A 1 Gigawatt geothermal-powered data center? That’s the bold vision behind Tuesday’s announcement that Google has agreed to buy 396 megawatts (MWs) of power from Fervo Energy, with an option for an additional 600 MWs

The companies say this is the world’s largest enhanced geothermal systems (EGS) contract to date, and could ultimately power a Google data center in southwest Utah.

Unlike solar and wind, which are intermittent, geothermal taps the Earth’s subsurface heat to generate electricity around the clock, regardless of weather or time of day. That’s a critical feature for data centers, which seek continuous, high-reliability power.

Even so, geothermal power has been a rarity in data centers, most notably used by operators in Iceland. The development of enhanced geothermal offers a broader pathway to the “firm,” always-on power profile that utilities associate with fossil fuels, without the emissions.

Interest in the technology has surged in recent years, with Meta, Microsoft and Amazon Web Services all pursuing geothermal pilots. This week’s Google deal represents the largest effort to power new data centers with heat from the earth.

“This agreement reinforces that EGS is ready to power the next generation of computing infrastructure,” said Tim Latimer, CEO and co-founder of Fervo Energy. “As demand for reliable electricity grows, customers like Google need energy resources that can be deployed at scale, operate around the clock, and deliver where power is needed.”

Building on a Nevada Blueprint

Fervo will provide Google with electricity from its Cape Station GeoCluster in Beaver County, Utah, which the company plans to bring online in 2028. Google has the option to expand the agreement by 600 MW, pushing the total toward nearly 1 gigawatt by June 2030. That would be enough dedicated geothermal capacity to run a major hyperscale data center campus around the clock.

The power will serve as a “foundational building block for a potential data center in Utah,” but Google says a development decision will depend on engineering studies and state and local approvals.

For Google, the deal represents a major step up in scale for its geothermal program. The Utah deal builds on a partnership that began in Nevada with Project Red, Fervo’s 2023 pilot.

That was followed in June 2024 by a 115-megawatt PPA with Google and NV Energy. That deal helped pioneer Nevada’s Clean Transition Tariff, allowing large energy users to directly fund clean power projects without shifting costs onto other ratepayers, a big priority amid the current wave of public scrutiny of digital infrastructure.

The Utah deal applies a similar structure using Utah’s SB132. Subject to regulatory approval, that law is expected to open the door to direct-to-load contracting between Cape Station and any Google facility built nearby, letting electrons flow more directly to the data center rather than solely through the utility grid.

"The next chapter of advanced power generation technology is being written in Utah," said Lucia Tian, director of advanced energy technologies at Google. "This project will drive meaningful economic benefit to the local community and help catalyze long-term energy cost reductions by making enhanced geothermal more affordable and accessible."

Google’s geothermal program also includes a separate deal with Ormat Technologies for up to 150 megawatts of geothermal capacity in Nevada.

Cape Station Scales Up

Fervo’s approach builds on techniques borrowed from the oil and gas industry. The company drills deep horizontal wells, creates fractures in hot rock formations, and circulates water through those fractures to capture heat and generate steam-driven electricity. Fiber-optic sensors installed in the wells provide real-time data on flow, temperature, and reservoir performance, enabling precise control over the system.

Fervo says its EGS can be deployed in locations where subsurface heat exists but natural permeability does not, which could expand the geographic range where geothermal power is viable.

The newest Google agreement moves Cape Station beyond its initial 100-megawatt phase. As we noted when Fervo went public in May, the company is permitted for up to 2 gigawatts at the site.

Fervo has also said it has cut both drilling time and cost per foot by roughly two-thirds since its earliest wells, a learning curve executives have compared to the one that reshaped shale oil and gas drilling.

Why It Matters

As AI boosts demand for data center capacity, power grids in many leading markets are constrained. That’s creating boom in on-site power generation, with about 60 data center projects announcing plans for 90 gigawatts in behind-the-meter power, according to data from Cleanview.

Most of that planned capacity will use natural gas, which has raised alarms about emissions from environmentalists and community leaders. Geothermal’s ability to deliver firm, around-the-clock, carbon-free power offers a cleaner way forward.

Google’s leadership on geothermal is part of its broader program to prioritize renewable energy and advancing promising technologies that can offer more options to power future data centers.

That could be crucial as Google seeks to build in Utah, where environmentalists led a heated campaign against a proposed data center project in Box Elder County backed by investor Kevin O’Leary, which highlighted concerns about usage of power and water resources.

Building hyperscale data centers at the speed of AI comes with a massive challenge: taking care of the thousands of tradespeople on the front lines. As project schedules compress and job sites push into remote locations, how do industry leaders build a culture that keeps workers safe, healthy, and genuinely supported?

Subscribe now

In this episode of Data Center Richness podcast, we sit down with Trey Dean, Senior Vice President of Global EHS at Vantage Data Centers. Trey shares how Vantage is prioritizing site welfare, tracking workforce mental health with actionable data, and using AI and robotics to remove workers from high-risk duties., like moving giant racks

Here is my conversation with Trey Dean.

This Data Center Download explores how building hyperscale data centers demands an unwavering commitment to worker welfare and site culture. As project schedules compress and developments move into remote regions, construction leaders must transform job sites into safe, supportive environments.

📌 Key Takeaways

• Elevate site welfare by providing climate-appropriate amenities and strong union partnerships to attract top trade talent.

• Demystify skilled trade careers for young workers through early engagement, clear wage paths, and structured mentorship.

• Measure workforce sentiment using modern data platforms to proactively support mental health and mitigate burnout on compressed schedules.

• Deploy AI visual tools and drone inspections to remove workers from high-risk hazards like live switchgear.

Rethinking Workforce Welfare and Skilled Trades Attraction

To compete effectively for skilled labor, construction sites must move beyond basic compliance and actively prioritize the daily experience of trade workers.

Creating Environment-Specific Job Sites

Job site welfare must adapt directly to harsh environmental conditions. In hot climates like Arizona, standard shade tents are insufficient; sites require climate-controlled welfare facilities that protect worker health. The same principle applies in cold environments like Quebec. Providing high-standard facilities creates a strong word-of-mouth reputation that naturally attracts skilled trades.

💡 Key Insight: A well-managed, safe job site that prioritizes worker welfare almost always finishes on time and on budget.

Demystifying Trade Careers for the Next Generation

The average age of trade workers across data center projects is well into the 40s, creating an urgent need to recruit younger talent. Simply telling 18-year-olds to skip college for the trades is not enough. Industry leaders must engage high school students early, explain complete career trajectories, and highlight real earning potential.

Operational Rigor and Industry Alignment

Scaling data center delivery requires aligning vendors, contractors, and owners around strict mission critical standards.

Elevating Safety and Quality Standards

Transitioning workers from standard commercial projects (such as retail malls) to data centers requires establishing a higher level of operational discipline. Mission critical environments demand strict protocols around high-risk activities, trade stacking (several specialists on the site at the same time), and simultaneous commissioning operations. Adopting a culture of accountability ensures that safety speeds up delivery rather than slowing it down.

💡 Key Insight: High operational discipline acts as a force multiplier, creating site-wide efficiencies that benefit all project partners.

Prioritizing Mental Health in Remote Environments

Modern data centers are increasingly built in remote locations, compounding labor logistics and taking workers away from their families for months at a time.

Supporting Traveling Workers and Work Camps

When workers travel far from home, project sites must provide quality amenities and comfortable living arrangements. Establishing well-equipped work camps with full infrastructure gives traveling crews a comfortable “home away from home.” Basic efforts like regular worker lunches, premium gear rewards, and open communication channels foster a supportive community.

Measuring Site Culture with Data

Traditional safety metrics track incidents after they happen, but modern management platforms actively measure real-time worker sentiment. Data reveals that sentiment drops during severe schedule compression, winter months, or after safety incidents. Tracking these trends provides actionable insights, enabling leaders to step in with support before burnout occurs.

💡 Key Insight: Mental health awareness must be simple, transparent, and accessible, offering both on-site support and confidential external resources.

Augmenting Safety with Robotics, AI, and Drones

Technology should never replace trade workers; its ultimate goal is to remove humans from high-risk job hazards.

Leveraging AI for Real-Time Risk Detection

Advanced camera systems powered by AI inspect active sites to flag missing PPE, exposed edges, and conflicting pedestrian or equipment pathways. Additionally, specialized robotic quadrupeds perform routine inspections in hazardous areas. These automated tools process real-time environmental conditions without creating unnecessary friction for site workers.

Aerial Drone Walkthroughs and Design Coordination

Since 2020, drone walkthroughs have evolved into essential tools for identifying structural risks and coordination issues before they cause delays on the ground. Teams capture aerial footage to flag upcoming installation conflicts and review them directly with engineers. This proactive approach keeps projects ahead of hazards and ensures long-term site safety.

💡 Key Insight: Robotics and AI act as force multipliers that protect workers, taking over dangerous tasks while leaving complex craft execution to human experts.

This text companion accompanies the video “Building a Safety Culture at Internet Speed” featuring Trey Dean of Vantage Data Centers. Watch on our YouTube channel for the full conversation and additional context.

