The Power to Build. The Bravery to Change

By Bill Kleyman, Apolo

---

This chapter is an excerpt from Greener Data: Volume Three, launched on Earth Day 2026. Featuring perspectives from 75+ sustainability leaders across the digital infrastructure ecosystem, the full book is available now on Amazon.

---

Watt’s up, everyone? 

You know it’s going to be a fun chapter when we kick things off with an energy pun. Trust me ,  it’ll all make sense soon. Look, writing an entire book is hard, and I have so much respect for the brilliant authors who can pull that off. Me? I’m just thrilled to carve out my own little section in this incredible collection and share a few pages with you.

I’m genuinely grateful that you’re spending a few minutes of your day here with me, because we’ve got a lot to talk about. From power to bravery (and a few volts of humor along the way), this chapter dives into the big shifts shaping how we build, sustain, and rethink our digital world. But this isn’t just a story about building more infrastructure, it’s about learning how to build it smarter, more efficiently, and with a clear understanding that every watt we deploy carries long-term consequences.

No ChatGPT here today. Just you, me, and a whole lot of Billawatts.

Let’s start here…

If you read the previous Greener Data publications (which I highly recommend), you’ll know my story started in a pretty unconventional way. Back in Soviet Ukraine, my brother competed in national telegraph competitions. That’s right. I grew up with the sounds of boops and beeps in our apartment. 

He used an old telegraph switch that our grandfather had modified so he could tap just a little bit faster. When he practiced, he’d let me sit on his lap and listen to the dots, dashes, and the rhythm of connection. He taught me numbers, letters, and how to communicate with people across the Soviet Union. Looking back, that’s probably where my fascination with connection and bringing people closer truly began.

Fast forward to today. About 30 years later, we’ve come an incredible distance as a connected world. From clicking relays to cloud regions, we’ve built an era where communication is instant and seemingly limitless. I was reminded of this during my keynote at AFCOM Data Center World POWER 2025. On the main stage, I shared a bit of breaking news: the night before, at midnight, AOL officially shut down its dial-up service. Forever. 

No more away messages, no more “You’ve got mail.” And yes, all those free AOL CDs we saved? Officially useless.

We’ve gone from the screeching symphony of a dial-up modem (you’re hearing it in your head now, aren’t you?) to the quiet hum of fiber optics threading through hyperscale data centers. It was the perfect moment to reflect on how far we’ve come. From dial-up to fiber, from copper to cloud, from the analog age to a completely digital society. Each leap forward brought incredible capability, and, whether we acknowledged it or not, a growing energy and environmental footprint that we’re now being forced to confront.

As a millennial, my generation became the bridge. We’re the last to remember 8 curfews, phones mounted on the kitchen wall, and life without connectivity. And now, we’re the ones helping to define what comes next.

As we explore what’s happening in our industry right now, I firmly believe you only need two things to be successful: Power and Bravery. Power without efficiency is waste. Bravery without responsibility is short-lived. In today’s data-driven world, sustainability sits right at the intersection of both.

Let’s explore what that means.

On November 30^th, 2022 – The AI Floodgates Opened

I don’t want this to become a lecture on ChatGPT. Everyone reading this can use their favorite search engine to type in a question and get a fun generative AI definition of this technology. However, we have to point out the meteoric event that happened merely three years ago. 

At the end of 2022, something extraordinary happened. OpenAI launched ChatGPT. For the first time, the world got a front-row seat to the power of large language models through a simple, friendly chat window. Within days, it felt like everyone was trying it. Teachers, engineers, marketers, even grandparents,  all experimenting, all amazed.

The growth wasn’t just fast; it was truly historic. ChatGPT became the most rapidly adopted technology in human history, and it hasn’t slowed down since. By late 2025, it’s serving more than 700 million weekly active users, powered by hundreds of integrated plugins and partnerships that make it feel less like a tool and more like an ecosystem.

But while ChatGPT was widely adopted, it’s not actually the end-all for this kind of AI. 

So here’s the thing: We now live in a world that’s discovered oil… but hasn’t quite invented the internal combustion engine yet. There’s too much raw material, and very few know how to actually use it. And just like oil, it’s suddenly everywhere. In our food systems, in medicine, in buildings, in manufacturing, and even in your shoelaces.

