El Paso is getting a $14 billion data center, courtesy of Meta and BlackRock. The city council heard hours of testimony from residents worried about groundwater, about air pollution from the gas plant being built to power the campus, and about why a desert city was subsidizing one of the richest companies on earth.¹ The council approved the deal anyway. A recall petition followed. When one councilmember later moved to cancel the agreement, city lawyers said it could not be undone.²

Even a cursory review of the project’s terms explains why El Pasoans are angry. Meta gets an 80 percent abatement on city and county property taxes, $12.5 million in city funds for roads, and a water agreement allowing up to 1.5 million gallons per day.³ Construction employs about 4,000 people at its peak. The finished campus will employ about 300 (while running 813 permitted generators).

City officials answer that, when completed, the campus will become the district’s largest taxpayer, contributing $15 million a year even under the sweetheart terms Meta and BlackRock secured.⁴ Taken at face value, that is a reasonable argument: the land was publicly owned and paid no property taxes, and a completed campus adds tax revenue and construction activity to a region seeking private investment. The perspective also conveniently ignores that the tax revenue arrives only if the project is completed and used as planned. The demands on land, water, roads, and the electric system arrive first and stay local if the business plan changes. Meta, BlackRock, and their lenders can refinance, sell, or leave. El Paso cannot move its aquifer, its grid, or its residents.

While the data center debate grows across campaign ads, internet discourse, and American towns, there is little debate among the tech executives racing to pour unprecedented sums of capex into the AI build-out. McKinsey estimates the global compute build-out may require $6.7 trillion of capital through 2030, including $5.2 trillion for AI workloads.⁵ Meta, Alphabet, Microsoft, and Amazon alone have outlined roughly $725 billion of capital spending and finance leases for 2026, and together the four hyperscalers account for roughly 18 percent of the S&P 500.⁶ Maintaining that market share requires defending the most profitable parts of their business: advertising and the cloud.

Four companies, four moats to defend

Meta generated $201 billion of revenue last year; $196 billion of it, 98 percent, came from advertising.⁷ AI already sharpens the recommendations that keep users inside Facebook and Instagram and the ad systems that monetize their attention. Meta is defending its control of the consumer interface and trying to make sure the next interface does not belong to somebody else.

Alphabet generated $403 billion; roughly $295 billion came from advertising, most of it search.⁸ An agent that answers a question directly bypasses the sequence that pays for all of it: query, results page, click, monetized destination. That makes Alphabet the clearest case of an incumbent whose most profitable product could be reorganized by AI, even as it builds a fast-growing cloud business.

Microsoft’s cloud segment grew 32 percent in the latest quarter; the widely noted 4 percent decline came in More Personal Computing, a segment of Windows devices, Xbox, and search advertising, rather than consumer software as a whole.⁹ Agents may cut the number of software seats a task requires, but the spending shifts toward compute, models, and usage-based services instead of disappearing. Microsoft can lose per-seat revenue and still capture more through Azure and its agent platforms.

Most of Amazon’s revenue is retail. Most of its profit is not: AWS produced 57 percent of 2025 operating profit on 18 percent of sales.¹⁰ Amazon’s spending is less a defense of retail than a bid to supply the infrastructure that the whole transition runs on.

What unites the four is a battle over the layer between users and everything they want to do. An agent can choose the search result, select the product, write the document, and place the order. Whoever controls that layer can redirect advertising, software, and commerce, and the ads are already following users into answer engines.¹¹

One early warning came in April 2025, when Google searches through Safari fell for the first time in 22 years. Apple’s Eddy Cue told a court that “People are using AI,” and Alphabet lost about $150 billion of market value that day. One channel and one month did not establish a decline, but the market understood the possibility, and Google has since rebuilt much of its search experience around AI answers.¹²

Alphabet is defending search; Meta, attention; Microsoft, software distribution; Amazon, cloud infrastructure. Each has arrived at the same fork: disrupt your own business model or be replaced by the frontier models. Meanwhile, the defended revenue is still growing. Search rose 17 percent last quarter; Azure grew 43 percent; AWS grew 37 percent; and the two big cloud backlogs hold a trillion dollars of signed contracts.¹³

Insurance, and possibly overinvestment

Amazon will spend $220 billion in capex this year. Alphabet: up to $205 billion, double what it spent in 2025. Microsoft: about $175 billion. Meta: up to $145 billion.

One reading says all this capacity answers insatiable demand. Another says it is a Potemkin build-out with no economic purpose. The too-frequently ignored reality is that the eye-watering figures cost much less than the alternative. In the words of Alphabet’s Sundar Pichai: “the risk of underinvesting is dramatically greater than the risk of overinvesting.“¹⁴ Mark Zuckerberg concedes the industry may be overbuilding, and spends anyway, because losing means missing “the most important technology for the next 10 to 15 years.”