There’s a big question at the heart of today’s tech headlines: is there an AI bubble? I sat down with someone who’s been building data centers since 1997, and I found his answer really insightful.

That’s Ron Vokoun. He’s had 37 years in construction, most of it in digital infrastructure - fiber networks in the ‘90s, the first Qwest data centers, a stretch dedicated to prefab and modular builds, and now cooling as the Senior Vice President of Data Center Cooling at SVL Inc.

Subscribe now

Ron has also become one of the most trusted voices curating data center news, first on LinkedIn, now with his Mission Critical Connections newsletter on Substack.

In this conversation, we get into why he thinks prefab isn’t just optional anymore, how rack densities have blown past anything we saw coming, and his interesting answer on whether this AI buildout is a bubble — and what’s actually keeping it in check.

Here’s my full conversation with Ron Vokoun:

This Data Center Download explores our podcast conversation with industry veteran Ron Vokoun, who shares his insights on modular prefabrication, the rapid shift to liquid cooling, and how physical constraints are helping prevent an AI market crash.

📌 Key Takeaways

• Adopt Off-Site Prefabrication: Move mechanical and electrical builds off-site to bypass severe job-site labor shortages and accelerate build speed.

• Plan for Hybrid Rack Densities: Prepare facility designs for an average of 200–250 kW per rack rather than assuming every rack will hit megawatt-scale densities.

• Secure Vendor Capacity Early: Protect supply chains by building long-term strategic partnerships with contractors and equipment vendors rather than commoditizing them.

• Monitor the Three Market Bottlenecks: Factor power availability, labor constraints, and community sentiment into site selection. All three can slow your growth.

• Commit to Daily Industry Reading: Dedicate time every day to studying technological shifts to maintain domain expertise and deliver value to clients.

37 Years on the Digital Infrastructure Frontier

Ron Vokoun’s career reflects the broader evolution of the digital infrastructure sector. Starting in environmental construction, his path pivoted to telecommunications during the 1990s fiber boom, leading to early facility buildouts for metropolitan fiber loops and nationwide networks.

From Local Fiber Loops to Early Data Centers

Entering the tech sector in 1992, Vokoun managed local fiber-optic loop deployments in Los Angeles before joining SP Telecom (later Qwest Communications). After the 2001 telecom downturn, Vokoun transitioned into general contracting, overseeing complex international builds in Tokyo and across the US.)

💡 Key Insight: Long-term success in digital infrastructure relies on maintaining strong industry relationships and adapting as technologies evolve across market cycles.

Solving the Construction Labor Shortage: Modular Prefabrication

Severe shortages in skilled job-site labor have made traditional field construction unsustainable for meeting current AI deployment schedules. Modular prefabrication and off-site assembly have shifted from optional efficiency tools to operational necessities.

Shifting Work Off-Site to Accelerate Speed to Market

Off-site prefabrication allows operators to tap into alternative regional labor pools operating in controlled plant environments. Building components in parallel with site work drastically reduces field congestion and accelerates delivery timelines.

• Parallel Processing: While foundation and shell work occur on-site, electrical rooms, mechanical skids, and cooling modules are fabricated off-site.

• Deployment: Modular units are transported to the facility and either wheeled through access doors or plugged directly into the exterior structure.

💡 Key Insight: Operators who fail to deeply integrate prefabricated modular strategies will struggle to keep pace with demand during ongoing labor shortages.

The Liquid Cooling Paradigm Shift

The rapid adoption of generative AI has transformed cooling requirements almost overnight. Legacy air-cooled architectures are rapidly giving way to direct-to-chip liquid cooling deployments.

Training vs. Inference Load Dynamics

While headline-grabbing announcements highlight theoretical rack densities reaching 600 kW or 1 megawatt, real-world facilities will operate across a spectrum of workload requirements.

Historical data centers in the mid-2000s often deployed 350 kW across an entire 3,000-square-foot white space. Moving forward, typical next-generation high-density data centers are expected to settle into an average of 200 kW to 250 kW per rack.

💡 Key Insight: Not every application requires megawatt-level single-rack density; future facilities will rely on hybrid cooling architectures optimized for specific workload profiles.

Market Dynamics: Why the AI Boom Is Not a Bubble

Jevons Paradox (an economic principle stating that increased efficiency in sing a resource leads to increased overall consumption) explains why compute demand continues to accelerate. As AI tools become cheaper and more efficient, new applications emerge, creating a self-reinforcing growth flywheel.

The Three Gatekeepers Throttling Overbuilding

Unlike the dot-com era, where speculative capital led to empty facilities, the current expansion is bounded by physical speed brakes that enforce market discipline:

1. Power Grid Limits: Long utility interconnection queues and required upfront financial deposits filter out speculative “dreamawatts” and “bragawatts.”

2. Skilled Labor Shortages: Physical capacity constraints prevent operators from building faster than the market can absorb.

3. Community & Resident Sentiment: Increasing local opposition creates regulatory and zoning checkpoints that slow speculative development.

💡 Key Insight: Infrastructure bottlenecks act as natural guardrails, preventing the unrestrained overbuilding that historically caused market crashes.

Content Curation and Thought Leadership

Transitioning from operations to business development highlights the importance of continuous learning. Adopting a structured daily reading routine builds authentic domain knowledge that can be shared across professional networks.

The Value of Disciplined Curation

Following advice from an industry futurist in 2006, spending the first hour of every workday reading industry developments builds deep domain expertise. Sharing curated, high-value insights on platforms like LinkedIn and Substack (such as the Mission Critical Connections newsletter) establishes long-term credibility.

💡 Key Insight: Share content only when it provides genuine value—curating high-quality insights builds a far more engaged audience than posting for frequency alone.

Taking Action

To adapt to current digital infrastructure trends, industry professionals should consider these immediate steps:

• Audit Construction Workflows: Identify mechanical and electrical systems in upcoming builds that can be transitioned from field assembly to off-site prefabrication.

• Evaluate Fluid Delivery Systems: Review facility plumbing architectures and define clear demarcation points for liquid cooling handoffs between colocation providers and tenants.

• Lock in Vendor Capacity: Establish strategic long-term supply agreements with key manufacturers and trade contractors to reserve production slots.

• Implement a Daily Reading Habit: Reserve time every morning to review technical trade publications, power grid policies, and cooling innovations.

This text companion accompanies the video Is the Data Center Boom Built to Last? featuring Ron Vokoun, Senior Vice President of Data Center Cooling at SVL Data Center Cooling. Watch on our YouTube channel for the full conversation and additional context.

The leading tech companies have a bold vision for the future AI infrastructure. But it requires more power than today’s data centers are designed to handle.

The solution is a gradual shift to using 800 volt DC power distribution in AI data centers. It’s a major undertaking that requires upgrades of the most mission-critical components, as well as the cooperation of an entire ecosystem of vendors.

NVIDIA, Google and Microsoft have teamed with the Open Compute Project (OCP) to provide an update on the 800 VDC transition. At the OCP APAC Summit in Taipei, the companies said they will work together to create an “open, standardized power architecture for next-generation AI data centers and to make sure the entire industry can adopt it safely, at scale.”

In a key step, the companies shared the Solid-State Transformer (SST) Specification v0.3 developed through collaboration among Microsoft, Google, NVIDIA, and the broader ecosystem. The SST is emerging as a critical piece of the transition to 800 VDC power distribution, a simplified pathway from grid to rack.

The end goal isn’t a single design, but a common set of interfaces and system requirements that lets suppliers build interoperable products. The collaboration was detailed in a joint blog post by the three companies:

We are aligning on system-level performance requirements, including power quality and power smoothing standards, and end-to-end system interfaces. We are publishing these requirements openly so that equipment manufacturers can develop interoperable products with confidence, without having to navigate conflicting proprietary specifications from different customers. We are engaging in parallel with UL Solutions, NFPA, IEEE, and IEC to advance the safety certifications and regulatory frameworks that are prerequisites for global deployment.

Addressing Power and Density Constraints

Today’s racks in AI factories rely on 54 VDC power distribution, where bulky copper busbars shuttle electricity from rack-mounted power shelves to compute trays. As racks exceed 200 kilowatts, this approach begins to hit physical limits, facing limitations in copper volume, thermal loss, and space efficiency.

That moment is already on the roadmap. The shift to 800 VDC is expected to get underway in earnest with NVIDIA’s Vera Rubin Ultra GPU platform, which is expected to arrive in 2027 and support power densities of up to 600 kilowatts per rack.

Moving to 800 VDC infrastructure from traditional 415 or 480 VAC three-phase systems enables higher densities, as well as a 5 percent improvement in power efficiency, and cooling benefits as power infrastructure is moved outside the rack.

“800 VDC unlocks the compute performance and power density required for AI at scale,” said Vladimir Troy, vice president of data center infrastructure at NVIDIA. “Through OCP, NVIDIA is working with more than 80 ecosystem companies to give AI factories a practical path forward — not just a future vision.”

