Cipher, Fluidstack and Anthropic Pledge $10 Million for a Data Center Town’s Water

Cipher Digital Inc. (NASDAQ: CIFR), the AI infrastructure builder Fluidstack Ltd. and Anthropic announced on September 14, 2026 that they had committed a combined $10 million to repair and upgrade the water system of Colorado City, Texas. The release, issued by Cipher that afternoon under a Colorado City dateline, says the money will fund projects city leaders identified after this summer’s shortages: in July, storm damage at the city’s well pump station left residents without reliable water for days. The projects it names are additional water storage capacity, replacement of aging valves and other water infrastructure needs. The funds will be managed through the Colorado City Water Infrastructure Foundation, a newly formed nonprofit that the release says is funded exclusively by the three companies and will work with the City Council under Mayor Ruben Hurt, City Manager Donna Madrid and the city’s contracted engineers. Texas House Speaker Dustin Burrows, the mayor and an executive from each company are quoted. Anthropic’s head of economic development, Ryan Wilson, put the premise in one line: “Communities that host AI infrastructure should be better off for it.”

The release makes two claims about the campus, and both are the company’s, not PSV findings. It says the Barber Lake data center runs on its own on-site water supply and draws nothing from Colorado City’s system, so every gallon of added capacity goes to the city, and that the facility’s water use will have no impact on local drinking water sources or residential utility bills. It also says Cipher built Barber Lake to meet the data center standards in Governor Greg Abbott’s directives to the Public Utility Commission of Texas and ERCOT. Cipher’s filings describe the asset. Its quarterly report for the period ended June 30, 2026, filed August 4, says the company signed the Barber Lake lease in September 2025 through its subsidiary Cipher Barber Lake LLC, is building the facility for Fluidstack USA II Inc. with an expected commencement date of September 2026, and collects rent for a 10-year term with two five-year extension options. Google agreed to make Fluidstack’s lease payments if Fluidstack does not, and in exchange Cipher issued Google warrants for 24,178,576 shares at $0.01 per share, with a make-whole if those shares are worth less than $430.0 million when the exercise period opens; Cipher carries that backstop as a $544.5 million asset. The investor presentation furnished with its Form 8-K the same day reports that the tenant has begun beneficial use of the building, including partial occupancy, and that rent is expected to begin in October 2026. Cipher’s website lists Barber Lake at 300 MW and energized, with a further 500 MW Barber Lake Expansion shown for 2030 or later. The filings name Fluidstack as the tenant and Google as the backstop. They do not name Anthropic, and the release does not describe Anthropic’s role at the site, so PSV does not characterize one.

Cipher’s own risk factors explain why a developer would pay for a system its campus says it does not use. The same quarterly report warns that a growing number of state legislatures, county boards, city councils and other local bodies have enacted or are considering moratoria, restrictive zoning amendments and heightened permitting for data centers, driven by concerns that include strain on local electrical grids and water supplies and the perceived limited local economic benefit of these facilities after construction is complete. If that happens where it builds, Cipher says, it may be forced to abandon a project, relocate to a less desirable or more costly site, or absorb unrecoverable costs, and community opposition could affect tenants’ ability to pay rent and the company’s ability to renew or re-lease a data center. The company names water and local benefit as the arguments towns use against data centers, and this pledge answers both in a town where one of its leased AI campuses is weeks from paying rent.

For a real estate reader, the consequence is that host-community spending is becoming a line item in the cost of entitlement and expansion rather than a public relations extra. Massachusetts made the point formally on September 8, when Executive Order No. 658 conditioned state permits for data centers above 25 MW of peak demand on a community benefits agreement. Texas has no such rule and the Colorado City commitment is voluntary, but it lands on the same thing a lender or buyer underwrites. Google’s backstop protects rent on the Barber Lake lease that exists. Nothing in the record protects the 500 MW expansion Cipher shows for 2030 or later, which its investor presentation says is subject to ERCOT’s batch process, and which will be argued in a town that will remember whether this summer’s pledge was kept. For owners of land and industrial property near a campus, the pledge is a signal about the local relationship that governs the next rezoning. Anthropic’s February 11, 2026 commitment points the same way: it said it would pay for the grid upgrades needed to interconnect its data centers, work to cover demand-driven electricity price effects, deploy water-efficient cooling and partner with local leaders, and that it makes those commitments directly where it works with partners to develop data centers for its own workloads. PSV arithmetic, not a company figure: $10 million against Barber Lake’s stated 300 MW computes to about $33,000 per megawatt, though the release ties the money to the city’s needs, not to any capacity figure, and does not say how the three companies divide it.

The useful artifact is a register kept for every data center campus a firm owns, lends against or sits next to, built only from public records and the firm’s own site files. Inputs: the developer’s releases and investor presentations; the lease, backstop and risk factor disclosures in its SEC filings; the host city’s council agendas and minutes and its public water system records; the state directives the developer says it meets; and any moratorium or zoning change on record in the county. Output: one row per commitment, with columns that decide something. The amount, and each company’s share, left blank when undisclosed. The vehicle that holds the money and who controls it, which here is a nonprofit funded by the companies rather than a city account. Whether the commitment is an executed, enforceable agreement or an announcement. The project list, each with the budget, schedule and completion status as the city reports them. The campus’s water source and annual volume, left blank where the developer has not disclosed them. And the expansion capacity whose approval the commitment will be judged against, with its regulatory dependency. Every populated cell cites a document.

An assistant handles the volume well: reading a year of council agendas, matching announced projects to contracts and line items as they appear, flagging a promised project with nothing on record months later, and comparing what a release claims with what the filings say. The reviewer is the development lead or the lender’s diligence counsel, and the approval gate is that no commitment is credited as reducing entitlement or re-leasing risk until it ties to an executed document with an amount and a schedule. The assistant should not decide whether a pledge is adequate, estimate a campus’s water use from its megawatts, or record “no impact on local drinking water sources” as an established fact. PSV ran no model on this record and tested no product.