Tech Intelligence · AI Infrastructure · October 13, 2026

One AI Data Center, About $30 Billion. Most of It Is the Chips.

Bloomberg’s Oct. 12 breakdown of a hypothetical one-gigawatt AI data center in Abilene, Texas puts the bill at roughly $30 billion. About $17 billion of that is chips, according to a headline summary of the Bloomberg item on Techmeme. Power infrastructure is about $3 billion. Construction of the building itself is about $2.4 billion.

Bloomberg’s full feature is paywalled, so every figure here is as reported by secondary sources and labeled that way. Where outside estimates disagree with Bloomberg, we say so.

~$30B — estimated cost to build and equip one 1 GW AI data center in Texas · Source: Bloomberg, via Techmeme and Business Report
~$17B — of that is chips, per Techmeme's summary; hardware is about half of total project cost · Source: Techmeme; Business Report
~$750B — planned 2026 capital spending by the four largest U.S. hyperscalers · Source: Bloomberg, via Business Report
~$3B — estimated cost of power infrastructure for the same site, against about $2.4B for construction · Source: Bloomberg, via Techmeme

§ 01 / Where the Money Goes

The land is almost an afterthought. Hardware is the story: server racks and compute equipment make up about half of the total, per Business Report’s read of the Bloomberg analysis, with power generation, cooling, fiber, construction and engineering adding billions more.

Keep Reading

AI · Safety & Governance

Anthropic and OpenAI Want to Embed Safety Evaluators. Can They Really Be Independent?

Read →

Bloomberg Podcasts — The Staggering Costs of a Single Data Center, From Land to Chips

§ 02 / Nobody Agrees on the Total

Outside estimates for a gigawatt run higher than Bloomberg’s. Bernstein estimates about $35 billion per gigawatt, which Investing.com notes is well below the $50 to $60 billion figure Nvidia gave on its fiscal Q2 2026 earnings call. Epoch AI, via Voronoi, says $37.2 billion with $21 billion in chips. An NBER paper by Stijn Van Nieuwerburgh puts it near $41 billion.

Cartoon satire: a cutaway of a data-center building showing a worker in a hard hat dwarfed by towering rows of glowing server racks above an underground level of blue cooling pipes and orange and red power cables set in brown soil — A cutaway of the idea: the racks tower over the plumbing and wiring beneath them. The drawing is illustrative, not to scale.

The summaries we could see do not spell out each estimate’s scope, and Bloomberg’s $30 billion sits below the low end of the outside range. Treat any single figure as a range.

§ 03 / Who Pays for the Power and the Debt

The grid is where the cost leaves the balance sheet. PJM’s independent market monitor found data centers accounted for $29.4 billion of $63.6 billion in capacity costs, or 46%, across the last four base auctions, and its Joseph Bowring said the only way to shield other customers is to run a separate auction, per Utility Dive.

Financing is the other pressure. On April 28, 2026, Bloomberg reported data-center-linked bonds slid after a report that OpenAI missed internal targets. Broadband Breakfast reports roughly $1.8 trillion of hyperscaler capex through 2028, about $600 billion more than the four generate in cash.

Keep Reading

AI · Antitrust

Nvidia Paid $20 Billion for a Company It Didn't Buy. The DOJ Wants to Know Why.

Read →

§ 04 / The Bottom Line

Bloomberg’s breakdown makes one point plainly: in a modern AI data center the buildings and the land are the cheap part, and hardware is the expensive part. Paperjam reports that memory, storage and silicon prices have soared. Whether the revenue arrives to match the spending is the open question, and no one can answer it yet.

What We Could Not Verify

Bloomberg’s full line-item table. The feature is paywalled and secondary summaries do not reconcile.

Bloomberg’s comparison of the $750 billion to national GDP, which we saw only in a feature excerpt, so we left it out.