Hydrogen Will Fuel the First Zero-Emission Gigawatt Data Center
Interview with Yuval Bachar of EdgeCloudLink
Episode Preview:
Episode Overview:
Is a zero-emission, gigawatt data center possible? The CEO of EdgeCloudLink says yes, but only if developers are capable of substituting hydrogen for natural gas as a source of energy.
Speaking on the sidelines of the DCD New York event in March, Bachar tells Cool Vector his company already runs a zero-emission data center in Mountain View, California. The success gave ECL and partners the ambition to aim for a gigawatt project in Texas.
Bachar estimates pipeline hydrogen runs at roughly six to seven cents per kilowatt hour - on par with natural gas, making the economics more competitive than widely assumed.
Key takeaways from Bachar’s interview:
Modular data centers are the only way to keep pace with GPU generation cycles. “Whatever we design right now to be delivered in 2028 is going to be too late — that’s three generations of Nvidia in the process,” says Bachar.
Hydrogen-powered data centers produce zero emissions and zero water waste by closing the loop between generation and cooling.
Speed to deployment has completely eclipsed sustainability as the primary purchase driver. Says Bachar: “The king is time to token, and people are willing to compromise on sustainability.”
A hydrogen-based gigawatt AI factory is achievable in three years from a committed customer.
Full Episode:
Full Audio:
Full Transcript:
Hydrogen Will Fuel the First Zero-Emission Gigawatt Data Center
Interview with Yuval Bachar, CEO, EdgeCloudLink
Yuval Bachar: My name is Yuval Bachar. I’m the founder and CEO of ECL — Edge Cloud Link.
The ECL platform initially started as a complete off-grid, hydrogen-based generation with zero emissions and zero water consumption to deliver high-performance data centers. How did you evolve from that starting point?
In 2023, when we shifted into the AI domain, we made a quick pivot and created an AI data center that is modular — able to be built in small form factors and larger form factors — enabling us to be highly competitive in a market where, overnight, 90% of the world’s data centers became obsolete. That’s unbelievable to most people.
Typical colocation data centers were running about seven to eight kilowatts per rack. We were at 75 when we started. So when the GB 300 came out requiring 150 kilowatts, for us it was a 2x increase. For the rest of the world it was 20x. There’s a very big difference in how you close that gap.
What are the benefits of modularity in data center development?
Yuval Bachar: Modularity gives you multiple things. First, you can build very quickly — turning around data centers within nine to twelve months instead of the twenty-four to thirty-six months required for traditional data centers.
Second, because of that fast turnaround, you can track the technology. Technology is now spinning every twelve months. At GTC last week, Jensen made it loud and clear: we will continue changing the technology every nine to twelve months. Data centers that are not modular and cannot be agile will always be behind. Whatever we design right now to be delivered in 2028 is going to be too late — that’s three generations of Nvidia in the process.
Third, modularity lets you scale on demand. There’s not a lot of predictability in the market right now. People don’t know how much they need or when they need it. We build modules when we need them and can expand a site very quickly with additional modules. These are structures — not containers — and they enable us to reduce what’s called the blast radius.
Large data centers today operate in an environment where the whole site is centralized around a single power station or cooling system. If that fails, it takes down the entire site. When you build independent, autonomous blocks and combine them, the biggest failure domain you have is a single block. Today, in a one-megawatt block, we can put hundreds — in some cases thousands — of GPUs.
How does your modular building process work?
Yuval Bachar: Every order we receive from a customer is for a number of identical modules, so we just replicate and copy exactly. But the second batch of modules can be completely different. The encapsulating system around it — the power system and the cooling system — will be the same, but on the inside we change completely according to the latest technology. From the original DGXs, to the GB 200, to the GB 300, to Vera Rubin — it all fits into the same module, just a different configuration inside.
How are your data centers able to accomplish zero emissions?
Yuval Bachar: When you’re running on hydrogen — as we are at our site in Mountain View, where we’ve been operating for two years — you achieve zero emissions and zero water consumption because when you generate energy with hydrogen, the byproduct is water. We use that water for the cooling system.
We don’t provide the hydrogen ourselves. It’s delivered to us by our suppliers, but the generation plant is ours.
If you draw hydrogen from a pipeline, the cost to generate energy is comparable to natural gas — roughly six to seven cents per kilowatt hour. If you’re operating in an environment without a pipeline for either fuel, hydrogen will push you to around sixteen to seventeen cents per kilowatt hour, while natural gas will run around twelve to thirteen cents. Natural gas is easier where a pipeline exists, and pipeline natural gas is available on virtually every street in the United States, which drives the cost down even further.
Why aren’t more data centers using hydrogen as a fuel?