New Architectures Emerge

The Open Compute Project has been the data center industry’s leading venture for collaboration between hyperscale operators. OCP has called on its members to raise their game to confront how AI is disrupting digital infrastructure, driving change in data center power, cooling and networking.

“Common 800 VDC interfaces can help the industry scale AI infrastructure while protecting the flexibility operators need in real deployments,” said Tom Garvens, Vice President of Data Center Technology and Systems, Google. “By working through OCP, we can align requirements early, reduce fragmentation and give suppliers a clearer path to build interoperable products.”

Hyperscalers have also been working on the transition to higher voltages. Microsoft, Meta and Google have collaborated on Mt Diablo, an OCP design for a “sidecar” that will shift power conversion into a separate rack, laying the groundwork for denser and more configurable server racks.

Power sidecars will serve as the entry point into the 800 VDC transition. But the companies emphasized that 800 VDC isn’t a replacement for existing AC infrastructure, but an additional deployment option that can coexist with existing AC systems and enables a gradual transition as AI power density continues to increase.

Read more

Anthropic continues to expand its options to add more capacity for more data crunching. Today the frontier AI lab said it is teaming with global investors Macquarie Asset Management and GIC to launch a dedicated data center platform to support Anthropic’s growth.

The three companies announced a strategic partnership to establish Theseus Infrastructure, a new platform that will develop custom infrastructure for Anthropic, which will lease the capacity under long-term agreements.

The partners did not indicate the number of planned sites or a dollar figure for planned investment, but said it would focus on projects in the U.S.

A Diversified Ecosystem of Compute Providers

Frontier AI labs have spent the past two years leasing their way to more compute. Anthropic’s latest move suggests leasing alone is no longer enough.

Today’s announcement adds to a massive expansion of Anthropic’s infrastructure, which includes:

• Anthropic has committed more than $100 billion to AWS over the next decade, securing up to 5 gigawatts of Trainium capacity through the Project Rainier cluster, where it already runs more than a million Trainium2 chips

• Anthropic’s agreements with Google Cloud have grown to cover up to a million TPUs and multiple gigawatts of next-generation capacity developed with Broadcom.

• In May, Anthropic announced a partnership with SpaceXAI to use all the compute capacity at its Colossus 1 data center in Memphis, which was built by xAI. The deal provides Anthropic with access to 220,000 NVIDIA GPUs and more than 300 megawatts of power.

• Last year Anthropic said it would invest $50 billion in data center capacity in the U.S. partnering with Fluidstack on custom data centers in Texas and New York, “with more sites to come.”

What Is Theseus Infrastructure?

The new partnership taps Macquarie’s global expertise developing, financing, and operating large-scale data centers, and GIC’s experience investing in infrastructure projects around the world.

The arrangement positions Anthropic in a role closer to that of a development partner than a traditional cloud customer, working with Macquarie and GIC to identify and build sites rather than simply signing capacity agreements.

The announcement also included a nod to local communities, who are increasingly important stakeholders in data center development approvals. The partners said their projects would create thousands of construction jobs and permanent operational roles in the communities where they locate.

Anthropic also affirmed that it will “pay its own way” on power, consistent with a commitment it made earlier this year.

Each week I curate 5 links from the data center sector that I find particularly interesting, with my commentary on why they merit your attention.

Sound interesting? Subscribe and get them every Saturday.

Subscribe now

NOTE: The “5 Links” Saturday newsletter will be taking a break for the balance of August. I’ll be back with more weekly insights in the first week of September.

5 Notable Data Center Links

• DCF Trends Summit: Execution Took The Stage - All the hype about massive AI data center projects has given way to the harder work of delivering all the promised physical infrastructure. I spent much of this week at the Data Center Frontier Trends Summit in Reston, Virginia, where the Big Topic was the shift from announcements to execution. “Power, entitled land, transmission, equipment, labor, capital, permitting and community acceptance determine how much of the announced pipeline can become energized capacity,” DCF Editor Matt Vincent writes in his roundup of the event. In coming weeks, I’ll have more coverage of the discussions at DCF Trends, but this overview is a good summary of the event’s major themes.

• AI Compute Crunch for the Hyperscalers? - We’re starting to see more signs of capacity constraints being felt in major AI computing operations. Recent earnings featured extensive discussion about capacity constraints at hyperscale operators, and this week reports emerged that AWS is scrutinizing the cloud usage of its own engineers. “In May, the company reportedly met with engineers and told them to reduce CPU waste to ensure AWS has enough CPU capacity to meet customer demand,” writes Tom’s Hardware, summarizing a report from The Information (subscription needed). I’ve been writing about the data center supply crunch for the past year, during which vacancy rates have continued to head closer to zero in leading markets. We’ve seen leading AI shops tighten usage limits and shift to credit-based billing for heavy users. It seems likely that the compute crunch will get worse before it gets better. It’s hard to assess how this will play out, but investment house Apollo is among those contemplating a compute-constrained world and its implications.

• Data Center Operators: Texas Audit Can Create a Filter for Quality - In a bit of a surprise, Texas-based data center operators are welcoming Gov. Greg Abbott’s decision to pause data center power approvals to conduct an audit of applicants, citing rampant speculation in the ERCOT grid’s interconnection queue, which now has applications for 475 gigawatts (GWs) of power, more than 5 times the peak demand on the ERCOT grid. “There was a ton of speculators in the queue that were going to get power,” Compass Datacenters CEO Chris Crosby said in an appearance on CNBC’s PowerLunch. “That was a problem because their whole intent was to sell. We had gigawatts presented to us by people who had no intention of putting a shovel in the ground.” Stream Data Centers President Michael Lahoud echoed that sentiment. “Our citizens deserve to be able to trust that these projects and developers are real, responsible and rational,” Lahoud said in a statement. “This action helps Texans ensure they’re prioritizing investments from developers that share their goals, protect their interests and are committed to being great long-term neighbors.” Crosby emphasized the opportunity for responsible developers to build alongside their communities. “Raise the bar,” said Crosby. “Ask for more.”

• Optical Networking Startup Lumilens Launches With $900M Funding - Investors are intensely focused on the potential for optical networking to accelerate AI compute workloads. This week startup Lumilens came out of stealth with $900 million in backing, including a $700 million in a new Series C round. Lumilens makes chips that make it possible to replace copper interconnects with faster fiber-optic cabling. That includes chips for co-packaged optics (CPO) that can be integrated directly into GPUs, as well as near-packaged optics (NPO) chips that reside on the same circuit board. NPO is seen as a likely interim step to broad CPO adoption. Here on Substack, Semiecosystem has coverage of Lumilens and several other networking startups, and Damnang has a good explainer on CPO and why it is attracting so much attention.

• Amazon Data Center Buildout Focuses on Massive Gas Projects - Data center developers have announced nearly 60 behind-the-meter gas power projects since the beginning of 2025 with a combined capacity of 90 GW, according to Cleanview. This week AWS made headlines for its interest in two huge developments that could bring up to 12 gigawatts of gas-fired power online. “Amazon is backing what could be the largest natural gas power plant ever built in the United States,” writes Michael Thomas at the Distilled Substack, confirming the company’s plans for a data center at a site in Pecos County, Texas which could feature 7.6 gigawatts of natural gas generation. This week AWS also confirmed that it is exploring the development of a data center at a natural gas plant in Homer City, Pa. as part of its planned $20 billion investment in Pennsylvania. “While we don’t have a definitive agreement in place, we are in discussions with Homer City Redevelopment,” an Amazon spokesman said this week. A 4.5 gigawatt natural gas plant is under construction on the 3,200-acre Homer City site, DCD reports.

• A Deeper Dive into “No Data Centers in My Backyard” - I’ve been tracking the tidal wave of headlines about community pushback against data centers for years. Substack essayist Jasmine Sun took two weeks to visit hotspots for data center controversy across the Midwest, and the resulting piece is a worthwhile read that adds genuine depth to the subject, with more balance and context than usually found in critiques of the industry. This is a long read (6,000 words) but if you’re genuinely confused about the intensity of the opposition to data center development, this will fill in the blanks. “If people are this angry at a building sitting and paying taxes on unused land, something fundamental has gone very wrong,” Sun writes, adding that “there is no compelling pro-data-center, pro-AI faction. … It is striking how little the AI industry understands its own unpopularity.”

If you found 5 Notable Links useful, please subscribe so you can get it each week!

In Case You Missed It …

Do higher voltages represent higher risk? It’s a critical question as NVIDIA and its largest cloud customers push toward 800 volt direct current (800 VDC) power architectures.

Schneider Electric has published one of the first studies examining arc flash risk in 800 VDC data centers, concluding that the hazard can be managed, and in many configurations may be comparable to the AC systems already deployed across the industry.

But the study also surfaces a significant challenge. The safety standards written for AC systems and battery-based DC systems don’t fully apply to the 800 VDC architectures now being designed for AI factories. Engineers are left adapting existing methods rather than working from an established rulebook.