Right now, the question isn’t “How do we find more oil?” It’s “How do we turn that barrel of oil into a shoelace?” And not just any shoelace. One that actually ties the shoe, lasts a while, and doesn’t cost as much as an entire shoe store.

That’s where what we do at Apolo.us comes in. 

Apolo was built to be that missing engine. We are the interoperability layer that actually makes AI work. Just as importantly, this approach enables right-sized AI deployments,  reducing overbuilt infrastructure, minimizing idle compute, and turning sustainability into an architectural outcome rather than an operational afterthought. It’s a powerful piece of software running on Kubernetes that lets data scientists build, train, and deploy any kind of AI or ML project, from classic machine learning to modern LLMs and generative AI. The magic? Apolo is one of the very few platforms on the market that can do this at an enterprise or multi-tenant data center scale.

You read that part right. Any data center operator with spare GPU or even CPU capacity can deploy Apolo, put their logo on the landing page, and suddenly it’s their own private, branded AI cloud. Basically, a neo-cloud in a box. It looks and feels like their very own mini-Amazon, except Apolo handles all the heavy lifting: MLOps, orchestration, model deployment, everything.

On top of that, Apolo AI Launchpad delivers ready-to-deploy AI applications designed for specific industries, allowing organizations to rapidly adopt AI without rebuilding their entire software stack. These solutions run securely within private clouds or data centers, ensuring sensitive enterprise data and models remain fully controlled and compliant with strict regulatory requirements. 

The result is something incredibly powerful: a data center that doesn’t just host infrastructure, but launches real AI applications, from secure enterprise copilots to analytics platforms and industry-specific agents, directly where the data already lives.

And because Apolo manages its own GPU clusters, platform stack, and enterprise-ready AI app delivery ecosystem, we’re not just powering the infrastructure, we’re enabling the real-world AI solutions being built on top.

In this rapidly fragmenting market, there are three main types of players:

1.Those with 500,000+ GPUs training massive foundational models.

2.Enterprises deploying 1,000–5,000 GPUs for private LLMs and specialized models.

3.And finally, the inference layer, where all that training turns into practical, usable AI applications.

That third category? That’s our home turf. And where we feel this market is absolutely going.

We focus on inference workloads. This is the AI that actually does something. Our models and data are deployed securely on-prem, in fully compliant and regulated data centers. Not a single bit or byte ever leaves the cluster to train someone else’s foundation model. That is why Apolo partners closely with organizations across critical industries such as government, defense, healthcare, and advanced manufacturing. These are industries that require total control of their data and models.

To accelerate real adoption, we built Apolo AI Launchpad. Launchpad allows organizations to rapidly deploy production-ready AI agents and applications directly inside their secure environments. Instead of starting from scratch, teams can launch industry-specific AI solutions on infrastructure they already trust and quickly turn raw AI capability into meaningful operational outcomes.

This is important because of a shift happening in our industry.

Here’s the fascinating part: right now, the market is about 80% training and 20% inference. But within the next 24–36 months, that ratio is going to completely flip. The world will have trained enough AI. The next era is about using it. That is, deploying it, scaling it, and making it work for people and businesses.

And this is where Apolo is already creating real impact (shoelaces that actually tie a shoe) with AI Agents and AI Apps that move the needle:

•A Specification AI Agent helping manufacturing clients save hundreds of hours every month by automating complex design and compliance documentation.

•A Manufacturing/Engineering Co-Pilot that captures human expertise and supports junior engineers with real-time, contextual guidance.

•Legislative AI systems that simplify and accelerate policy research for government agencies.

•And AI-powered student portals helping universities modernize curriculum design and drive student success through personalized learning.

These use cases aren’t science fiction! They’re inspiring, life-changing, and transformative. But let’s be honest, none of this came without a few bumps and bruises along the way.

Because while we were ready for the data center industry… the industry wasn’t quite ready for us.

And that’s exactly why, in this new era of AI-driven infrastructure, the two most important ingredients for success are Power and Bravery.

Let’s dive into what that really means.