Satya Nadella has pointed the industry to the railroad bubble of 1873. The historian he cites, Liaquat Ahamed, sees the parallel plainly: “Everyone has decided they want to be in the railroad business.” Nobody’s plan, he adds, accounts for everyone else’s.¹⁵

This is insurance thinking: each company’s budget is set by its competitors’ budgets rather than by realistic forecasts of the total addressable market. The TAM claims have inflated alongside the capex. SpaceX cited a $28.5 trillion market ahead of its June IPO, and Anthropic is reportedly pitching more than $30 trillion, a quarter of global GDP.¹⁶

The insurance can be individually rational. If a $100 billion investment lowers the odds of losing a trillion-dollar franchise, the campus does not need an attractive stand-alone return; it preserves position, locks up scarce power before a rival does, and buys an option on future demand. Rational decisions by the individual can still produce collective excessive capacity, without fraud or incompetent management. And when that capacity requires a city to reserve scarce water, power, and tax capacity, a private strategic hedge becomes a public commitment.

The financing behind the build-out has also turned circular: the chipmaker funds the customers who buy its chips. In Ohio, Nvidia has agreed to backstop a single campus with up to $105 billion, against the $150 to $200 billion it expects to earn there as the exclusive supplier.¹⁷ The Bank for International Settlements counts roughly $1.65 trillion of hyperscaler debt sitting off balance sheet in special-purpose vehicles. Michael Burry, who made his name shorting the subprime bubble, reads the pattern bluntly: “almost all customers are funded by their dealers.”¹⁸

Built for the race, not the load

The four hyperscalers have signed more than $1.5 trillion in data center leases. The build-out will approach 2.8 percent of American output, a bigger share than the railroads reached at their peak.¹⁹

The danger to watch is the wrong compute in the wrong place under the wrong contract. DeepSeek and Moonshot’s Kimi models show how fast model economics can change; cheaper inference can invite more use, not less, but it changes which chips, cooling systems, and locations are worth paying for.²⁰ A data center is not one asset. The land, grid connection, and shell may last decades while the accelerators inside go obsolete in a few years. A well-located powered shell can be reused; a specialized campus built around a weak tenant or an uneconomic power contract may not be. The railroad analogy holds because infrastructure races overbuild. It breaks because compute gets replaced and repurposed at very different speeds.

Whatever gets built, the hyperscalers have largely offloaded their financial exposure onto other market participants: construction risk to bondholders, load risk to utilities, capacity cost to ratepayers, and the generators to community residents. From here, the build-out has more than two endings. Demand may stay strong and justify the spending. Demand may grow while prices fall and returns disappoint, a profitable technology inside a poor investment cycle. Or a slowdown may strand projects and expose weak counterparties. The middle outcome may be the most likely, and investors can diversify against either tail risk. A host community, unfortunately, cannot, and nobody knows what the bad tail looks like up close. As one capital markets analyst notes, there is simply no playbook for restructuring a bankrupt data center. ²¹

Terms communities should demand

The principle should be simple: the party making the forecast should pay for being wrong. No tax abatement, public road, water reservation, or expedited permit should rest on a nonbinding promise from a thinly capitalized project company. The tenant or a creditworthy parent should guarantee the public obligations. Collateral should cover completion, remediation, and decommissioning. Minimum annual payments should protect the tax base if utilization falls, take-or-pay commitments should keep unused power capacity off other customers’ bills, and water should be priced for local scarcity. Community-benefit agreements should be enforceable contracts, not press releases.

None of this is anti-growth. These are the terms any prudent lender already demands before committing capital to a long-lived, technology-sensitive project, and a developer with a creditworthy tenant and durable demand can provide them. A project that cannot bear a guarantee, a minimum payment, and transparent ownership is not strong enough to justify public subsidy or a long-lived claim on scarce water. Where a signed deal cannot be reopened, as in El Paso, the city still controls permits, utility approvals, expansion phases, and infrastructure decisions. Each one is a chance to secure protections that survive changes in ownership, financing, and technology.

So, who really needs all these data centers? The four hyperscalers do, and their customers already use the capacity. Society’s need is a different question, answered by the value of the services against the full cost of land, water, power, carbon, and public risk, not by how much defensive capacity dominant companies can afford to reserve. Local consent is part of that cost, not a box checked after the financing closes.

Nobody has to stop the build-out. The bargain to refuse is the one where four companies choose the scale and host communities hold the residual risk. The companies asking a town to host the physical consequences of their race should stand behind their forecasts with their own balance sheets. That much discipline is owed to the people who will live with the infrastructure long after the models and the financing have changed.

Note on sources

Figures are drawn from company filings, earnings releases, and the reporting cited in the footnotes. Revenue shares are calculated from reported segment data.