“800 VDC power distribution represents a significant shift in data center design, but it also introduces safety considerations that need to be studied extensively,” said Manish Kumar, EVP of Secure Power and Data Centers at Schneider Electric. “Our work with some of the world’s leading hyperscalers provides engineers and safety professionals with one of the first practical frameworks for evaluating arc flash risks, providing a structured approach to understanding fault behavior, establishing safe work practices, and designing effective protection schemes.”

The Schneider study, conducted with power software specialist ETAP, is part of a larger industry effort to ensure that as power densities increase, worker safety issues advance alongside technical innovation.

Subscribe now

What is Arc Flash?

Arc flash incidents represent a significant threat to worker safety. An arc flash is an electrical explosion that generates intense heat that can reach 35,000 degrees F, which can damage and even melt electrical equipment. An arc flash can be a dramatic event in a data center, where the power infrastructure handles enormous amounts of electricity.

Arc flash incidents typically occur during equipment failure, improper maintenance, or human error around energized systems. Studies show that adherence to safety standards such as NFPA 70E significantly reduces the frequency and severity of arc flash incidents.

Data compiled by the Electrical Safety Foundation International (ESFI) shows 5,180 non-fatal U.S. workplace electrical injuries involving lost work time over 2023–2024 alone. Arc flash incidents represent a subset of those injuries but are among the most severe, frequently causing catastrophic burns, blast trauma, hearing damage, and permanent disability.

In addition to the human safety issues, arc flashes have led to expensive outages in data centers, most notably a 2009 incident at Fisher Plaza in Seattle, which knocked a major online payment gateway offline, slowing global e-commerce.

The Shift to 800 VDC

Moving to 800 VDC infrastructure from traditional 415 or 480 VAC three-phase systems will enable higher power densities at the rack level, enabling more efficient and powerful AI Factories. But it also prompts new design considerations for data center power distribution.

Read more

Each week I curate 5 links from the data center sector that I find particularly interesting, with my commentary on why they merit your attention.

Sound interesting? Subscribe and get them every Saturday.

Subscribe now

5 Notable Data Center Links

• AWS Boosts CapEx, Citing Cloud Precedent - Big CapEx spending up front, followed by customers, then profitability. It’s a model we’ve seen before with cloud computing, as Amazon CEO Andy Jassy reminded us this week as the company reported strong earnings. Amazon is boosting its 2026 capex by $20 billion to $220 billion, and says it has so much AI business that it expects to be capacity constrained through 2028. AWS is the largest cloud player, building as fast as it can, and it simply cannot keep pace with demand and doesn’t expect to do so anytime soon. Jassy now believes AWS could “very possibly” reach $1 trillion a year in revenue. Wow. I’ve always believed that cloud offers perhaps the best template for how AI may develop as a business, and Jassy drew direct comparisons in the earnings call, expressing high confidence that the capex math on AI is sound. This tracks for me on the demand side … I think most of the risk is in the cost side of the equation and how that evolves. There’s a good summary of Jassy’s cloud commentary in Jamin Ball’s Clouded Judgement here on Substack.

• $100 Billion Data Center Planned at DOE Kentucky Site - Will the largest concentrations of AI infrastructure be built on land owned by the Department of Energy? This week a coalition led by Brookfield and NextEra Energy unveiled plans to build a $100 billion, privately funded AI campus at the DOE’s Paducah Site in Western Kentucky, backed by up to 4.6 GW of dedicated new power generation. Brookfield will develop the data center, while NextEra will own and operate the on-site generation. This follows the same model as a 10 gigawatt project being planned at the Portsmouth Site in Piketon, Ohio, where SoftBank and utility AEP Ohio will partner on 10 gigawatts of on-site power to support a data center, with OpenAI rumored to be the likely tenant. One rationale for the interest in these sites is the potential for streamlined permitting and buildout on the DOE sites. The initial RFI seeking bids for these DOE sites was titled “Accelerating Speed to Power / Winning the AI Race.”

• Newspaper Printing Plant to Become Data Center - The new economy often rises in the footprint of the old. A good example is the growing number of newspaper printing plants being redeveloped as data centers. The latest project arrived this week, as real estate firm Legacy Investing agreed to purchase the printing facility for the Minneapolis Star Tribune, with plans to develop the building as a data center. Newspaper printing plants are good candidates for data center conversions, as they typically come with heavy floor loading, high ceilings, large open floor plates, abundant electrical service, and loading docks. Previous examples include QTS Chicago (Sun Times printing plant), Iron Mountain New Jersey (former NY Times printing plant), Digital Realty TOR1 (once a Toronto Star printing plant) and Patmos in Kansas City (former KC Star printing facility).

• Can Data Centers Build Their Own Transmission Lines? - Amazon recently asked Virginia regulators to allow hyperscale data center operators to voluntarily fund transmission infrastructure serving their campuses. The proposal would be one way to ensure that data centers cover the costs of the transmission infrastructure for large new campuses, rather than ratepayers. Hyperscalers building their own power transmission seems like a big step, but there’s an interesting precedent in subsea fiber, where Google, Meta and Amazon have all built dedicated subsea cables to connect their global data center campuses. One potential motivation might be faster deployment of new transmission lines, which takes 7 to 12 years on average. It is one of the toughest bottlenecks in the ambitious AI buildout, and it will be interesting to see if hyperscalers can navigate the regulatory maze required to make this strategy happen.

• 350-Megawatt Microgrid for Data Center in Ohio - Microgrids are one of the technologies we’ve been tracking for years that is finally seeing traction as data centers embrace on-site power generation. This week Veolia said it has been selected to build and maintain a 350-megawatt microgrid to power an AI data center campus in New Albany, Ohio, one of the sector’s largest cloud clusters. The installation will combine on-site generation with large-scale battery storage. “Our work on this project shows how Veolia can help developers accelerate speed-to-power while maintaining the reliability required for mission-critical AI operations, all while protecting the community,” said Karin Hamel, CEO of Veolia Energy, North America.

If you found 5 Notable Links useful, please subscribe so you can get it each week!

In Case You Missed It …

Among emerging data center markets, Kansas City has had a fairly low profile. But that’s about to change.

The Kansas City market got a major vote of confidence this week from Digital Realty, one of the largest and most experienced players in data center real estate. This week the company revealed that it has acquired 1,440 acres of powered land in the Kansas City metro where it plans to develop up to 2 gigawatts (GW) of power capacity.

Subscribe now

The new campus in De Soto, Kansas could become one of the largest in the Digital Realty portfolio, rivaling its huge footprints in Northern Virginia and Dallas, which have been assembled across multiple properties over several decades.

“Our belief is that this is going to quickly become the seventh-largest data center market in the U.S.,” said Greg Wright, the Chief Investment Officer of Digital Realty, on the company’s earnings call. “We’re seeing very strong customer demand here. It’s really becoming a Midwest hub for both AI and cloud workloads.”

The arrival of Digital Realty follows a series of expansions by hyperscale players, as well as neoclouds focused on AI computing. Those projects build on an existing base of colocation and connectivity service providers, laying the groundwork for a larger growth story.

A Pathway to Power and an Enterprise Base

We’ve entered an era where data center campuses have become massive, and secondary regional markets can grow quickly. The new data center geography is deeply influenced by access to power.

Regions with available utility capacity have quickly moved to the fore in data center site selection, creating a higher profile for markets like West Texas, Indiana, Pennsylvania, and Reno.

The Kansas City metro market has power, but it also is an established market for business computing, offering growth avenues for both hyperscale and enterprise customers.

Most analysts tracking data center real estate have not positioned Kansas City as a huge growth market. The exception is datacenterHawk, which projects that Kansas City will become the seventh-largest market in North America. Cushman & Wakefield sees Kansas City as a contender on list of its emerging markets in the Americas, trailing locations like Austin, Iowa, Pennsylvania, Reno and Minneapolis.

Digital Realty secured 600 megawatts (MW) of utility power for 2028 delivery, with the ability to ramp up to 2 gigawatts over time.

Read more

A power grid event last week in Northern Virginia’s data center cluster has focused fresh attention on the potential risks of the large energy usage of AI infrastructure.

A transmission line fault Wednesday in Ashburn, Virginia, prompted hyperscale facilities to automatically transfer to backup power, dropping more than 3 gigawatts (GW) of demand off the PJM grid in seconds.

A day earlier, a leading data center developer announced plans to roll out a solution to this exact problem.

Subscribe now

The two events highlight the challenges posed by the rapid growth of digital infrastructure, along with the data center sector’s race to address them through innovation.

Crusoe announced plans to deploy 5 GW of ON.energy AI UPS systems across multiple hyperscale AI campuses, which would be one of the largest deployments yet of battery-backed UPS infrastructure for AI data centers.

The ON.energy system functions as a facility-level UPS, a departure from traditional data center power architecture. It keeps the data center online through grid disturbances while using batteries and power electronics to prevent AI workload fluctuations from propagating back onto the grid, presenting a steadier load to the utility.

This ability to “ride through” power instability can address events like last week’s problems in Virginia. It is also becoming part of the policy for grid access in Texas, one of the focal points for development of new AI data centers.

What Happened in Ashburn?

