AI Spending Has Changed Orders of Magnitude: $725B Across Four Firms, $500B for Stargate

At some point the AI-spending conversation gained an extra digit. Just a few years ago, hundreds of millions made headlines; now the unit is billions, and tens of billions per company barely raises an eyebrow. The 2026 capital-expenditure (capex) guidance disclosed by the largest tech firms makes that shift vivid.

Combined capex at Microsoft, Alphabet (Google), Amazon and Meta is set to swell to roughly $725 billion in 2026. Since the four spent about $410 billion combined in 2025, that is a roughly 77% jump in a single year—most of it flowing into AI data centers, GPUs and power infrastructure.

This is not just an upbeat investment story. When money on this scale moves all at once, it reshapes supply chains spanning semiconductors, power, construction and real estate—and surfaces side effects like surging memory prices. Below, we map the overall picture, then look at why the giants are betting this hard and where the risks lie.

The reality of "$725 billion" in 2026

First, a sense of scale. According to reporting, the four firms' 2026 capex guidance breaks down roughly as follows (all 2026 outlook).

The reality of 2026 is several companies each planning close to $200 billion in a single year—a sum on the order of a nation's annual budget, and an annual figure at that. Goldman Sachs expects cumulative capex for these four hyperscalers from fiscal 2025 to fiscal 2030 to reach about $5.3 trillion. Investment in AI infrastructure has begun running not as a passing boom but as a multi-year "construction project."

$500 billion in a single project—the sheer size of Stargate

Alongside the individual-company capex, the most emblematic effort is "Stargate," led by OpenAI, Oracle and SoftBank. Announced in January 2025, the plan calls for $500 billion and 10 gigawatts (GW) of compute capacity built across the US over four years. This is not one company's capex—it is the figure a new company, created solely to build AI infrastructure, has set out on its own.

Since then, starting in Texas, five new sites have been added, lifting planned capacity to about 7 GW and over $400 billion across three years (as announced in September 2025). OpenAI and Oracle have also struck a partnership for up to an additional 4.5 GW worth more than $300 billion over five years. SoftBank itself has pledged $30 billion of investment in OpenAI, of which $19 billion is earmarked for Stargate. Because securing electricity is the binding constraint, the players have begun investing in power generation and transmission in parallel with the data centers.

Why are the giants betting this hard?

The reason for a gamble this size comes down to one thing: compute has become a precondition for competing. The performance of generative AI depends heavily on the volume of compute poured into training and inference. The smarter a model gets, the more compute it needs; the more users it has, the more that demand swells. Compute itself has become the scarce resource that will decide the next contest for dominance after search and cloud.

For each firm, this is offense and defense at once. For cloud providers, AI demand is the growth engine of the core business; fail to secure capacity here and customers go to a rival. So they move to lock in capacity ahead of demand they cannot fully forecast. Even when the spend looks excessive, they cannot stop—this structure is what pushes them all to floor the accelerator simultaneously.

The emerging risks—"memory inflation" and cash drain

Spending on this scale carries a backlash, though. One is component-price inflation. As AI servers soak up demand, prices for memory chips and other components have climbed, inflating the spend itself. Microsoft CFO Amy Hood explained that $25 billion of the capex outlook stems from rising memory and component costs. Money spent does not translate one-for-one into capacity; the slice eaten by price increases is no longer negligible.

The other is pressure on cash. The giants have funded investment from ample operating cash flow, but as capex approaches $700 billion, shrinking free cash flow and greater reliance on debt become real. Indeed, Stargate has reportedly scrambled to arrange massive financing, and the question of "whether the demand to recoup this investment really continues" is raised repeatedly among investors.

What business readers should watch

So how should you read this enormous investment wave for your own business? First, the benefits of AI-infrastructure spending spread widely to the "backstage" industries—semiconductors, power, cooling, construction and real estate. Equating AI plays with model-developers alone is now too narrow a view. Second, as the supply of compute grows, the cost to those who use AI tends to fall over the medium term. Today's seemingly steep generative-AI fees are likely to decline as capacity expands.

The flip side to watch is what happens if the "continued expansion of AI demand" this investment assumes breaks down. Should demand fall short, the accumulated capacity becomes a glut and depreciation eats into profits. AI-infrastructure investment is a grand growth story and, at the same time, an enormous bet on demand. Rather than being dazzled by the size of the numbers, business readers should read the assumptions—and the payback scenario—beneath them.

Key takeaways

• 2026 capex at Microsoft, Google, Amazon and Meta totals about $725 billion (up 77% year on year), with several firms each near $200 billion. Goldman projects about $5.3 trillion cumulative for the four from fiscal 2025–2030 (all as forecast as of 2026).
• OpenAI, Oracle and SoftBank's Stargate targets $500 billion and 10 GW over four years; added sites lift the plan to about 7 GW and over $400 billion (as of September 2025). SoftBank has pledged $30 billion to OpenAI, $19 billion of it for Stargate.
• The driver is structural: compute is now a precondition for competing. Offense and defense overlap, so the giants all floor the accelerator at once.
• Risks are component-price inflation (Microsoft attributes $25 billion of capex to component costs) and cash drain. Benefits spread widely to backstage industries like semiconductors and power, but a break in AI demand leaves a risk of overcapacity.