Anthropic's $45 billion compute deal works out to $16 million a megawatt, per year. The building costs $11 million, once

Anthropic just agreed to pay Nscale roughly $45 billion over six years for 460 megawatts of AI computing capacity. Divide that out and it comes to about $16.3 million per megawatt, per year. JLL's 2026 data center outlook puts the average global cost to actually build a data center's shell and core at $11.3 million per megawatt — once. Anthropic's annual lease payment alone is bigger than the entire one-time construction budget. That gap is the number worth bringing into your next hyperscale pricing conversation, not the headline dollar figure.

What exactly did Anthropic agree to?

Anthropic will pay Nscale, a UK-based AI infrastructure developer, about $45 billion over six years for 460 megawatts of capacity running on Nvidia's upcoming Vera Rubin chips. The capacity sits inside Nscale's Monarch Compute Campus, a roughly 2,250-acre site outside Point Pleasant in Mason County, West Virginia, that Nscale acquired in March 2026 and is building toward as much as 8 gigawatts of total capacity long-term, with an initial 2-gigawatt phase targeted for early 2028. Reporting on the Anthropic agreement points to this specific 460MW block coming online around the end of 2027. It's the latest in a string of massive compute commitments Anthropic has made this year, including a separate roughly $10 billion deal for capacity in Norway, as the company reportedly works through more than a dozen letters of intent with data center developers.

Why run the math per megawatt instead of looking at the $45 billion headline?

Because the total dollar figure tells you Anthropic is spending a lot of money; it doesn't tell you what that spending means for the people who actually pour the concrete and pull the cable. Breaking the deal down to a per-megawatt, per-year basis makes it comparable to construction cost benchmarks that estimators already use.

The physical building — the part a GC, electrical sub, or mechanical contractor actually bids — gets paid for in less than nine months of Anthropic's lease payments. Over the full six-year term, the lease is worth roughly 9 times the average one-time shell/core construction cost, and still 2.5 to nearly 5 times the cost of a fully built, AI-ready facility with liquid cooling and dense electrical distribution.

What should an estimator actually do with that comparison?

• Stop discounting contingency to look competitive on hyperscale bids. If a data center owner's compute-lease economics dwarf the construction budget by an order of magnitude, a few extra points of contingency for tariff exposure or trade-labor escalation is not going to sink the deal — and shaving it to win the job transfers real risk onto your company for no reason tied to the owner's actual ability to pay.
• Price change orders against the owner's economics, not your fear of losing the account. A $5–10 million cost increase on a several-hundred-million-dollar campus build is immaterial next to a $45 billion compute commitment. Owners and their construction managers know this; contractors negotiating change orders should too.
• Treat the compute lease term as a proxy for how locked-in the owner is. A six-year, tens-of-billions commitment signals the tenant has no realistic walk-away option if the building slips a quarter. That's leverage for negotiating schedule-driven cost recovery, not just a reason to hit the date.
• Separate "what the building costs" from "what the deal is worth" in every client conversation. Owners will talk chip supply, IPO timing, and compute contracts. Bring the construction-cost comparison back to the table so contingency and change-order conversations stay anchored to actual buildout numbers, not the size of the press release.

One more thing worth naming: Anthropic filed a confidential IPO registration in June 2026, and investors are reportedly targeting an October 2026 offering near a $2 trillion valuation — the largest IPO ever, if it happens. That doesn't change what 460 megawatts costs to build. It does mean the owner side of these deals has every incentive to hit dates and very little incentive to fight a contractor over contingency that's a rounding error against the lease.

For more on how this compute boom is reshaping data center construction risk, see our look at what Nvidia's $1.3 trillion 2027 capex forecast means for bidding the backlog.

Friday one chart. Every week, one piece of data that should change a decision on your project. Subscribe at constructionaibrief.com.