Dominion Energy said the July 22 event began when a transmission line in Ashburn experienced a fault and automatically went out of service.

“Because of this normal system operation, data centers elected to transfer load off the system,” Dominion spokesperson Jeremy Slayton told Data Center Knowledge. “To be clear, no load was shed and Dominion Energy did not disconnect area data centers from the grid; the data centers’ own control systems transferred them to backup power for a very short period of time.”

The disconnections happened in stages on the PJM Interconnection regional grid, which reported more than 3 GW going offline, or about 3% of the grid's total demand at the time, according to PJM operations data.

The data centers operated as designed, acting to protect IT equipment and treat a voltage event as a reason to disconnect. The challenge is that response happened simultaneously across a cluster of large campuses in a single market. The result was a sudden loss of gigawatts of demand at the precise moment operators are working to restore stability.

The voltage disturbance was felt from Washington to Chicago, with customers reporting flickering lights and noise from appliances. It’s not the first time this has happened. Virginia saw about 1.5 GW disconnect during a similar transmission fault in July 2024.

Moving the UPS Upstream

Conventional uninterruptible power supply systems keep IT systems running until generators come online, and are typically deployed downstream to protect individual electrical distribution blocks or data halls. The ON.energy AI UPS is a medium-voltage battery system installed inline between the electrical supply and the data center, buffering the entire facility rather than individual IT loads.

Read more

Each week I curate 5 links from the data center sector that I find particularly interesting, with my commentary on why they merit your attention.

Sound interesting? Subscribe and get them every Saturday.

Subscribe now

UPCOMING EVENT: DCF TRENDS: I will be part of the opening keynote panel at the Data Center Frontier Trends Summit, which will be Aug. 4-6 in Reston, Va. On August 4, I’m joining EdgeCore Digital Infrastructure CEO Lee Kestler and DCF Editor in Chief Matt Vincent to discuss key trends in data center development, and the landscape in Northern Virginia - where is a LOT happening, as you’ll see in this week’s links.
Check out the DCF Trends Summit and event schedule.

5 Notable Data Center Links

• Loudoun County Supervisors Explore Data Center Moratorium - The home of Data Center Alley is considering a moratorium on data center development. In a 6-1 vote on Wednesday, the Loudoun County Board of Supervisors directed county staff to prepare options by Sept. 15 to pause data center applications, while also asking the county attorney to review if a pause is legal. Northern Virginia’s Loudoun County is home to Data Center Alley in Ashburn, the world’s largest concentration of digital infrastructure. “I think it's time that we just press pause,” said Supervisor Julie Briskman, who proposed the motion. “It's what the community has been kind of begging us to do.”
  

  The supervisors passed the measure even though county attorney Leo Rogers said a moratorium would likely not be legal without a specific rationale. Briskman noted that three other counties in Virginia have undertaken pauses in data center approvals. “I’m looking to test the waters, because it could be argued that we are in a crisis,” said Briskman. The measure passed near the close of a lightly-attended 5-hour meeting. In a statement to the Washington Business Journal, the Data Center Coalition’s Nicole Riley said a pause on development “would send a signal that the area is closed for business, both for data centers and for other significant economic development projects.” The supervisors will weigh next steps at their Sept. 15 meeting.
  
  The board also is seeking options on regulating the operations of diesel backup generators, citing resident complaints about noise and emissions. Data centers in Loudoun operate about 4,700 emergency generators. “We need the ability to address existing backup generators,” said Supervisor Michael Turner, who noted the availability of SCR (selective catalytic reduction) systems and diesel particulate filters that could limit health impacts of older generators. Possible measures also included setting limits on genset usage, including “requiring pauses during generator operation in emergencies,” according to board agenda documents.
  

• As Deal Closes, Aligned Will Get $5 Billion in Growth Capital - Investor group The Artificial Intelligence Infrastructure Partnership (AIP) this week completed its $40 billion acquisition of Aligned Data Centers from Macquarie Asset Management. Aligned operates a portfolio spanning 51 campuses and more than 6.4GW of operational and planned capacity. Chief Executive Officer Andrew Schaap and the existing management team will continue to lead Aligned. AIP (which includes MGX and BlackRock’s GIP) will commit an additional $5 billion of growth capital to fund Aligned’s continued expansion. The Aligned team has been one of the most efficient players in creating new hyperscale campuses. The precedent worth noting is QTS Data Centers, which has accelerated its development pace with the backing of Blackstone.

• TECfusions to Go Public Through SPAC Merger - Data center firm TECfusions plans to go public via a merger with special purpose acquisition corporation (SPAC) Apex Treasury Corp, the companies announced. Once it completes the transaction, TECfusions will join the Nasdaq and trade under the symbol TECF. TECfusions focuses on the adaptive reuse of industrial properties, and has data center projects underway in three locations - Clarksville, Va.; Tucson, Ariz.; and New Kensington, Pa. The deal values TECfusions at around $4 billion and a post-transaction value of $4.2 billion. The deal has been approved by the boards of both firms and is expected to close in Q4 2026. See the Apex Treasury web site for more details.
  

• Transmission Fault in Northern Virginia Triggers 3 GW Load Drop - Here’s an event you’ll hear much more about (including from us). A transmission line fault Wednesday in Ashburn prompted hyperscale facilities to automatically transfer to backup power, causing more than 3 GW of demand to drop off the PJM grid in seconds. Dominion Energy said the event began when a transmission line in Ashburn experienced a fault and automatically went out of service. “Because of this normal system operation, data centers elected to transfer load off the system,” Dominion spokesperson Jeremy Slayton told DCK. “To be clear, no load was shed and Dominion Energy did not disconnect area data centers from the grid; the data centers’ own control systems transferred them to backup power for a very short period of time.” The grid disturbance was large enough - about 3 percent of the PJM grid - that it took nearly 10 minutes to fully stabilize the PJM system, DCD reported. This is a real-world manifestation of a problem driven by AI growth that has captured the focus of both the data center industry and utilities. This week Data Center Richness will have coverage of one solution being applied at scale, part of a broad effort to address these risks.
  

• Synergy: Despite Headwinds, Data Center Growth Will Continue Apace - Despite power constraints and public opinion obstacles, U.S. data center capacity will double in the next three years, according to Synergy Research Group (now part of TechInsights). Growth will be led by an aggressive build-out of hyperscale operators’ owned data centers, which will double in operational capacity within just two years. "It is indisputable that constrained availability of power and rising local concerns over data centers are crimping many new plans for data centers,” said John Dinsdale, a Chief Analyst at Synergy Research Group. “But it is also clear that data center developers will continue to find ways around those issues and that booming demand will continue to drive aggressive capacity growth.”

If you found 5 Notable Links useful, please subscribe so you can get it each week!

In Case You Missed It …

The double-wide racks are going mainstream, offering another sign of how AI is altering the shape of the data center.

This week AMD officially launched its Helios Rackscale Solution, an extra-wide rack bringing together AMD’s GPU, CPU and networking technologies in a unified platform for AI computing. At 7,000 pounds and power densities of up to 246 kW per rack, Helios will demand attention as it enters hyperscale data centers.

Helios is based on the Open Compute Project’s Open Rack Wide standard, which features rack frames of 47.25 inches wide and 94 inches in height - about 4 feet wide by nearly 8 feet high.

That’s a big change from the traditional 19-inch racks used in most enterprise and colocation data centers, as well as the 21-inch wide equipment slot for the Open Rack v3.

To help customers with the adoption of the larger racks, AMD and Schneider Electric have teamed on a reference design for Helios that can support densities of up to 246 kW per rack enclosure. This also results in a 7,000-pound rack.

There’s a lot here to talk about. This story covers:

• An overview of Helios;

• The shift to larger rack designs, and why it is happening;

• A look at the design issues addressed in the reference design via a Q&A with a Schneider Electric executive.

What is Helios?

The double-wide racks are part of AMD’s initiative to compete with NVIDIA in the massive, fast-growing market for AI computing power. Helios represents AMD’s response to the unified AI platform approach that has given NVIDIA a commanding market position in AI computing.

AI systems are driving big changes in data center design, from the grid to the chip, as operators seek to push the boundaries of new hardware and models.

“Frontier AI is raising the bar for what infrastructure needs to do,” AMD CEO Lisa Su said Thursday in her keynote at Advancing AI. “It takes more than a single chip or a single server. You actually have to design the rack as one system.”

AMD jointly developed the racks with Meta, and showed them off last fall at the Open Compute Summit. Here’s our reporting on that event:

The official launch was Thursday at the company’s Advancing AI 2026 conference, with a customer list including OpenAI, Anthropic, Meta, Microsoft and Oracle.

The Helios rackscale solution brings together AMD Instinct MI455X GPUs, 6th Gen AMD EPYC server CPUs, AMD Pensando networking and AMD ROCm software into a unified platform built on open standards.

AMD says Helios can deliver exaflop-class AI performance in a unified architecture optimized for large-scale inference and frontier-model development. Each rack integrates 18 open rack wide-aligned 4-GPU compute trays (72 total GPUs).

