Big Tech's AI Spend Just Eclipsed Apollo and Manhattan Combined

Ask anyone to name the largest infrastructure bet in history and you’ll hear Apollo, the Manhattan Project, maybe the interstate highways. In 2026, four companies — Google, Microsoft, Meta, and Amazon — are quietly outspending all of them in a single calendar year. Not with taxpayer money. With shareholder capital.

The numbers, plainly

Combined 2026 capex from the Big Four hyperscalers is tracking around $520 billion. Alphabet near $125B. Microsoft around $140B. Meta close to $110B. Amazon, mostly AWS, near $145B.

Now the comparisons. Apollo cost roughly $25.8B between 1961 and 1972 — about $260B in today’s dollars. The Manhattan Project ran $2B over four years, or roughly $30B today. The Interstate Highway System: about $500B in current dollars, spread across 35 years.

In other words, four companies are spending in twelve months what it took the United States decades to spend putting humans on the moon, splitting the atom, and paving a continent — combined.

Where the money actually goes

Three buckets absorb almost all of it.

First, Nvidia silicon. H100s, B200s, and the Rubin generation that became the 2026 workhorse — tens of thousands of dollars per accelerator, ordered in lots of hundreds of thousands. Meta alone is reportedly spending over $40B on GPUs this year.

Second, land and concrete. Gigawatt-class campuses are rising simultaneously in Abilene, Ashburn, Des Moines, and Phoenix. Single-site construction budgets routinely cross $10B.

Third — and this is the quiet bottleneck — power. Substations, transmission, gas peakers, and increasingly nuclear: Microsoft restarting Three Mile Island, Amazon’s behind-the-meter deal with Talen Energy. The compute story is now an electricity story.

Why this isn’t Apollo

Apollo and Manhattan had finish lines. Land before 1970. Beat the Germans to fission. Win or lose, the project ended.

Hyperscaler AI capex has no such terminus. The compute required to serve a GPT-5 or Claude Opus 4.7-class model grows geometrically each year, and inference demand is now scaling faster than training. What’s being built isn’t a goal — it’s a platform. There’s no version of this story where someone declares it done.

The other structural difference: who eats the loss. Taxpayers absorbed Apollo’s risk. Today it’s four sets of shareholders, which is why every earnings call now opens with some variant of “when does this capex translate to revenue?” If Meta’s Llama doesn’t lift ad efficiency, if Gemini doesn’t monetize search, those hundreds of billions amortize quietly into the income statement over the next six years.

The bubble argument that won’t die

The dot-com fiber overbuild keeps coming up for a reason. Telecoms laid tens of billions of dollars of glass in the late ’90s. Most of it sat dark until the mid-2000s. The counter-argument — that YouTube and Netflix wouldn’t exist without that overshoot — is also still valid. Both things were true.

AI infrastructure has the same shape. Some of the 5-gigawatt campuses breaking ground today will meet a demand air-pocket around 2028. A real secondary market for used GPUs will emerge. Power contracts will get unwound. Useful-life assumptions will quietly shrink from six years to four, and the resulting depreciation hit will land on a few quarters all at once. That’s the moment we’ll know whether 2026 was foresight or mania.

The lingering question

The fact that something larger than Apollo plus Manhattan is now being executed annually, by four private companies, is itself the headline of this decade. The era of state-led megaprojects didn’t end with a bang — it ended with platform companies quietly stepping into the role.

Two questions worth sitting with. If the bet pays off, who captures the productivity gains AI delivers — the platforms that built the rails, or the economy that rides them? And if it doesn’t, does the damage stop at shareholders, or does it cascade through the power grid, the construction industry, and the entire semiconductor supply chain that has reorganized itself around this one customer? The concrete and silicon going up right now will draw the map of the next ten years.