A San Francisco startup called Marengo surfaced recently with a pitch that sounds either wildly optimistic or perfectly timed, depending on whom you ask: the company claims its AI-powered platform can design data centers in half the time and at half the cost of traditional engineering firms. The assertion arrives just as developers scramble to build nearly 100 gigawatts of new data center capacity by 2030—an infrastructure supercycle that JLL estimates will require up to $3 trillion in investment.
The startup, backed by Y Combinator along with scout funds from Andreessen Horowitz and Index Ventures, says its software can explore more than 1,000 concept designs in parallel, a stark contrast to the three or four concepts that traditional engineering processes typically manage. Marengo's platform runs site due diligence through front-end engineering design and permitting packages, delivering what the company describes as "validated outputs" that professional engineers can trust.
Whether that claim holds up in practice remains to be seen. But the market Marengo is entering leaves little room for patience.
When the calendar becomes the constraint
North American data center vacancy has hovered at or near 1 percent through much of the past year, with tightness expected to persist. JLL reported in mid-2026 that 66 gigawatts of capacity were under construction globally, with 95 percent of it already pre-committed. Many tenants now contract for deliveries scheduled in 2028 or beyond, a timeline that would have seemed absurd just three years ago.
Equipment lead times stretched 50 percent longer than pre-2020 baselines, averaging 33 weeks, according to JLL data. Large grid-scale transformers reached delivery windows of 160-plus weeks by early 2026, Reuters reported. In June, the Federal Energy Regulatory Commission issued orders directing regional transmission operators to justify or reform rules for large load interconnections—data centers prominent among them. FERC stated in its release that by requiring RTOs/ISOs to either defend or revise their tariffs, the commission is moving to ensure that Americans have reliable, affordable power.
The hyperscalers are pouring capital into this buildout at a pace that defies easy comprehension. Microsoft has indicated it expects to spend roughly $190 billion on AI-related infrastructure for calendar year 2026, CFO Amy Hood said on an earnings call. Alphabet spent $80.6 billion on capex in the first half of 2026 alone, according to a regulatory filing, with future lease payments for data centers not yet commenced totaling an additional $85.2 billion. Meta, in its Q4 2025 prepared remarks, guided 2026 capex at $115 billion to $135 billion, driven almost entirely by AI infrastructure.
The constraint, it turns out, isn't just money or real estate.
Design complexity meets thermal reality
When NVIDIA's GB200 NVL72 rack-scale systems pushed power densities to around 132 kilowatts per rack, the physics of cooling became unavoidable. "An AI data center equals liquid cooling," Peter Huang, global president of thermal management and data centers at BP Castrol, told Data Center Dynamics in a March interview. AWS developed new liquid-cooling approaches specifically to deploy GB200 clusters quickly, a shift necessitated by Blackwell's thermal profile.
Industry experts note that traditional design workflows weren't built for this pace or this iteration count. Developers holding six to twelve months of strategic equipment inventory need to make siting and design commitments earlier than ever. Average data center construction costs climbed from roughly $7.7 million per megawatt in 2020 to $10.7 million per megawatt in 2025, according to a report from GFH citing JLL data. Turner & Townsend's construction cost index identified a 7 to 10 percent premium in the United States for AI-ready, high-density liquid-cooled facilities versus traditional designs.
That's the environment Marengo is betting it can crack open.
High-consequence credentials
The founders' backgrounds suggest they understand the stakes. Emil Ares, the CEO, holds a physics degree from Cambridge and worked on 56-megawatt magnets at the 1.4-gigawatt National MagLab facility before joining the UK Space Agency, according to the company's Y Combinator profile. Gad Marconi, the CTO, built and deployed on-site data center infrastructure at Switzerland's 25-megawatt Flumenthal hydroelectric plant after working as an AI software engineer at a nuclear and energy engineering firm.
The company says it has "developed internal AI design tools that our professional engineers use to accelerate their engineering design work," according to its Y Combinator launch text. The startup's website shows deliverables spanning site due diligence, feasibility studies, concept design, and front-end engineering design and permitting packages for a 500-megawatt campus concept. Multi-objective optimization surfaces risks instantaneously, the company claims.
Marengo says it works with some of the largest data center builders in the world but has declined to name customers publicly. The startup, founded recently with a small team, positions itself as validation-first rather than automation-first—outputs engineers trust, not just outputs that look plausible. Whether that distinction proves meaningful will depend on adoption.
The acceleration race
Marengo enters a market where incumbents and new entrants alike chase design speed with varying degrees of success. Autodesk expanded its Design and Make Intelligence capabilities and launched Forma Building Design, according to company news releases. Bentley Systems maintains a data centers solution page highlighting tools like Synchro and OpenBuildings. Jacobs released a digital twin solution for gigawatt-scale AI data centers built on NVIDIA's Omniverse DSX platform. ALICE Technologies, which sells generative scheduling software for capital projects, announced a collaboration with McKinsey to expand adoption.
Smaller players are moving too. TestFit added a data center typology to its generative site and feasibility platform, with updates including "Data Center Yard Depth" and new U.S. power plant and terrain layers. Zenerate added "Data Center Planning" capabilities. Hypar offers a cloud platform to encode and generate building logic applicable to repetitive program generation.
The hyperscalers, unsurprisingly, move faster still. Digital Realty partnered with NVIDIA on an AI Factory Research Center in Manassas, Virginia, advancing the DSX blueprint and liquid cooling innovation. Equinix announced it would deploy NVIDIA AI Factories across its global footprint in collaboration with Cisco. Robert Thiel, principal architect at Continental AG, said in an Equinix blog post that placing an IBM storage and NVIDIA GPU cluster in an Equinix "AI-ready" data center "in just two weeks" enabled the company "to increase the number of AI experiments by 14x."
Friction points remain
The buildout cycle shows no sign of stalling. McKinsey projected global data center spending could reach $7 trillion by 2030 in a March analysis—a figure that includes chips, power systems, and facilities. Dell'Oro Group forecast data center capex alone would surpass $3 trillion by 2030. The International Energy Agency reported data center electricity demand grew 17 percent in 2025, with sustained higher load growth in data-center-dense regions projected into 2027.
Design and permitting remain friction points that regulators and industry players openly acknowledge. The Federal Permitting Improvement Steering Council added its first data center project—QTS Richmond DC5—to FAST-41 permitting coverage in April, seeking to streamline timelines under the 2015 statute. AECOM called in a February report for a sovereign data center framework in the United Kingdom to maximize national economic value from AI-driven growth. "The UK risks losing strategic control and economic value unless growth is guided by clearer national priorities," Adrian Del Maestro, vice president of global energy advisory at AECOM, said.
Marengo's half-time, half-cost pitch hinges on whether parallel AI-driven exploration can replace iterative human revision cycles without sacrificing engineering rigor. The company's wager is that, in a market where vacancy sits at 1 percent and Microsoft alone plans to spend $190 billion this year, time bought is revenue secured. The developers now contracting for 2028 deliveries with transformer orders already locked in and 95 percent of the under-construction pipeline already spoken for will decide if the bet pays off.
For now, the startup remains a small team with large ambitions entering a market that rewards speed but punishes mistakes. That's either the perfect moment to arrive or precisely the wrong one.
