The Data Center as the New Factory

Every industrial revolution eventually reveals its real factory floor. For AI, that floor is a windowless building full of server racks, and the capital being poured into it now rivals — and in some readings already exceeds — the scale of the interstate highway system. Amazon, Alphabet, Microsoft, and Meta are on track to spend over $650 billion combined in 2026 alone, up 60% from $410 billion in 2025. That is not a technology budget line item anymore. It is the largest corporate investment cycle in history, and its physical bottlenecks are now the actual story.

The Numbers Behind the Headline

Amazon leads with roughly $200 billion in planned 2026 capex, followed by Alphabet at $175–185 billion, Microsoft at $145 billion annualised, and Meta at $115–135 billion — at the midpoint, Meta's single year of spending exceeds the entire GDP of more than 120 countries. Across six hyperscalers including Oracle and the Stargate consortium, more than $690 billion has been committed to 74 new data centre projects spanning 28 US states, targeting 37 gigawatts of capacity by 2030 and a projected 4.7 million jobs. Hyperscalers now spend 45–57% of revenue on capital expenditure, up from just 10–15% in 2020 — a structural shift in how these companies allocate capital, not a temporary surge.

Why the New Factory Is Also the New Grid Problem

This letter connects directly to Letter 28 (The Grid): the physical electricity network was already under strain before AI arrived, and data centres have made the mismatch acute. A single AI-optimised server rack now draws 30 to over 100 kilowatts, against 5–15 kilowatts for traditional infrastructure. Nearly half of planned US AI data centres face delays tied to a 7-gigawatt capacity gap, with transformer and switchgear lead times stretching to five years — a hard physical constraint no amount of capital can simply buy its way past on a faster timeline. A data centre can go from groundbreaking to operational in 9 to 12 months. A new power plant takes two to five years. That mismatch is the actual investment thesis.

The Nuclear Answer Nobody Expected This Fast

The clearest sign of how physically constrained this has become: Constellation Energy is restarting Three Mile Island Unit 1 — the same plant partially shut down after America's most famous nuclear accident — under a $1.6 billion project backed by a $1 billion Department of Energy loan and a 20-year power purchase agreement with a single customer, Microsoft. Meta has separately signed its own nuclear power purchase agreement, and Amazon has expanded its nuclear offtake commitments. This is not an ESG preference anymore. It is an operational requirement for firm, 24/7, carbon-free baseload power that intermittent renewables cannot supply at the density AI training demands.

Where the Actual Moat Sits

The companies that already control permitted, grid-connected capacity hold a structural advantage that new capital cannot quickly replicate. Equinix operates more than 260 permitted, grid-connected data centres — a position analysts describe as a "five-year-to-replicate physical moat" given current interconnection queue times. That is the real lesson of this cycle: in a power-constrained world, existing grid connections are worth more than existing chips, because chips can be manufactured faster than substations can be permitted.

Nvidia's Position, Restated Honestly

Nvidia's $197 billion in data centre revenue remains the most visible number in this entire cycle, and deservedly so — but it is also the easiest part of the story to already have priced in. The physical infrastructure layer this letter focuses on — grid capacity, cooling systems, permitted real estate, nuclear power purchase agreements — is less discussed, less obviously "AI" to a casual observer, and consequently less likely to already be fully reflected in every relevant stock price.

The Honest Risk: A Revenue Gap Nobody Has Closed Yet

Bain's own modelling is the most important caveat in this letter: sustaining the current investment trajectory requires roughly $500 billion in annual spending to generate approximately $2 trillion in revenue — a four-times revenue multiple on invested capital that has not yet been demonstrated at this scale, by anyone, in this cycle. Morgan Stanley separately notes that 21% of S&P 500 companies now cite AI benefits in earnings calls, but markets are increasingly discriminating between companies citing AI and companies actually monetising it. The technology sector faces an estimated $1.5 trillion in new debt issuance over 2025–2027 to fund this buildout — a genuine leverage question sitting underneath the excitement, not a footnote to it.

The Human Cost Nobody's Modelling Into the Stock Price

Local communities in Georgia, Indiana, Missouri, and Washington have already imposed moratoriums on new data centre grid connections, citing noise, water usage, and the optics of reserving scarce power capacity for servers instead of housing or industry. This is a genuine, underpriced political risk to the buildout timeline — not because any single community can stop a hyperscaler, but because enough local moratoriums, compounding across enough of the 28 states already hosting projects, could meaningfully slow the very capacity expansion this entire investment thesis depends on.