“AI infrastructure is rapidly moving to full-scale AI factories, and that requires compute, networking, power and cooling to be designed together from the start,” said Forrest Norrod, executive vice president and general manager, Data Center Solutions Business Group, AMD. “AMD Helios provides an open, rack-scale architecture built to deliver the performance, efficiency and flexibility required for next-generation AI workloads.”

Why Larger Racks?

The larger racks reflect the latency requirements of new AI strategies, especially “mixture of experts” architectures that reduce the computation required for larger LLMs, but are more demanding on the network.

Larger rack designs are guided by the need for AI apps to “scale up” rather than “scale out.” A quick and broad description:

• Scaling up means adding more resources to a single machine - making one node server or domain powerful. In this case, 72 AMD GPUs operate in the same high-speed fabric.

• Scaling out means adding more machines (nodes) to share the workload. Instead of one giant server, you use many smaller ones working together.

At Meta they refer to the new hardware as BFRs - short for something similar to “Big Freaking Racks.” They think racks could get even better, as seen in this slide from their OCP presentation.

The shift to wider racks is not universal among AI specialists. Notably, NVIDIA is sticking with a single-rack spec for its next-generation Kyber rack for the Vera Rubin Ultra platform.

Rubin uses NVIDIA NVLink technology for ultra-high-speed networking optimized for GPU-to-GPU traffic, while the AMD design relies on Ethernet, which is a more versatile open standard but has higher latency.

How to Cool Helios

Server rack design matters, as they are the key building blocks in data center design, guiding the size of rows, containment systems, and data hall layout. New form factors can be disruptive.

Read more

The business impact of AI is about far more than algorithms and models. Demand for AI capacity is driving a boom in manufacturing across the U.S., which we’ve been tracking here at Data Center Richness.

The latest huge project comes from Taiwan-based electronics manufacturer Wistron, which has invested $700 million in a state-of-the-art factory in Fort Worth, Texas to make the latest GPU Grace Blackwell and Vera Rubin chips for NVIDIA AI systems.

Subscribe now

The 324,000 square foot D1 facility is Wistron’s first U.S.-based manufacturing facility, and was unveiled Tuesday by Wistron Chairman Simon Lin and NVIDIA Founder and CEO Jensen Huang.

“Together, NVIDIA and Wistron are restoring US advanced manufacturing capacity in Texas, creating skilled jobs and strengthening America’s AI supply chain,” said Huang.

Lin noted that the D1 facility is producing the NVIDIA GB300 Grace Blackwell Superchip, and will soon produce the NVIDIA Vera Rubin Superchip as well.

The Fort Worth site “is not typical manufacturing,” Lin said. “It is new, very comprehensive, and high-tech … In the next couple of years, this location will be one of the most important, as we build AI infrastructure here in the United States.”

Wistron is an original design manufacturer (ODM) which has built systems for many of the world’s largest tech companies, including NVIDIA, Apple, Microsoft, Dell and HP. The company shared a video for the Fort Worth opening:

Delta Expands in Fremont

Wistron wasn’t the only Taiwan-based ODM announcing a U.S. expansion this week. On Tuesday Delta Electronics said it is buying an additional building to expand its U.S. campus in Fremont in the East Bay, which includes R&D, lab testing for its data center infrastructure and smart energy solutions.

"Fremont and the Bay Area have played an important role in Delta’s U.S. growth story, and this investment represents our continued belief in this region’s talent, customers and innovation ecosystem," said Austin Tseng, President of Delta Electronics Americas. “We're proud to deepen our roots here.

Delta says the 55,000 square foot building located at 47200 Bayside Parkway will serve as a live demonstration environment where customers can experience Delta's full grid-to-chip capabilities, from smart energy infrastructure to AI data center solutions.

Earlier this year, Delta unveiled a grid-interactive microgrid at its Detroit-area facility, developed in partnership with DTE Energy, creating a testbed for demonstrating real energy infrastructure under live conditions and advancing grid resilience for commercial and data center customers.

Additional Resources

These announcements are only the latest in a growing expansion of U.S.-based manufacturing to support the growth of AI. We’ve tracked projects from Micron, Corning, Vertiv, Schneider Electric, Eaton, Crusoe and DHL to support data center and chipmaking capacity.

Here’s additional reading:

Google is researching new ways to tap idle compute capacity spread across its global data center network to train AI models.

The approach, detailed in a paper from Google DeepMind and Google Research, introduces an architecture called Decoupled DiLoCo (Distributed Low-Communication) that allows AI training to run across remote facilities connected by standard networking, while tolerating hardware failures that would shut down conventional training runs.

The research points toward a future where distributed compute can supplement the massive clusters that power today’s frontier AI training, though not replace them. The tightest training loops for the largest models still require enormous, purpose-built facilities with high-bandwidth interconnects.

Subscribe now

What Decoupled DiLoCo offers is a way to put additional capacity to work, wherever it sits, using the networking already in place between data center sites.

“As frontier models continue to grow in scale and complexity, we’re exploring diverse approaches to train models across more compute, locations and varied hardware,” the Google team wrote.

How Training Works Today

To understand the shift, it helps to understand today’s approach. Frontier AI models are typically trained using a programming model called Single Program, Multiple Data (SPMD) combined with synchronous distributed training. Thousands of accelerators execute the same training program on different portions of the workload, periodically synchronizing their results before advancing to the next step.

That tight synchronization is one of the primary reasons today’s AI training campuses concentrate massive GPU clusters behind ultra-high-bandwidth networking fabric and enormous power infrastructure. It also means that a single slow accelerator, or in many systems a hardware failure, can delay or interrupt work across tens of thousands of GPUs, making resilience and cluster-wide coordination critical to efficient training.

Read more

Each week I curate 5 links from the data center sector that I find particularly interesting, with my commentary on why they merit your attention.

Sound interesting? Subscribe and get them every Saturday.

Subscribe now

5 Notable Data Center Links

• Lancium Lines Up Partners for Two Texas MegaCampuses - It was a huge week for Lancium, which develops powered MegaCampuses for gigawatt-scale data centers. On Monday, QTS Data Centers and Lancium announced a campus in Hall County, Texas, anchored by a 1 gigawatt (GW) grid connection. QTS says it expects to build 11 data centers across 465 acres at the site. The Hall County campus is expected to bring more than $10 billion in capital investment to the region, with Lancium and QTS funding 100 percent of the energy infrastructure
  Then on Wednesday, Crusoe and Lancium announced a 1 gigawatt campus in Childress, Texas, that extends a partnership the two companies built in Abilene. Both deals lean on the same model: Lancium owns the land and the power, while a data center operator designs, builds and runs the campus.

• Csquare Raises $1.05 Billion in Data Center IPO: Colocation provider Csquare went public this week on the New York Stock Exchange, raising $1.05 billion in an IPO that valued the company at about $3.25 billion. Shares priced Wednesday at $21 and began trading on Thursday, with shares closing the week Friday at $20.67. Csquare is a holding of Brookfield Infrastructure Partners, and as we noted last month, the company is focused on retail colocation, not the wholesale and hyperscale construction that has absorbed most of the capital and attention during the AI boom. But its IPO is notable at a time when stock market investors have few “pure play” data center options. Digital Realty (DLR) and Equinix (EQIX) are the only U.S.-based data center operating platforms that are public and not in the process of being acquired (as is Digital Bridge).

• CleanSpark Signs $6.6 Billion Lease with Investment Grade Tenant - CleanSpark is the latest crypto miner to execute a transition to the AI data center business. CleanSpark says it has leased 175 MW of IT capacity to a “high-investment grade, leading global technology company” at its Sandersville, Georgia, campus. The agreement also offers the tenant exclusivity on CleanSpark’s entire Texas portfolio, spanning 885 MW of potential capacity. "This lease is a transformational moment for CleanSpark as we complete our evolution into a diversified digital infrastructure platform and begin monetizing our power portfolio at institutional scale," said Matt Schultz, CleanSpark CEO and chairman, who called the tenant win a “tremendous validation of our land-and-power strategy.” For background on the trend, see The Crypto to AI Power Shift Accelerates.

• Data Center Builders Are Racing to Offload Stakes Worth Billions (WSJ) - Potential sales of data-center operators are on the rise as owners of these firms seek exits amid strong investor interest in the sector. Bankers are reportedly working to sell stakes in firms including Netrality Data Centers, DataBank, Edged, and EdgeCore Digital Infrastructure, according to the Wall Street Journal. Always take M&A rumors with a grain of salt. That said, there’s no question that interest in data center M&A is heightened following last year’s $40 billion acquisition of Aligned Data Centers by a partnership including Blackrock, GIP, MGX, Microsoft and NVIDIA. It’s no surprise that operators are seeking exit opportunities, and there appears to be strong interest from buyers, with the new KKR-backed Helix Digital Infrastructure being named by the WSJ as a potential acquirer. The rest of 2026 could be pretty interesting … (If you don’t subscribe to the Journal, Bisnow has a summary).