Yuval Bachar: It’s difficult. I always like to quote JFK, who said we went to the moon not because it’s easy, but because it’s hard. It was a complicated process to build an environment that can actually run on hydrogen. It required us to completely redesign the power system — we didn’t only move to hydrogen generation, we also eliminated the UPS system and the diesel generators and went to an active-active environment. That worked very well for us because AI requires power all the time.
It’s not simple. In today’s world, data centers are getting more and more complex to build, so operators are trying to minimize complexity in the power stack. But I think more and more people are interested in hydrogen now and working to deliver it, because hydrogen is available at high volumes, and if you access it from the right place at the right time you can generate energy at a very reasonable price.
How much of your sales are driven by enthusiasm for sustainability versus other factors?
Yuval Bachar: When we started, most of the interest and enthusiasm was around the zero-emission, zero-water story. Nowadays, when we go to customers and tell them we can deliver in nine months, it’s AI-ready, it’s at a relatively low cost, and it’s sustainable — the answer we get is, “Whatever you said, nine months. That’s all we care about right now.”
The king is time to token, and people are willing to compromise on sustainability — unfortunately, at least for this short bridging period where we don’t have enough grid capacity.
If you were to start building a new AI gigafactory with your modular, zero-emission approach today, what would the project look like?
Yuval Bachar: It would look like an array of modules connected together — one-megawatt, two-and-a-half-megawatt, and five-megawatt blocks — built in stages. Our stages are typically 150 to 200 megawatts. To build a gigawatt is about five stages, which takes roughly three and a half years. But the first delivery to the customer comes after nine to twelve months.
Give us an update on the zero-emission AI gigafactory you are building in Texas.
Yuval Bachar: We have a site in Texas called the TerraSite, and that site is capable of taking hydrogen and building a gigawatt of data center based on hydrogen. We are the first to run completely off-grid on hydrogen, and we’ll be the first large-scale hydrogen-based data center in Texas. We can’t say the name of the customer, but it’s one of the large neo-clouds.
The move to build an AI factory of one gigawatt was a big deal. We were sitting with our advisors and board asking: are we building twenty-five megawatts? After a lot of discussion, we said, no — we are building a gigawatt. Let’s go for it.
We started, and it was very ambitious at first. We’re a startup, and building a gigawatt data center is complicated. But we believe we did the right thing, even though it’s taking a little longer than expected, because you have to show that you can scale. Going from one to twenty megawatts is great, but the factories right now require hundreds of megawatts.
On the other side of it, inference is becoming very dominant. For us, inference is the most natural fit — instead of a hundred blocks, you put five or six, and it’s the exact same block. Our customers will have the same experience on the inference side and on the factory side because it’s the same block, just a smaller number of them.
If you started today, how long would it take to build a zero-emission AI gigawatt factory?
Yuval Bachar: If we have a customer committing to one gigawatt, it’s going to take us about three years to bring it up. We have to make sure we can deliver the hydrogen fuel cells — we’re talking about tens of thousands of fuel cells that we need from our suppliers — and secure access to the hydrogen.
Hydrogen today gives us the ability to scale initially to 250 megawatts, but that’s growing. We’re in Houston, where there’s a significant amount of hydrogen available in pipelines from multiple suppliers. Hydrogen availability is pretty straightforward at this point. The fuel cell and the rest of the system is what we need to work on, but we believe that in around three years we can bring a high-gigawatt, zero-emission site online.
What is challenging about working with hydrogen in digital infrastructure?
Yuval Bachar: The looks you get when you say “hydrogen” — people immediately tell you it’s probably very dangerous, and you have to explain why it’s not. It’s a safe process and we know what we’re doing.
Beyond perception, you’re penetrating a whole new market. The hydrogen market was never designed for stationary power for a 24/7 data center with continuous availability and cost requirements. It took us about a year and a half to break through with all the hydrogen suppliers and get them to recognize this as a real offtake — a different kind of offtake, not a refinery or petrochemical factory. It’s a new kind of customer that does not accept downtime on the pipeline and does not accept maintenance cycles.
When we started talking to hydrogen suppliers about stationary power, they had no idea what we were talking about. Now they’re much more versed in it, and the same evolution is happening with natural gas.
What did you learn while building data centers for Facebook and LinkedIn from 2013 to 2019?
When I joined LinkedIn, we were at a very early stage of our data center buildout, and LinkedIn gave us the freedom to do what we thought was the best thing. We were pushing our racks to twenty kilowatts when others were running four to five kilowatts per rack, and we were trying to project into what the future would require.
Facebook gave me a much bigger experience: just do whatever you can. I always like to quote the statements on the walls at Facebook that came from Zuckerberg: “What would you do if you were not afraid? Go do it.” That’s exactly what ECL is. What would we do if we weren’t afraid? We’re not afraid. We’re just going to do it.
