at LU-VE Group, on a panel titled "Building and Servicing the Next Wave of Data Centers" at the recent Energy Investors Forum in Irving, Texas. The discussion was moderated by BlocksBridge Consulting.

The following are Buzzell's remarks, lightly edited for length and clarity.

Converting Mining Sites into AI Data Centers

Jesse (BlocksBridge): Bitdeer is in the process of converting a number of sites to AI and HPC. What does that conversion look like from an engineering and design perspective?

Joshua: I think about it in three portions. There's the engineering component, when we're thinking about lines and drawings. There's also a process component. And the last portion is human capital.

Crypto miners are used to a one-nine (90%) of reliability. The worst thing that happens if you interrupt a crypto miner is you may make less Bitcoin. With an AI data center, you have a more significant problem, where you need three nines (99.9%) to six nines (99.9999%) of reliability depending on the problem set.

Once you interrupt service, you have the potential of impacting another business, and there are reputational issues. But more importantly, you're impacting a human experience improvement for someone. A human experience improvement that excites me is hearing aids, which enabled my mother to hear her grandkids again. That's the type of impact you can have.

Back to the engineering layer, it's important to think about reliability. How do you have redundancy inside that system? From a process standpoint, you're typically going to need a change review board when you're implementing changes, whether electrical, mechanical, or network, on an operational site.

And then lastly, there's the human capital component. If you look at a crypto mining site, the human capital and processes required will be quite different. Think about something as simple as housekeeping, which is maintained to a much higher degree because there are thousands of cables and wires that have to be protected from dust and particles.

Behind-the-Meter Generation

Jesse: Behind-the-meter generation has been one of the major trends of the past couple of years. How are you thinking about it, as opposed to just plugging into the grid, and do you see it as a long-term trend?

Joshua: The easiest answer is that it depends. You need to think through the problem set, because you can't think of the market as monolithic. Even the United States market is broken up into probably three or four large chunks. Then there's the question of what level of reliability you and the end user might be willing to accept. Those gates will determine whether or not a site may be suitable for on-site generation.

Depending on each of those problem sets, there are possibilities to do it. If you're talking about a thermal solution, which would be stranded natural gas, going simple cycle first and then phasing into a combined cycle might get you the higher efficiency rates to support that. But at the end of the day, it all comes down to whether the power is going to be on when the customer wants it.

Water Cooling and Heat Reuse

Jesse: Bitdeer has facilities all over the world, from cold Norway to hot Texas. How do you think about cooling?

Joshua: If you're a crypto miner, you're probably familiar with a hydro miner. Hydro mining gives you more hash because you have a warm water cooling space.

Then you have the DLC market, with PG-25 (25% propylene glycol), which is similar to the hydro miners and was the first implementation of direct-to-chip. Initially, because of the way the chips were designed, the colder it was, the more efficient it was. But the silicon has since been redesigned in the GPU market, so that's not always true anymore, and that allows you to go up to 40°C now.

What that opens up is energy reuse effectiveness, one of the KPIs data centers are measured on. When you're warmer, you can probably get 70°C to 90°C output. That opens up things like growing mangoes in Norway.

Instead of just thinking about cooling as blackbody radiation, where we exhaust it to the atmosphere, the question is how you can reutilize energy you're already going to use, to further monetize it and reduce the overall amount of therms utilized by the full system.

Latency, Data Gravity, and Where Compute Gets Built

Jesse: What about the data component of a data center, specifically latency and where the data rests, and how the industry is evolving around that?

Joshua: There's been an order of operations associated with data center placement and anything in the digital economy. Typically you look at where the eyeballs are, where the data gravity is, the latency component, grid power, and finally fuel that can be converted into electrons.

From a crypto perspective, it's about how cheap the power is, either connected to the grid or as fuel that can be converted into electrons. Creating the hash creates a coin; it's a token. Same thing with GPUs: you're creating tokens.

In traditional data centers, you care about where the people are, where the cell phones and routers are. That's why you have data gravity, which determines where the business clusters are, and latency.

This latest GPU boom is not as latency sensitive as traditional loads. The big question now is how latency sensitive it will be with inference, which has different workloads.

The Bottleneck Nobody Is Pricing

Jesse: It's 2028. What's the biggest bottleneck that nobody is pricing today?

Joshua: The cooling S-curve and the power S-curve from the shift to direct current. That's probably going to peak based on what has been released about where the GPUs are going right now. And then you'll probably get into financial constraint after that.