Power and Bravery: The New Currency of Innovation

It’s almost impossible to capture just how fast this revolution is moving. For the better part of the last 25 years, every one of us, whether you’re a technologist, a student, or a CEO, has been trained to interact with information the same way. You’d open your favorite browser, type a question, and wait for that familiar list of blue links. Decades of habit built around a simple hyperlink.

Just like our AOL reference, those blue links are fading into history. Today, when you ask Bing or Google a question, the first thing you see isn’t a list of websites. Now, you see an answer. It’s generative AI. The interface of search itself has changed right in front of our eyes, and it happened practically overnight.

ChatGPT hit one million users in just five days. Nine months later, it was nearing a billion unique users.¹ 

Now, as internationally recognized tech and energy expert Peter Gross once said, “AI isn’t a revolution. It’s an acceleration.” And he’s right. The moment ChatGPT went live, a storm rolled through the data center industry. Everyone wanted in, hyperscalers, colos, enterprises, everyone, but most weren’t ready for the kind of demand that was coming. And this is where Peter’s statement comes into focus. How ready were we for this AI acceleration?

Supporting AI workloads is nothing like managing SQL databases or Exchange servers. It’s a completely new physics of computing, and here’s what we’ve learned:

Density is skyrocketing.

When we released the very first AFCOM State of the Data Center report nearly a decade ago, the average rack density was roughly 6 kW per rack. Over the years, that number steadily climbed as virtualization, cloud adoption, and early AI workloads pushed infrastructure limits. But the most recent findings show something very different. In the 2026 study, average rack density has surged to approximately 27 kW per rack, representing the largest year-over-year jump we have ever recorded and nearly four times what many facilities were running just a few years ago.

(For context, a typical Microsoft rack in the late 1980s consumed roughly 1 kW. We have come a very long way.)

And this is not just a capacity challenge. It is fundamentally a sustainability challenge. As rack densities accelerate and AI workloads dominate infrastructure planning, operators are being forced to rethink how much compute we deploy, how efficiently we cool it, and whether every workload truly needs maximum power all the time. The data center industry is no longer simply scaling infrastructure. It is redesigning how digital infrastructure consumes energy in the first place.

Cooling is in trouble.

Cooling has quickly become one of the most pressing operational constraints in modern data centers. In the 2025 AFCOM State of the Data Center Report, about 34% of respondents reported that their cooling solutions did not meet all of their operational requirements. Just one year later, the 2026 report shows that number has climbed to 39%, highlighting how rapidly AI workloads are pushing beyond the limits of traditional thermal designs.

This trend reinforces a reality many operators are already experiencing. Traditional air-based cooling alone is increasingly insufficient for high-density AI infrastructure. As GPU clusters and AI training environments drive rack densities higher, facilities must rethink how heat is managed and removed from the environment.

One approach we’ve found particularly effective is leveraging hybrid cooling strategies. Instead of ripping out and replacing entire environments, many operators are adapting existing infrastructure. Our partners have successfully modernized legacy deployments using rear-door heat exchangers (RDHx), improved containment, and smarter airflow optimization to support new AI workloads.

Even smaller facilities can accomplish a great deal with the right engineering approach. In many cases, the most sustainable move is not building something entirely new. It is extending the life, efficiency, and capability of the infrastructure that already exists. The industry should be proud of that progress. But the reality is that it’s still not enough.

To put this in perspective, a single NVIDIA DGX system with eight H100 GPUs takes up six rack units and draws between 10–16 kW on its own. That means the average data center rack can only handle one of those machines before maxing out. One node. That’s it.

And that’s the wake-up call.

Most data centers were built for yesterday’s workloads. That is, ERP systems, email servers, and web apps. Now, they’re facing racks filled with AI training clusters that devour power, demand liquid cooling, and weigh half a ton. Racks are changing. Cooling is changing. Power delivery is being completely redefined.

Let’s stay on that last point … power.

Let’s talk about power… and I don’t just mean metaphorically (though we’ll get there). I mean real electricity, gigawatts, megawatts, racks humming with compute, and grids groaning under the load. It really feels like our industry isn’t even getting out of bed these days unless we’re talking about a gigawatt data center.

