ARK Founder Cathie Wood: Anthropic Agrees to $50 Billion per Gigawatt
ARK Invest founder Cathie Wood stated in a recent program that Anthropic has agreed to pay SpaceX AI approximately $50 billion per gigawatt for computing power rental. Elon Musk's construction costs are estimated to be in the mid-200 billion range, allowing for immediate investment returns.
According to hard numbers disclosed in SpaceX's prospectus, Anthropic will pay $1.25 billion monthly starting from the agreed amount, totaling about $15 billion annually until May 2029; the months of May and June will have a ramp-up discount. The contract covers Colossus and Colossus II, located at the Tennessee-Mississippi border, with either party able to terminate with about 90 days' notice. Musk has previously stated not to get overly confident, as this arrangement can end in the short term, and the termination rights should be included at SpaceX's request.
ARK's Chief Futurist Brett Winton previously calculated that Anthropic could generate about $24 billion in revenue per rated gigawatt, while SpaceX AI's construction costs are about $29 billion per gigawatt; if the lease lasts for five years, the pre-tax cumulative cash flow could exceed $50 billion. If the rental includes H100/H200 from Colossus I and some GB200/GB300 from Colossus II, the gross cash flow over five years could exceed $60 billion. This waterfall chart has been criticized by accounting scholars for making overly strong assumptions, which do not automatically align with the contract duration and depreciation.
The construction speed has public timestamps: the first batch of about 100,000 H100 clusters for Colossus lit up in 122 days, then doubled in about 92 days; the first batch of about 110,000 GB200-level clusters for Colossus II took about 91 days. The power source uses gas turbines for on-site generation to bypass grid queuing. Subsequently, Google has agreed to rent for $920 million monthly starting in October until June 2029; there are also clients like Reflection. The total nominal amount of multiple rental agreements publicly advertised could reach hundreds of billions of dollars, but all are subject to duration and termination clauses.
Mechanically, the buyers are cutting-edge labs lacking chips, while the sellers are SpaceX AI, which has established the power plants and cabinets first. The pricing is determined by the combination of computing power shortages and prospectus disclosures: rental fees are transferred from Anthropic and Google to the established clusters; the beneficiaries are the lessors who quickly convert capital expenditures into monthly cash, while the pressured parties are traditional data center vendors still adhering to two-year construction timelines and tenants who must accept 90-day withdrawal clauses.
Public information has not provided a complete conversion table between Wood's stated "$50 billion per gigawatt" and the prospectus's "$15 billion per year," nor has it locked in the lease to be executed until 2029.
Source: Public Information
ABAB AI Insight
Wood's historical trajectory is: early bets on Tesla were correct, later raising target prices far from transaction ranges; ARKK has seen significant declines since its 2021 peak, with five-year annualized returns lagging behind benchmarks. This time she is not using a new story but connecting the monthly rent in the prospectus with the speed of self-built power plants into a recovery table. The numerical basis jumps from Winton's $24 billion revenue/$29 billion construction cost to her stated $50 billion rent/$20 billion construction cost, with the difference being whether it is annualized per gigawatt or total contract price, rated or actual load.
The capital path is for xAI to first build training clusters, rent out unused capacity to competitors, and then integrate with SpaceX for orbital computing. Ground rent transforms the narrative of Colossus's losses into new cloud profits, while orbital data centers remain a backup. Resource mobilization relies on self-installing turbines and cabinets, changing "waiting for the grid for two years" into "generate power first and then supplement permits."
Comparable cases include AWS early on renting excess capacity as cloud, Jensen Huang turning chip shortages into prepaid orders, and independent power producers locking in traditional electricity prices. In terms of industry position, the computing power market has shifted from self-use training to emergency rentals: whoever connects power first gets to collect monthly payments; model labs pay a premium above spot prices for locked-in capacity.
Structural judgment indicates a transfer of pricing power: pricing power temporarily shifts from "whose model is stronger" to "who can deliver gigawatt-level racks within months." The mechanism is that training windows cannot wait for permits, and tenants are willing to pay three to five times traditional recovery logic to buy time; builders apply rocket-like vertical integration to power plants, making speed itself a multiple of rent.
ABAB News · Cognitive Laws
- The first to connect power collects capital expenditures as rent.
- Short-term contracts can also rewrite balance sheets.
- Grid scheduling is someone else's profit window.