The world's largest technology companies are set to spend roughly $500 billion on AI infrastructure over the next few years, but a less obvious constraint is slowing the buildout. It's not chip supply or power availability. It's the design process itself.
Data centers increasingly require liquid cooling for high-density racks, custom power delivery at rapidly escalating densities, and grid connections that can take longer to secure than the buildings take to construct, according to multiple industry reports published in 2026. The mechanical, electrical, and plumbing workflows required to make these facilities work have become a critical path—and a growing business opportunity.
Enter Marengo, a San Francisco startup from Y Combinator's summer 2026 batch that says it can use AI to design data centers faster and at lower cost than traditional methods. The company, which has between two and ten employees and was founded this year, says it's working with some of the largest data center builders in the world. It has not named those customers. No case studies appear on its LinkedIn company page as of August.
CEO Emil Ares, who studied physics at the University of Cambridge, previously steered the company toward space mission design before pivoting to data centers in 2026. That shift mirrors a broader pattern. Construction technology startups are racing to automate the workflows that now account for 40 to 55 percent of data center construction costs, according to design automation firm Augmenta.
The race matters because data center capacity in the United States will double in the next three years despite mounting headwinds, Synergy Research Group said in July. JLL projects nearly 100 gigawatts of new capacity from 2026 to 2030, part of what the firm calls a $3 trillion "investment supercycle." Average global construction costs reached $11.3 million per megawatt in 2026, up 6 percent year-over-year, JLL said in January.
The Hyperscaler Arms Race
Meta narrowed its 2026 capital expenditure guidance from an initial range of $125 billion to $145 billion down to $130 billion to $145 billion in second-quarter results released July 31. Microsoft projected roughly $190 billion in calendar 2026 capex during its fiscal third-quarter call in May. Alphabet's second-quarter 2026 capex hit $44.9 billion, more than double the prior year, according to an S&P Intelligence brief published in July. Amazon CEO Andy Jassy wrote in May that the company's AWS front-loaded 2026 capex covers "land for the data centers, power, the buildings themselves, the hardware, the chips, networking gear."
Those hundreds of billions are colliding with power grid constraints. Projects in the PJM Interconnection queue now face extended wait times after approval due to equipment shortages, particularly large power transformers, and transmission upgrades, according to a May Data Center Knowledge report. The Department of Energy noted in March that the electric grid carries more than 80,000 distribution transformer variants, complicating procurement timelines.
CBRE data shows Northern Virginia, the largest U.S. data center market, delivered more than one gigawatt in 2025 yet still posted a 0.3 percent vacancy rate in mid-2026. Demand, in other words, remains ferocious.
The Federal Energy Regulatory Commission has issued a series of orders to speed large-load integration. In June and July 2026, it directed all regional transmission organizations and independent system operators to reform how data centers and other large loads connect. Those actions follow FERC Order 2023, finalized in July 2023, which overhauled interconnection queue procedures, and Order 1920, issued in May 2024, which established long-term regional transmission planning frameworks now cycling through 2029 to 2031 implementation.
Software Meets Steel and Concrete
The design bottleneck has opened wedges for software companies. Augmenta, which automates electrical raceway routing and MEP coordination, said it cut modeling time by roughly 50 percent across three projects in three months in a case study involving Owen Electric. ALICE Technologies, which offers generative scheduling, said it saved 63 days in a data center delivery by optimizing sequencing and removing soft logic. The company announced an alliance with McKinsey in April to transform capital project delivery, citing adoption by an unnamed leading data center provider.
Jacobs, the global engineering and construction firm, released a data center digital twin solution in March built on NVIDIA's Omniverse DSX blueprint for gigawatt-scale AI facilities.
"Owners need greater certainty before committing capital," Jacobs CEO Bob Pragada said in the release. "We're creating solutions that help owners plan more efficiently, improve performance and achieve more secure and resilient long-term operations."
NVIDIA's DSX ecosystem has grown to include validation work at Digital Realty's Manassas campus, reference architectures from equipment partners including Vertiv, and simulation stacks from Cadence, PTC, and Siemens, according to NVIDIA blog posts published between November 2025 and June 2026. Google released its Brazos liquid cooling retrofit system in June and committed to open-sourcing designs through the Open Compute Project, drawing on more than seven years of gigawatt-scale liquid-cooled TPU deployments that achieved roughly 99.999 percent uptime.
Other construction-tech entrants include nPlan, which has trained AI models on approximately 750,000 schedules representing $2 trillion in spend. Entangl, a Y Combinator summer 2024 company, focuses on AI for data center engineering and operations issue detection. Livio claims 75 percent faster builds with panelized, engineered AI data centers.
