Raoul Pal: Energy as a Constraint for AI is Now Consensus, but Energy Itself is Also Growing Exponentially
Raoul Pal stated that it is now a consensus that energy is a constraint for artificial intelligence, but energy itself is also growing exponentially, and is not currently the limiting factor for expansion; major cloud providers are projected to spend about $600 billion on capital expenditures this year, comparable to railroads. Some are calling it a bubble, but they are building in response to pent-up demand, with demand exceeding deliverable infrastructure, and capital for construction is not lacking.
Independent research pushes the same spending into higher ranges: Goldman Sachs predicts that large U.S. hyperscale firms may spend about $800 billion on capital expenditures in 2026, with a consensus of about $1.1 trillion in 2027, estimating that around $300 billion in annual AI revenue is needed to break even on investments; cloud contract backlogs are reported to exceed $1.5 trillion. Other institutions estimate that the combined spending of five firms in 2026 will be around $750 billion to $820 billion, with projections for 2027 ranging from $1 trillion to $1.3 trillion. Operating cash flow is no longer sufficient to cover expenditures, leading to a shift towards bonds and project financing. Reports indicate that grid access, transformer, and gas turbine delivery times are becoming more significant physical bottlenecks than cash.
Pal and David Mattin monthly analyze the intersection of energy, AI, and robotics in a study called The Exponentialist. He has another framework of "more intelligence per unit of energy" and has discussed the spillover effects of data center electricity usage and electricity prices. The $600 billion figure is his speaking point, not the latest upper limit of consensus from sell-side analysts.
The railroad analogy points to fixed capital being sunk, serving later traffic. The backlog argument rewrites spending from "vision" to "contracted but undelivered." The exponential growth of energy refers to the generation and efficiency curves, which do not automatically shorten the queue for grid connection. Lunch discussions at the White House about data centers and communities, as well as the politics of electricity prices in various states, reflect the same constraint projected onto the ballot.
In market mechanisms, the narrative being bought is one of "demand queued up, cash still sufficient to write checks" for expansion, while the narrative being sold is a rebuttal to bubble claims. Funds are flowing from the operating cash flow of hyperscale firms and new debt into data centers, the grid, and chips; beneficiaries are those delivering power equipment, energy storage, and accelerators, while local users bearing the spillover costs are under pressure. Event-driven factors are framed by investors on a monthly basis, not as a new spending guideline.
The constraint shifts from chips to joules and then to the grid connection calendar. Ample capital only proves the ability to place orders, not the ability to deliver power next year. The larger the backlog, the more sensitive unfinished assets are to interest rates and electricity prices.
Source: Public Information
ABAB AI Insight
Pal frames the macro cycle as a story of the exponential age: first liquidity and crypto, then intelligence per watt. The $600 billion railroad analogy aims to shift AI spending from concept stocks to infrastructure positions. Sell-side numbers have already surpassed his speaking point, indicating that the expansion narrative itself is also being revised upwards. Backlogged contracts are a key tool—having $1.5 trillion in cloud commitments means capital expenditures are no longer like the fiber vision of 2000, but like delivery orders. However, the $300 billion annual revenue needed to break even remains a projection, not already realized profit.
The capital path involves five cloud and social giants fronting operational cash flow, then passing the gap to the debt market and project financing. Chip manufacturers and utilities are locked into the same delivery chain: cards, cabinets, power, and cooling must arrive simultaneously. The motivation is that whoever secures power and transformers first can sell their model reasoning as cloud contracts. Pal writes energy as not yet bottlenecked to continue over-allocating to intelligent assets; the grid queue and five-year gas turbine lead times are the counterarguments. Robotics is included in the same convergence because agents and physical labor will ultimately compete for the same kilowatt-hour.
Comparative objects are 19th-century railroad capital expenditures and the first time in the 2010s that hyperscale clouds treated capital expenditures as a moat. The industry phase is transitioning from "can we buy cards" to "can we get power on time." Nvidia remains the shovel, while power generation and substations become the new shovels. Local electricity price politics will determine the next round of permits, more so than research subscriptions can rewrite timelines.
Structural judgment is that technological substitution is subject to physical layer pricing. The mechanism is that software demand is indexed, while power generation and grid connection climb project by project, with the gap filled by capital expenditures. Substitution occurs at the model layer, while bottlenecks occur at the joule layer. Whoever controls dispatchable power will have veto power in the next round of pricing reasoning.
ABAB News · Law of Cognition
The constraints in consensus are often not the real bottlenecks at the moment.
In the face of backlog construction, there may still be a wall at the meter.
Ample capital shortens order time, not grid connection time.