Big tech will spend $725 billion this year. Generative software made $25 billion.

Amazon, Microsoft, Alphabet, and Meta guided roughly $725 billion in combined 2026 capital expenditure, up 77% from about $410 billion in 2025. Add Oracle's separately guided $50 billion and the five-company figure clears $700 billion by any measure you pick. Every dollar of it is chasing a market that, by Bank of America's count, generates something like $25 billion a year in direct generative-software revenue today.

Twenty-eight to one.

That is the ratio the entire data-center site-selection industry is currently underwriting against.

How did the number get this big, this fast?

By raising guidance four times in six months. In December 2025, industry trackers pegged combined hyperscaler capex for 2026 above $600 billion, a 36% increase over 2025. By February 2026, reported guidance for the five largest US cloud and compute infrastructure providers had moved to $660 billion to $690 billion: Amazon at $200 billion, Alphabet at $175 billion to $185 billion, Microsoft at $120 billion or more, Meta at $115 billion to $135 billion, Oracle at $50 billion. By May, the four hyperscalers alone (excluding Oracle) had pushed their combined figure to roughly $725 billion, with Meta specifically citing higher memory-chip prices as one reason it raised its own range. Analysts now project the combined figure topping $1 trillion in 2027.

Guidance that rises every time a company reports is not a forecast. It is a company telling you, quarter by quarter, that it committed to more than it originally planned, and that the plan before that one was also wrong.

What does the market think about all this spending?

It got scared, briefly and specifically. Amazon, Microsoft, Nvidia, Meta, Google, and Oracle shed a combined $1.35 trillion in market capitalization in the week ending February 6, 2026, after Amazon's $200 billion capex guide (56% higher than the prior year) landed ahead of market expectations. Mamta Valechha of Quilter Cheviot summarized the mood shift as investors going from "the fear that you cannot be last" to questioning, in her words, "every single angle" of the entire buildout. Michael Field of Morningstar put it even more starkly: "The bet is becoming binary. Either a big pay off if these investments come good, or a huge waste."

The arithmetic behind that nervousness is not subtle. Bank of America calculates that hyperscaler capex now consumes roughly 93% to 94% of operating cash flow after dividends and buybacks, up from about a third in 2023. Direct generative-software revenue across the industry runs around $25 billion, against $600 billion to $700 billion in infrastructure spend. That is the 28-to-1 ratio this piece opened with, and it is the number every optimistic capex headline quietly steps around.

Hyperscaler capex now consumes roughly 93 to 94 percent of operating cash flow after dividends and buybacks. In 2023 it consumed a third.

Where does the debt show up?

Everywhere the cash flow doesn't reach. Morgan Stanley expects hyperscaler borrowing to top $400 billion in 2026, more than double 2025's $165 billion, and JPMorgan's CMBS research desk projects data-center securitization volume reaching $30 billion to $40 billion a year in both 2026 and 2027, up from about $27 billion in 2025. Capital that used to come from cash on the balance sheet is now coming from bond desks and structured-credit shelves, which changes who is exposed if the spending does not convert to revenue on schedule.

Oracle is the clean stress case, because its balance sheet shows the industry's pattern in exaggerated form. The company's remaining performance obligations reached $523 billion, up $68 billion in a single quarter and up 438% year over year, with 40% to 60% of that backlog reportedly tied to OpenAI's roughly $300 billion, five-year Stargate commitment. Against that backlog, Oracle posted negative free cash flow of about $10 billion in a single quarter, with cumulative negative free cash flow projected at $60 billion to $70 billion through fiscal 2029. Credit-default-swap spreads on Oracle debt widened from roughly 50 basis points in September 2025 to more than 130 basis points three months later, the highest since 2009, even as Moody's, S&P, and Fitch held official ratings at Baa2/BBB with negative outlooks. A bondholder lawsuit landed on January 14, 2026. Oracle's stock was down more than 40% from its September 2025 peak by the time that suit was filed.

None of that means Oracle collapses. It means the market is already pricing a scenario where the backlog does not convert to cash on the schedule everyone modeled, for at least one company in the group, and that company's exposure is concentrated in a single counterparty five years into a much longer set of lease obligations.

Haven't we seen this shape before, at smaller scale?

Yes, and recently. Microsoft announced $1 billion across three Licking County, Ohio sites in October 2024 and paused all three by April 2025, not because the entitlement failed, but because the anchor thesis underneath the timing (Intel's semiconductor megafab) slipped to at least 2030. The capital was real, the announcement was real, and the demand assumption underneath it was wrong on a corporate timeline nobody in Licking County controlled. Scale that dynamic up from one county's $1 billion to an industry's $700 billion and the mechanism does not change. What changes is how much of the economy notices when a piece of it repriced.

The counterargument deserves a fair hearing, because it is not baseless. Crusoe's Abilene campus shows real revenue-generating capacity landing on schedule: Phase 1 went live on Oracle Cloud Infrastructure within fifteen months of construction start, backed by an $11.6 billion financing round and, eventually, a $15 billion joint venture targeting roughly 400,000 GPUs. Capacity built against signed offtake and delivered on time is a different risk profile than capacity built against a demand forecast, and a meaningful share of the $700 billion is the former. The honest read of this cycle is that both things are true inside the same number: some of it is Abilene, funded and delivered against real contracts, and some of it is Licking County, funded against a story that has not yet had to prove itself.

Watch who is buying that securitized paper, because it tells you how the risk is being repackaged rather than removed. Data-center-backed structured credit is landing in pension funds and insurance-company portfolios that traditionally buy investment-grade paper, not venture-style compute risk, on the strength of ratings built from the same lease and offtake contracts underwriting the physical campuses. That flow only works if the underlying tenants perform on schedule. It is the same hyperscaler-revenue bet this whole piece describes, just wearing a bond rating instead of an equity ticker.

So where does that leave a site-selection call?

Sorting which category a given site falls into is now the actual job, more than the zoning question this platform was built to answer. A campus backed by a signed, multi-year hyperscaler lease with contracted offtake behaves like Abilene: real revenue, real schedule, real risk of execution delay rather than demand collapse. A campus riding a corporate capex announcement with no signed tenant, in a market whose thesis depends on someone else's fab or someone else's model-training roadmap, behaves like Licking County, and the difference will not show up in the zoning file. It shows up in whether the backlog behind the site belongs to a company whose own cash flow can carry it.

My call: at least one of the five companies named in this piece meaningfully cuts its own 2027 capex guidance, in absolute dollars rather than just growth rate, before the end of calendar 2027, and Oracle is the most likely candidate given the concentration in its backlog. I expect the cut to get framed as "capital discipline" rather than a demand miss, the way Microsoft framed Licking County as project-specific rather than thesis-wide. If all five hyperscalers instead raise 2027 guidance again with no pullback anywhere in the group, the supercycle argument beat the skeptics clean, and I will have underestimated how deep this demand actually runs. Watch the guidance, not the ribbon-cuttings. The guidance moves first.

This analysis is a source-cited research summary drawn from public records, not legal advice. It can contain errors and should be verified independently before any investment decision.