Because if you’re going to succeed in this new era of AI-driven infrastructure, you’ve got to master the watts as much as the code.

According to S&P Global, US data centers are projected to need roughly 22% more grid-based power by the end of 2025 compared to 2024.² That’s a leap. 

And when you jump ahead to 2030, the increase is nearly three times what we’re using today.³ Meanwhile, a recent write-up from the Pew Research Center shows that US Data centers consumed about 183 terawatt-hours of electricity in 2024, which is more than 4% of the nation’s total electricity use,  and it’s expected to grow 133% by 2030.⁴ 

According to McKinsey analysis, the United States is expected to be the fastest-growing market for data centers, growing from 25 GW of demand in 2024 to more than 80 GW of demand in 2030.⁵ Other estimates, like those from the Boston Consulting Group, put demand upwards of 130 GW.⁶ The latest metrics just put that demand number at 200 GW.⁷ Instead of just throwing around some numbers. Let’s put 100GW of power into perspective. 

•1GW is enough for a city of 1 million people

•10GW is enough for an LA or NYC

•100GW is about 10% of all of the lights currently on in the whole world.

So, when we see estimates of 100GW over the next four years, my immediate thought is, “How are we going to build enough power for twenty LAs or twenty NYCs?” It’s extraordinary. The real question isn’t just where that power comes from, but how responsibly we generate it, distribute it, and avoid wasting it through overbuilt or underutilized infrastructure.

What’s driving this surge? Simple: the explosion of AI and the compute density that comes with it. Coupled with the overall population digitalization. Where once a rack with 6 kW (or less) was acceptable, our most recent data shows average densities around 16 kW per rack and climbing. We’re seeing compute nodes with GPUs pulling 10-16 kW each; that means one rack can be consumed by just one such node in many facilities.

And here’s the key point: you can’t separate power from innovation. If your facility’s grid connection, UPS system, cooling infrastructure, or power-distribution architecture isn’t built for 10x what you planned even a few years ago, you’re already behind. The world we’re building isn’t asking “Can we handle it?” It’s asking “How much more can we take? How much faster can we spin?”

So remember: the “power” in power and bravery has two faces. One is the literal force: the megawatts, the racks, the cooling loops, the reliability, and the grid-access. The other is the figurative: the power to make choices, to lead, to invest, to design infrastructure that isn’t just incremental but exponential.

Now, let’s move into the other half of the coin: bravery, and explore what it takes not just to have power, but to wield it.

When Power Meets Bravery: How Bold Choices Build a Greener World

Since the ChatGPT floodgates opened, we’ve worked closely with clients across nearly every sector. This includes fascinating companies across manufacturing and healthcare, as well as higher education and government. And while the use cases vary, the lessons are strikingly similar:

1.Control is everything. Organizations want sovereignty over their data, models, and outcomes. They’re tired of their information training someone else’s LLM.

2.Privacy and cost go hand in hand. Running secure, private AI environments gives organizations control over their data and compute usage. Keeping workloads on dedicated infrastructure helps avoid unpredictable cloud fees, reduces data transfer costs, and keeps budgets more predictable.

3.Openness fuels innovation. The most successful clients demand transparent, interoperable platforms that empower their developers.

4.Efficiency matters. Sustainable AI deployment means optimizing ,  not overbuilding ,  power, cooling, and resources. Done right, sustainability becomes an architectural advantage, reducing cost, improving resilience, and enabling AI to scale without scaling waste.

5.Sustainability is a strategy. The smartest companies now view energy efficiency and responsible design not as PR moves, but as long-term competitive advantages.

These are the real conversations reshaping enterprise AI.

Back in 2021, IEEE warned in its report Deep Learning’s Diminishing Returns that “the cost of improvement is becoming unsustainable.”⁸ Fast-forward to 2025, and IEEE’s AI Index paints an even sharper picture. The cost to train leading models like Gemini 1.0 Ultra has soared to nearly $192 million, while parameter counts and data sizes have exploded.⁹ Yet, there’s good news: inference efficiency has improved dramatically. The cost of using AI is dropping as hardware performance and energy efficiency rise.