• Protesters Pelt Data Center Site With Acid-Filled Balloons - It’s getting dangerous out there. This week Dutch environmental group Extinction Rebellion protested a data center project in Amsterdam by throwing acid-filled balloons onto the property. “Yesterday morning, balloons were thrown over the perimeter fence of Pure DC’s Amsterdam site which, protestors claim, contained corrosive chemicals,” the company told DCD. “While we have not yet tested the substances to independently verify these claims, if true, these actions could have seriously injured protestors or workers and are irresponsible and dangerous. … On this occasion there has been no damage, and no delay to construction, but activities of this nature would amount to serious criminal damage, for which we would pursue prosecution to the fullest extent of the law.”

If you found 5 Notable Links useful, please subscribe so you can get it each week!

In Case You Missed It …

Datacenter demand is enormous. Data center capacity is limited. So how does the digital infrastructure sector solve these challenges?

“Can We Build the Infrastructure We Keep Promising” was the topic as I made a guest appearance on the Shared Everything podcast from VAST Data. It’s hosted by Nicole Hemsoth Prickett, who for many years has been a leading voice in coverage of HPC, supercomputing and enterprise data centers.

We discussed how physical limitations are forcing a massive reality check. As power grids reach capacity and communities push back, the industry is shifting from raw expansion to hyper-efficiency and radical engineering. Read on to learn how top investors filter real projects from hype, how advanced density is reshaping system architecture, and when next-generation power sources (nuclear, space data centers) will actually arrive.

Here’s the conversation from my guest appearance on the Shared Everything podcast with Nicole Hemsoth Prickett.

This Data Center Download provides key insights from the Shared Everything podcast.

📌 Key Takeaways

• Filter the Hype: Evaluate project viability by focusing on experienced operators rather than speculative builders who lack proven operational capacity.

• Prepare for Liquid Cooling: Prepare to transition infrastructure to support high-density cooling as upcoming chip roadmaps push rack power demands toward 600 kW.

• Optimize Existing Grids: Leverage load flexibility, virtual power plants, and battery storage as the strategies to accelerate “speed to power” instead of simply waiting for new grid interconnections.

• Set Realistic Timelines: Anticipate a 2030s horizon for small modular nuclear reactors (SMRs) to realistically scale in the commercial market.

The New Era of Data Center Constraints

The massive demand for AI compute has run into physical, geographical, and social limitations. The industry is no longer operating in an era of unlimited expansion, requiring developers to navigate a complex web of power deficits and community pushback.

Sifting Reality from Hype in Project Funding

A major structural imbalance exists between proposed data center capacity and what can actually be built. Analysts from Aurora Energy Research tracking 280 gigawatts of proposed projects reveal a stark division: only 80 gigawatts come from experienced operators, while 150 gigawatts are proposed by developers who have never built a data center before.

Investors are becoming highly discerning, favoring established teams or funding platforms like Helix Digital Infrastructure (led by former AWS CEO Adam Selipsky) that may acquire existing, proven operators.

💡 Key Insight: Access to power is the ultimate differentiator. Projects with secured power and experienced teams will secure funding, while unproven greenfield proposals will face challenges. Consolidation is likely.

Navigating Community and Political Resistance

Data centers are facing unprecedented local opposition, a trend that has escalated dramatically over the last decade. Many communities are actively rejecting new builds due to concerns over noise, aesthetic impact, and strain on local utility grids. This resistance acts as a major constraint on development, proving that technology and capital are no longer the only hurdles to clear.

Engineering for Extreme Density and Efficiency

With hardware demands skyrocketing, data center operators are forced to rethink everything from silicon layout to facilities engineering to keep systems running without starting electrical fires.

The Inevitable Transition to Liquid Cooling

Air cooling is no longer sufficient for most AI workloads. With future hardware roadmaps pointing toward staggering 600-kilowatt racks, the industry is rapidly transitioning to advanced liquid cooling. While direct-to-chip liquid cooling using cold plates is the current standard, operators are preparing for two-phase (phase-change) cooling packaged in cold-plate form factors, and even microfluidics built directly into the silicon.

💡 Key Insight: According to AFCOM survey data, 40% to 50% of data center operators are actively evaluating liquid cooling, realizing their legacy infrastructure will soon be obsolete.

Redesigning Power and Software Architectures

Efficiency is being designed into every layer of the technology stack. On the hardware front, operators are shifting to 800 volt DC power distribution to minimize energy loss from the grid to the chip. Simultaneously, because physical space is limited, software engineers are focusing heavily on optimizing code and AI tokens to get more compute out of existing hardware footprint.

Solving the Power Crisis

Because power is the primary bottleneck for deploying new GPU clusters, the industry is desperate for fast, scalable energy solutions.

Optimizing the Existing Grid for Speed to Power

Building new power plants takes too long for AI companies wanting to monetize their hardware immediately. Instead of relying solely on on-site gas turbines, forward-thinking operators are focusing on grid flexibility. By utilizing virtual power plants, solar integrations, and large-scale battery storage, data centers can curtail their loads during peak grid demand, allowing utilities to grant them power access much sooner.

💡 Key Insight: Optimizing the grid we already have through load curtailment and battery storage is the fastest, most sustainable path to securing energy without burdening local communities.

The Reality of SMRs, Space, and Water-Based Facilities

While alternative concepts sound promising, their practical timelines vary widely:

• Small Modular Reactors (SMRs): The industry is highly receptive to SMRs, but regulatory hurdles and construction timelines mean nuclear-powered data centers won’t scale until the 2030s.

• Space-Based Compute: Despite the hype of “free solar power” in orbit, physical limitations on launching massive solar arrays, thermal management, and radiation make space data centers a niche solution rather than a mass-scale fix.

• Water-Based Data Centers: Floating or ship-mounted data centers (such as those being deployed by Samsung Heavy Industries) are seeing a resurgence as viable, rapidly deployable alternatives for coastal regions.

New York Gov. Kathy Hochul has signed an executive order calling for a one-year moratorium on new data center construction, becoming the first state to place a temporary ban on AI infrastructure development.

Hochul will pause state-level environmental permits for data center projects exceeding 50 megawatts for up to a year to provide time for her administration to develop a regulatory framework that protects ratepayers, the environment, the energy grid and communities across the state.

“This is a critical moment, and New York is going to take the time to get it right,” Hochul said in a speech in New York City.

“As data center development threatens to hike up utility bills, deplete our natural resources, and create uncertainty for New Yorkers, it’s my responsibility to take action and lead,” Hochul said in a related press release. “New York will lead the way in creating the strongest standards in the nation for data center development, ensuring that when companies succeed because of New York, New Yorkers succeed too.”

Subscribe now

Crafting A Policy Response

Hochul chose to act through an executive order rather than signing The Responsible Data Center Development Act (A11560/S10642), a broader regulatory framework for large data centers that the New York Legislature passed on June 4, 2026.

It would establish new requirements governing permitting, electric service, environmental review, labor standards, energy efficiency, and community benefits for large data center projects.

The executive order differs from the bill in several important respects. It applies to data centers of 50 megawatts (MW) and larger, rather than the bill’s 20 MW threshold, and it relies on existing executive authority rather than creating the broader statutory framework contained in the legislation, including a new utility tariff.

One of the bill’s sponsors, State Senator Kristen Gonzalez, joined Hochul for the announcement in New York.

Here’s a video from Hochul announcing the moratorium.

Roughly 14 to 15 states have considered some type of data center moratorium in 2026, and dozens of counties and municipalities have adopted or debated temporary pauses while they develop new rules for siting large AI data centers.

Hochul is among a group of governors seeking to navigate the strong public pushback against data centers.

• In Maine, Gov. Janet Mills vetoed a bill calling for a data center ban, calling it too restrictive.

• In Virginia, Gov. Abigail Spanberger recently signed a state budget introducing new taxes on data center power usage. The tax may impose up to $1.2 billion in new taxes on the state’s huge data center sector over the next two years, and was part of a compromise with legislators calling for a complete repeal of Virginia’s data center incentives, which run through 2035.

• New Jersey Gov. Mikie Sherrill has signed into law the Data Center Fair Share Act,

  which creates a new ratepayer structure specifically for data centers, requiring them to pay for their own energy use and the grid infrastructure needed to serve them.

Hochul’s decision to use an executive order is likely to be noted by other governors facing pressure to act on data center growth, most notably Gov. Josh Shapiro in Pennsylvania, a key growth area for AI infrastructure.

Governors in two other huge data center markets, Texas and Illinois, are also balancing how to respond to public sentiment against data center growth.

• Illinois Gov. Jay Pritzker announced a pause on new state tax incentive agreements after lawmakers failed to pass comprehensive legislation regulating data center growth.

• In Texas, Gov. Greg Abbott was initially a strong supporter of data centers, but recently shifted gears and called for regulatory policy reviews and perhaps legislation next year.

What’s in the Executive Order?

Here’s a summary of the details of Executive Order 62, outlining the action items, responsible agencies, regulatory processes and timelines for completion:

• A statewide environmental review.
  The Department of Public Service (DPS) will prepare a statewide Generic Environmental Impact Statement examining the environmental effects of large data centers through a public process that includes hearings, public comment and consultation with other state agencies.