Density as Design Challenge
The push for automation stems partly from technical complexity. High-density AI racks now routinely exceed 50 kilowatts per rack. In some configurations they approach 100 kilowatts or more, according to the Uptime Institute's 2026 Global Data Center Survey, released in July. An IEA 4E EDNA report published in February found that power densities associated with NVIDIA's B200-class GPUs make liquid cooling a design requirement for many AI deployments, not an option.
Liquid cooling introduces new trade-offs. AWS announced in December 2024 that it had developed modular multimodal cooling and power delivery innovations capable of increasing rack density sixfold in two years, with potential for a further threefold increase. Google engineering teams wrote in February 2025 about vertical power delivery, water cooling, and power-aware workload allocation as integrated codesign challenges across the infrastructure stack.
Academic research highlights the system-level implications. A June preprint on ArXiv identified the need for high-voltage ratio DC-to-DC converters, low-voltage DC distribution, and medium-voltage solid-state transformers to support extreme transients and densities. A March paper on liquid cooling optimization at the Frontier supercomputer, validated per ASHRAE Guideline 14, reported up to 28 percent total energy savings with ramp-constrained joint optimization using digital twin models. An April study on generative design for direct-to-chip cold plates showed temperature reductions exceeding 5 degrees Celsius on average and more than 35 degrees Celsius at maximum hotspots compared to baseline designs.
ASHRAE has updated standards to reflect the shift. Standard 90.4-2025 is now available, and Technical Committee 9.9 has published resources on AI data center energy and thermal frameworks. ISO registered a new work item in April for guidance on liquid cooling applications.
Speed as Competitive Weapon
Carnegie Endowment for International Peace wrote in a June report that "the decisive variable is speed" in data center development, comparing timelines and policies across geographies. The report noted valuation premiums for jurisdictions that enable faster project completion. It also highlighted the behind-the-meter trend, in which data centers colocate with generation to bypass grid connection delays.
Cleanview, a research firm tracking behind-the-meter projects, identified 59 such developments and expects approximately 3 gigawatts online by the end of 2026 under current schedules, with cumulative capacity potentially reaching 5 to 13 gigawatts in 2027 depending on slippage.
State regulators are responding. The Virginia State Corporation Commission has introduced initiatives to shift more upstream costs to data centers, including a new GS-5 rate class, according to commission fact sheets published between February and June. The measures aim to reduce cross-subsidies as Northern Virginia's data center footprint expands.
EPRI's "Powering Intelligence 2026" report, updated in February, projects data center electricity demand could reach 200 to 800-plus terawatt-hours by 2030 depending on AI adoption scenarios, up from an International Energy Agency estimate of 17 percent growth in 2025 alone. EPRI emphasizes grid and utility collaboration, flexible operations, and frameworks for integrating data centers with transmission planning. The North American Electric Reliability Corporation updated its Large Loads Action Plan in March, setting a December 31 deadline for new reliability standards addressing large loads including data centers.
Proof Points Matter
The convergence of AI-driven design automation, digital twin validation, and modular construction strategies reflects an industry under time pressure. DPR Construction's market condition reports from the first and second quarters of 2026 noted that owners are locking design decisions earlier, increasing use of prefabrication and modular methods to protect schedules, and applying lessons from the "hyperscale data center playbook" to other sectors.
For Marengo and its competitors, the opportunity hinges on proving claims with data. Augmenta has published multiple case studies with named contractors. ALICE has released schedule compression metrics and announced a McKinsey partnership. Jacobs has tied its digital twin to NVIDIA's ecosystem and named a validation site at Digital Realty.
Marengo's pitch of faster, cheaper design resonates with documented industry pain points: MEP coordination delays, equipment procurement lead times, grid connection backlogs. But the startup has not yet published third-party validation or customer names. In an industry where billion-dollar projects hang on engineering credibility, that matters.
The firms best positioned to capture the automation opportunity will likely be those that integrate across disciplines, combining MEP spatial design, thermal and power simulation, schedule optimization, and real-time commissioning feedback rather than solving point problems in isolation. Johnson Controls released a second data center reference design guide in May to advance industrial-scale AI facilities. Vertiv and other equipment vendors are publishing DSX reference architectures. Google, AWS, and other hyperscalers are open-sourcing liquid cooling designs.
Dell'Oro Group forecast in February that the multi-year AI cycle will push data center IT capital expenditure to $1.7 trillion by 2030. How much of that spending accelerates, and how much stalls in engineering backlogs, will depend in part on whether automation tools can deliver the productivity gains their makers promise. Perhaps more importantly, it will depend on whether buyers trust those tools enough to compress timelines on projects where a single mistake can cost tens of millions of dollars and months of delay.