That’s bravery in action: confronting unsustainable growth (both financially and environmentally), learning, adapting, and finding smarter paths forward.

Bravery in business means more than risk-taking. A lot of it has to do with self-awareness. It’s the willingness to ask, “Who are we, and who are we becoming?” As The Innovator’s Dilemma reminds us, what once made us great can make us obsolete if we refuse to evolve.

And in sustainability, bravery is redesigning what’s possible, such as:

•Deploying liquid and hybrid cooling systems that handle megawatt racks without wasting water.

•Embracing microgrids, hydrogen, and SMRs to build energy independence.

•Partnering across the ecosystem to create shared models of efficiency, not silos of consumption.

At Apolo, that bravery takes tangible form. We’re embedding a green tracker directly into our software so users can visualize the carbon impact of every AI job. We partner with data centers powered by sustainable energy sources, and we source second-life hardware through circular-economy suppliers to extend compute lifecycles responsibly. And perhaps most importantly, we’re laser-focused on selecting the right infrastructure for the right task, no overkill, no waste.

As you think about your AI journey, consider these three points:

1.Design for flexibility. Build platforms that can pivot from training to inference without rebuilding the foundation.

2.Commit to sustainable density. Power and cooling must scale intelligently,  not endlessly.

3.Be fearless in experimentation. True innovation comes when we’re willing to rethink comfort zones, partnerships, and paradigms.

I truly believe that we’re standing at the intersection of power and bravery. At this crossroad, we see technology meet conviction. The choices we make now won’t just define data centers; they’ll define how humanity powers intelligence itself. Ultimately, all the GPUs, grids, and gigawatts in the world won’t matter without the guts to use them differently. That’s the heart of real innovation: power and bravery, working together to light a greener, smarter world.

RESOURCES

1. Hu, Krystal. “ChatGPT Sets Record for Fastest-Growing User Base—Analyst Note.” Reuters, 2 Feb. 2023, https://www.reuters.com/technology/chatgpt-sets-record-fastest-growing-user-base-analyst-note-2023-02-01/

2. Hering, Garrett, and Susan Dlin. “Data Center Grid-Power Demand to Rise 22% in 2025, Nearly Triple by 2030.” S&P Global Market Intelligence, 14 Oct. 2025, https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/101425-data-center-grid-power-demand-to-rise-22-in-2025-nearly-triple-by-2030.

3. Skidmore, Zachary. “S&P Global: US Data Centers to Require 22% More Grid-Based Power by End of 2025.” Data Center Dynamics, 15 Oct. 2025, https://www.datacenterdynamics.com/en/news/sp-global-us-data-centers-to-require-22-more-grid-based-power-by-end-of-2025/.

4. Leppert, Rebecca. “What We Know About Energy Use at U.S. Data Centers Amid the AI Boom.” Pew Research Center, 24 Oct. 2025, https://www.pewresearch.org/short-reads/2025/10/24/what-we-know-about-energy-use-at-us-data-centers-amid-the-ai-boom/.

5. Green, Alastair, et al. “How Data Centers and the Energy Sector Can Sate AI’s Hunger for Power.” McKinsey & Company, 17 Sept. 2024, https://www.mckinsey.com/industries/private-capital/our-insights/how-data-centers-and-the-energy-sector-can-sate-ais-hunger-for-power.

6. Green, Alastair, et al. “Power Moves: How CEOs Can Achieve Both AI and Climate Goals.” Boston Consulting Group, 8 Nov. 2024, https://www.bcg.com/publications/2024/ceos-achieving-ai-and-climate-goals.

7. Kleyman, Bill. “Power, Bravery, and the Wild Data Center Future Ahead.” AFCOM, 4 Dec. 2025, https://afcom.com/news/715847/Power-Bravery-and-the-Wild-Data-Center-Future-Ahead.htm.

8. Thompson, Neil C., et al. “Deep Learning’s Diminishing Returns.” IEEE Spectrum, 24 Sept. 2021, https://spectrum.ieee.org/deep-learning-computational-cost.

9. Strickland, Eliza. “12 Graphs That Explain the State of AI in 2025.” IEEE Spectrum, 7 Apr. 2025, https://spectrum.ieee.org/ai-index-2025.