• Pausing key environmental permits.
  The Department of Environmental Conservation (DEC) will suspend processing of state permit applications for data centers of 50 megawatts or more until the environmental review is completed, unless those applications have already been completed.

• Create a community benefits framework.
  Empire State Development will publish a Community Investment Framework outlining best practices for community benefit agreements, infrastructure investments, workforce commitments and transparency to guide negotiations between developers and host communities.

• Evaluate new grid funding mechanisms.
  DPS will ensure data centers fund the electric infrastructure needed to serve them, including potential grid investment funds, clean energy resources, battery storage and ratepayer protections.

• Establish an interconnection working group.
  DPS will create a Data Center Interconnection Working Group to improve the process for connecting large electric loads, streamline interconnection planning and evaluate how infrastructure costs should be allocated. It will also review how transmission owners assess the impacts of large load.

• Assess water withdrawal regulations.
  The DEC is tasked with reviewing whether New York’s water withdrawal regulations adequately address large data centers, and will have a year to develop recommendations for changes to policies and reporting requirements.

Who May Be Affected?

Importantly, the executive order should have no immediate impact for data centers in New York City, including those supporting the financial sector.

Hochul’s order is most relevant for large hyperscale developments that have not yet completed the environmental review process.

It’s early to speculate about specific projects, but in coming weeks we will likely hear far more about the implications for pending developments.

As of May 2026, nearly 12 gigawatts of data center load requests are in the New York Independent System Operator interconnection queue, with more than eight gigawatts entering the queue in 2025 alone.

In a meaningful step towards better transparency, Microsoft has shared details on energy and water usage at each of its data center campus locations. The site-level data was included in the company’s 2026 Environmental Sustainability Report, which was released Thursday.

The report identifies which campuses are using the most power and water, and also provides data on Microsoft’s use of non-potable water and replenishment programs.

Subscribe now

The disclosures are a response to community concerns about industry secrecy around data center energy and water usage, which has been a flashpoint in local pushback against AI infrastructure development. Microsoft said the data offers “a deeper understanding of how digital infrastructure interacts with local energy and water systems.”

This follows Microsoft’s recent decision to end the use of non-disclosure agreements in site development, a practice that has long been a trust-killer in community relations.

Taken together, the steps suggest Microsoft is willing to reconsider long-held business practices to improve its chances of delivering data center capacity.

The report highlighted the challenges of meeting sustainability goals during a rapid expansion of global infrastructure. Microsoft’s 2025 emissions were 25% higher than a year earlier, which the company attributed to its data center expansion and a decision to stop buying renewable energy certificates (RECs) to offset its emissions.

Power Usage by Campus

Hyperscale companies regularly disclose their overall data center energy usage and efficiency, including Power Usage Effectiveness (PUE), a leading metric for energy efficiency. But the companies have long asserted that site-level power usage was a trade secret, and disclosing it could impact their competitive position.

What’s also true is that data centers, by their nature, use huge amounts of power and water, reflecting the growing role of technology in all our lives.

The growth of the digital economy has concentrated enormous amounts of business processes in data centers, with users generating more data over time as adoption grows. Online services have become an essential part of everyday work and entertainment, including email, online banking, remote work, Netflix, YouTube, telehealth - all of which are powered by data center infrastructure.

Shifting all this activity in data centers has enabled faster and more efficient delivery of services, like the digital delivery of books, movies and music. Nonetheless, the volume of focused resource usage is significant, which has almost certainly been a factor in efforts to limit disclosures.

This is the first time I can remember seeing a campus-level disclosure of power usage, so there’s lots of interesting data.

Let’s dive into the trends from the Microsoft report:

Read more

Each week I curate 5 links from the data center sector that I find particularly interesting, with my commentary on why they merit your attention.

Sound interesting? Subscribe and get them every Saturday.

Subscribe now

5 Notable Data Center Links

• NVIDIA Partners With AI Clouds, Offering Revenue Share - NVIDIA is launching a new infrastructure business model to accelerate the deployment of AI factories, offering a revenue-sharing and credit-support framework to its neocloud partners. Participating AI cloud providers purchase NVIDIA systems with credit support and share revenue from the capacity, giving NVIDIA hardware sales plus recurring usage-linked income. The strategy could boost the availability of compute resources for AI-native startups, while supporting the growth and maturity of neocloud operators. SemiAnalysis notes that NVIDIA’s move can also support evolution of the GPU financing market, making it easier for lenders to ”get up the learning curve and adopt tools to price and manage risk” for neoclouds. Early adopters include Sharon AI and Firmus, which is developing a 360-megawatt, 170,000-GPU campus in Indonesia.

• Meta Announces 1GW Alberta Project, Affirms Data Center Spending - Just a week ago, pundits were concerned that Meta might be pulling back on data center spending and planning to rent excess compute. On Wednes1ay, Meta announced plans for a 1 gigawatt data center campus in Alberta, Canada and an investment of CAD $13 billion ($9.2 billion US). A day later, Mark Zuckerberg told Bloomberg that the company “needs all the computing power it can get.” But the Meta CEO also provided some context for the reports about plans to rent compute. “The offers that you get for using the compute are so high that it may make sense, in some cases, to rent out or consider those kind of deals instead of your own internal uses,” Zuckerberg said, adding that “I don’t know anyone in the industry who feels like they have too much compute.”

• Google Will Develop Huge Wyoming Data Center Campus - Google will own and operate a 2.7-gigawatt data center project in Wyoming, the Cowboy State Daily reported this week. The company is taking over a 716-acre campus south of Cheyenne, which is seeing a surge in data center development. Microsoft has been in the Cheyenne market for many years, while Meta has a campus under development.

• Blue Owl Launches Dark Fiber Platform - Funds managed by Blue Owl Capital have launched Kirkwood Infrastructure Group to capitalize on strong demand for dark fiber routes from AI data centers. Kirkwood IG aims to expand into new markets alongside key data center development partners, with an initial focus on Florida and the Southeast U.S. "The digital infrastructure buildout in the U.S. is just beginning, and while a new company, our team has a tremendous track record of deploying industry-leading network infrastructure that enables the next wave of compute requirements and digital transformation,” said Scott Bergs, CEO of Kirkwood IG. Bergs previously led DF&I, a Blue Owl holding which was acquired June 30 by Summit IG.

• Ares Funds Invests in Sabey Data Centers - Ares Secondaries funds have made a minority equity investment in Sabey Data Centers, which operates 251 megawatts of capacity across six U.S. data center campuses. Sabey says it has an opportunity to triple that output by 2036 on existing land holdings.
  The deal was announced by Sabey and its existing financial partner, National Real Estate Advisors. “With Sabey, National and now Ares aligned behind the business, we are well positioned to execute our disciplined growth plan and deliver the mission-critical capacity our customers need,” said Tim Mirick, President of Sabey Data Centers. For background on Sabey and its growth, see our podcast with Mirick and NREA’s Jeff Kanne.

If you found 5 Notable Links useful, please subscribe so you can get it each week!

In Case You Missed It …

Micron Technology said this week it is raising its planned U.S. investment to more than $250 billion through 2035, a figure that would have been unthinkable for a memory chipmaker a decade ago.

The company’s expanded U.S. investment plan is the clearest sign yet that high bandwidth memory has become one of the most valuable commodities in the artificial intelligence buildout. It also reflects how AI has reordered the economics of an industry long defined by brutal boom-and-bust cycles.

Another signal of the memory boom came from Wall Street, where U.S.-listed shares of South Korean memory specialist SK Hynix made their debut this morning, and were up about 15% in midday trading.

Micron marked its expansion announcement with the first “concrete pour” at its Clay, New York campus, arriving more than a quarter ahead of schedule and signaling that the company’s spending plans are translating into physical construction, not just projections on a slide.

“As America celebrates its 250th anniversary, data and memory are foundational to the modern economy, and Micron is increasing our U.S. investments to more than $250 billion through 2035 to meet that moment,” said Sanjay Mehrotra, Micron’s chairman, president and CEO.

How AI Rewrote the Memory Business

The scale of Micron’s commitment tracks a broader repricing of memory driven by high bandwidth memory (HBM), the specialized chips that sit alongside GPUs in AI training and inference clusters.

AI workloads consume far more DRAM (dynamic random-access memory) and HBM per server than conventional cloud computing, and that demand has turned what was once a commodity, cyclical business into one of the tightest bottlenecks in the AI supply chain.

Micron isn’t the only beneficiary. Its chief HBM rivals, SK Hynix and Samsung Electronics, have both ridden the same wave, with all three companies now racing to lock in years of future capacity for AI customers.

SK Hynix, which holds the largest share of the global HBM market, listed its American depositary receipts (ADR) at $149 per share on the Nasdaq today under the ticker SKHY, raising $26.5 billion. At midday shares traded at about $172.

The company has said proceeds will fund expansion of its South Korean manufacturing facilities and purchases of extreme ultraviolet lithography equipment.

Read more