190 GW of announced data center capacity is reshaping the hardware investment stack

Bessemer Venture Partners published a detailed roadmap of the AI data center stack on May 19, mapping six areas where it sees venture-scale opportunity. The firm’s thesis is straightforward: AI’s insatiable demand for compute is creating a physical infrastructure crisis that startups, not incumbents, are best positioned to solve.

The numbers are staggering. As of early 2026, 190 GW of hyperscale data center capacity has been announced across 777 projects. Of that, roughly 148 GW is planned, 21 GW is in construction, and 12 GW is operational. Global data center electricity consumption is projected to more than double by 2030. In the U.S., data centers will soon consume more electricity than all energy-intensive manufacturing combined.

But the headline figure masks a deeper problem. Data centers can be built in 12 to 18 months. Connecting them to the grid takes five to seven years. Of the 110 data center projects slated to come online in 2025, more than a quarter were delayed due to power, permitting, and construction constraints. That gap between construction speed and grid readiness is the central investment thesis of the entire Bessemer roadmap.

The firm identifies six layers of opportunity: permitting and site selection, power generation, transmission and power conversion, software and orchestration, construction and labor, and cooling technologies. Each represents a bottleneck where existing processes are manual, slow, or capacity-constrained.

The permitting bottleneck is a software opportunity

Bessemer estimates that over $5 billion is spent annually on complex infrastructure permitting in the U.S., representing more than a third of total permitting spend. Over $1.5 trillion of infrastructure capital is currently stuck in the permitting pipeline. Between March 2024 and 2025, 16 data center developments were delayed or denied due to permitting restrictions, with local community pushback a leading cause.

The incumbent providers are large consulting firms like Tetra Tech and Quanta Services, alongside a fragmented market of more than 50,000 smaller firms, many with fewer than 15 employees. Their processes are largely manual. Bessemer highlights two startups attacking this: Lorica, which is building an AI-native permit preparation and execution service, and Paces, which unifies grid, permitting, and environmental data into a single platform. Paces customers report closing roughly three times as many deals because risks like grid constraints and zoning conflicts are surfaced upfront.

This is not a glamorous corner of the AI stack. But it is where billions of dollars of capital are currently frozen. Software that compresses permitting timelines from years to months has a direct line to revenue.

The “Bring Your Own Power” movement is real

The most consequential shift Bessemer documents is the move from grid-connected data centers to on-site generation. While grid-connected sites still account for the largest share by project count at 45%, on-site generation and hybrid approaches together account for close to half of all announced capacity. Approximately 50 GW of behind-the-meter gas generation projects were announced in 2025 alone.

Bessemer calls this the “Bring Your Own Power” movement. Data center operators are willing to accept the additional complexity of building and managing on-site power in exchange for certainty on capacity, timelines, and emissions. This is a structural shift in how the industry thinks about energy.

The portfolio companies Bessemer names in this category are telling. Boom Supersonic, known for developing supersonic passenger aircraft, has adapted its jet engine core into Superpower: a 42 MW natural gas turbine purpose-built for data center power generation. Arbor is building a next-generation gas turbine using supercritical CO₂ as a working fluid with built-in carbon capture. Inertia, co-founded by Twilio founder Jeff Lawson, is commercializing the laser-based inertial confinement fusion approach pioneered at Lawrence Livermore National Laboratory.

Fusion remains the longest-horizon bet. But the presence of a supersonic jet company repurposing its engine for data centers signals how quickly the power generation landscape is being reshaped.

The transformer bottleneck is structural

The third layer Bessemer identifies is transmission and power conversion, and this is where the numbers get most alarming. Transformer demand has increased 119% from 2019 to 2025. Manufacturing capacity has not kept pace. Lead times from incumbents like General Electric, Siemens, and Mitsubishi have stretched to as long as five years, up from roughly one year pre-COVID.

These are large, highly engineered devices built to order by a small number of manufacturers. Adjacent grid hardware faces the same crunch: switchgear lead times now stretch beyond 60 weeks. The bottleneck is structural and will not resolve quickly.

Bessemer points to American Terawatt, which is building a private wire high-voltage direct current transmission network that connects data centers to power sources without waiting for the public grid. This is a bet that the grid itself is the bottleneck, and that private infrastructure can bypass it.

Software and orchestration is the sleeper layer

The fourth layer is software and orchestration, and it is where the Bessemer roadmap most directly intersects with the AI software stack. Data centers running at 50 to 100 MW are complex physical systems. Software that manages power distribution, cooling optimization, workload scheduling, and predictive maintenance is increasingly critical.

Bessemer does not name specific startups in this category in the excerpt, but the opportunity is clear. As data centers push toward higher power densities driven by NVIDIA’s 800V DC architecture, the software layer that manages power delivery to individual racks becomes a competitive advantage. The firm notes that the most durable businesses will own the hardware relationship with the data center operator and layer dispatch optimization or predictive controls on top.

What this means for AI builders

The Bessemer roadmap is a useful corrective to the narrative that AI progress is primarily a software and model capability story. The physical infrastructure that powers AI training and inference is facing constraints that will shape which models get built, where they get deployed, and at what cost.

For AI researchers and engineers, the implications are practical. Model architecture choices that reduce power consumption per token will become increasingly valuable. Inference optimization that allows models to run on less power-dense hardware will open deployment options that are otherwise closed. The cost of compute is not just a function of GPU availability. It is a function of power availability, permitting timelines, and transformer lead times.

The Bessemer roadmap makes clear that the next decade of AI infrastructure will be defined by companies making electrons cheaper, faster, and smarter. The firms that succeed will be those that treat power as a first-class constraint, not an afterthought.