Bessemer Venture Partners has published a roadmap of the AI data center infrastructure stack across six categories: permitting and site selection, power generation, power transmission and conversion, software and orchestration, construction and labor, and cooling. The BVP Atlas analysis, written by Lindsey Li, Brielee Lu, Josh Hechtman, and David Cowan, frames the current moment against a historical parallel: Henry Bessemer's 19th-century steel process enabling civilization-scale infrastructure change. The investment thesis is that the hardware and software enabling the AI buildout remain in early innings. The infrastructure enabling the hardware is further behind.
The scale context: 190 gigawatts of hyperscale capacity has been announced across 777 projects, of which approximately 148 GW is planned, 21 GW is under construction, and 12 GW is operational. In 2025, 78 percent of built-environment venture capital, $4.5 billion of $5.7 billion total, was deployed to data centers. Global data center electricity consumption is projected to double by 2030.
Data centers are constructible in 12 to 18 months. Grid connections take 5 to 7 years. That mismatch is the single largest constraint on buildout velocity: more than 25 percent of 110 projects slated for 2025 were delayed due to power, permitting, or construction issues. The response has been a shift toward behind-the-meter generation. Roughly 50 gigawatts of on-site gas generation was announced in 2025. On-site plus hybrid approaches now account for nearly half of announced capacity.
The power hardware supply chain has not kept pace. Transformer demand increased 119 percent between 2019 and 2025. Lead times have stretched to five years for some classes of equipment, from roughly one year pre-COVID. Switchgear backlogs exceed 60 weeks. Grid congestion costs U.S. consumers an estimated $11.5 billion annually.
BVP identifies cooling as one of the six critical investment categories because rack density evolution has made it a hard technical constraint rather than a design preference. The trajectory the analysis describes: 20 to 40 kW per rack during the cloud era, 500 to 600 kW in current AI builds, and a 1 megawatt per rack target associated with Nvidia's Rubin Ultra architecture. AI workloads are projected to drive a 50x power density increase within one year of the Rubin generation entering production.
The analysis cites direct-to-chip cooling and immersion as the two modalities dominating new AI data center architectures. Within direct-to-chip, BVP highlights Corintis as developing AI-designed microfluidic cold plates. Within orchestration, Phaidra is cited for using real-time power draw signals to optimize cooling system response. The industry is standardizing on 800V DC power architecture, which carries implications for cold plate and CDU design across the distribution chain.
The workforce constraint appears in the construction and labor section but applies to cooling commissioning as well. The construction industry shortage stands at approximately 439,000 workers as of late 2025. Peak crew sizes for current data center builds are running 4,000 to 5,000 people, compared with roughly 750 during the cloud era. The average data center workforce age is 53. The skilled labor required to commission liquid cooling systems, CDUs, and chilled water plants is a subset of a workforce already under pressure.
About this sponsor
Bessemer Venture Partners is one of the longest-standing venture capital firms in the United States. The BVP Atlas publishes research on infrastructure, energy, and technology investment across the AI buildout stack.
The AI Data Center Stack roadmap covers permitting, power generation, transmission, software orchestration, construction, and cooling. It is the firm's most comprehensive published analysis of where capital is flowing and where the bottlenecks are forming in the AI infrastructure buildout.
Read the full BVP Atlas roadmap →