---

Footnotes

1. New York Times (Tabuchi), Aug. 5, 2026: El Paso council record and recall petition; $14 billion Meta/BlackRock campus; 813 permitted generators; approximately 4,000 construction jobs at peak and 300 permanent. ↩
2. El Paso Matters, June 3 and June 9, 2026: opponents’ groundwater and air concerns; the $500 million, 366-MW El Paso Electric gas plant dedicated to the facility; the attempted termination and city officials’ position that the agreement cannot be undone. ↩
3. El Paso Matters, Apr. 3, 2026 (podcast), and Oct. 15, 2025: 80 percent abatement of city and county property taxes; $12.5 million of city funds for road improvements. El Paso Matters, June 3, 2026: water supply agreement permitting average use up to 1.5 million gallons per day, against a citywide average supply of 108 million gallons per day. ↩
4. El Paso Matters, June 9, 2026: city estimate of about $15 million a year in city taxes, projected to make the data center the city’s biggest taxpayer. See note 2. ↩
5. McKinsey & Company, “The cost of compute: A $7 trillion race to scale data centers,” April 2025: $6.7 trillion of projected data center investment through 2030, of which $5.2 trillion for AI workloads. ↩
6. Combined 2026 hyperscaler capital spending and finance-lease guidance of roughly $725 billion, per Financial Times analysis of earnings-call guidance, as reported. S&P 500 weights as of Aug. 2026, per Stock Analysis data compiled by The Motley Fool: Microsoft 5.3 percent, Alphabet 6.3 percent, Amazon 4.2 percent, Meta 2.1 percent. ↩
7. Meta Platforms Form 10-K, fiscal 2025 (filed Jan. 2026); Q4 and full-year 2025 results, Jan. 28, 2026: revenue of $201.0 billion, advertising of $196.2 billion. ↩
8. Alphabet 2025 Annual Report, revenue table: total revenues of $402.8 billion; Google advertising of $294.7 billion. ↩
9. Microsoft FY2026 Q4 results and earnings call, July 29, 2026: Intelligent Cloud revenue of $39.3 billion, up 32 percent; More Personal Computing down 4 percent. ↩
10. Amazon Q4 and full-year 2025 results, Feb. 5, 2026: AWS at 18 percent of net sales and $45.6 billion of $80.0 billion in operating income. ↩
11. OpenAI began selling ads in ChatGPT in early 2026 (announced Jan. 16; live for U.S. free and Go tiers from February), as reported by CNBC. ↩
12. Testimony of Eddy Cue, U.S. v. Google remedies proceeding, May 7, 2025, as reported by Bloomberg and CNBC: first decline in Safari searches in 22 years; Alphabet market value decline of roughly $150 billion. Google disclosures on AI Overviews reach, as reported. ↩
13. Alphabet Q2 2026 results; Microsoft FY2026 Q4 results; Amazon Q2 2026 results: Search revenue up 17 percent, Azure up 43 percent, AWS up 37 percent; AWS backlog of $496 billion and Google Cloud backlog of $514 billion. ↩
14. Sundar Pichai, Alphabet Q2 2024 earnings call. Mark Zuckerberg, Bloomberg interview, July 2024. ↩
15. Liaquat Ahamed and Satya Nadella’s citation of his railroad history: Financial Times, Aug. 26, 2026. ↩
16. SpaceX prospectus, May 2026 ($28.5 trillion total addressable market, $26.5 trillion attributed to AI), ahead of its June 12 Nasdaq debut. Anthropic’s reported TAM above $30 trillion: Wall Street Journal, Aug. 26, 2026; figures not final. For scale, FactSet data cited by the Journal put combined 2025 revenue of the 191 technology companies in the S&P 1500 at $2.4 trillion. ↩
17. Nvidia Form 8-K and Ohio regulatory filing, Aug. 17, 2026: credit support capped at $105 billion; $150–200 billion of expected revenue as exclusive supplier. ↩
18. Michael Burry, X post, Nov. 2025, as reported by Benzinga, Nov. 20, 2025. Bank for International Settlements Annual Report estimate of roughly $1.65 trillion in hyperscaler off-balance-sheet debt through special purpose vehicles, as reported, Aug. 2026. ↩
19. Hyperscaler leases above $1.5 trillion and Stijn Van Nieuwerburgh (Columbia Business School) estimate of 2.8 percent of economic output: Financial Times, Aug. 26, 2026. ↩
20. DeepSeek R1 and the July 2026 semiconductor selloff: Financial Times, Aug. 26, 2026. Moonshot AI’s Kimi K3: Yahoo Finance, July 27, 2026. ↩
21. Advait Arun, Center for Public Enterprise, to the Financial Times, Aug. 26, 2026. ↩