the edit, vol. 36
the $1 trillion bet nobody voted on
On June 1, Alphabet announced it would raise $80 billion by selling stock — the largest equity offering in the company's history. The capital will fund data centers, chips, and the infrastructure required to run its AI models at scale. Alphabet's 2026 capital expenditure guidance is now $180 to $190 billion. That is roughly double what it spent in 2025, which was itself double what it spent the year before.
Investors reacted by selling the stock.. The market's interpretation was not that Google had found a new engine of growth. It was that Google had concluded it could no longer fund its own future from the cash it generates — and needed to dilute its shareholders to keep pace.
Alphabet is not alone.
Microsoft will spend approximately $190 billion on AI infrastructure this year.
Amazon, $200 billion.
Meta, $135 billion.
Together, the four hyperscalers are expected to spend over $700 billion on AI capital expenditure in 2026.
Goldman Sachs projects that figure will exceed $1 trillion in 2027 and reach $1.6 trillion annually by 2031 — implying $7.6 trillion in cumulative AI infrastructure spending between now and then.
These numbers are so large they function like abstractions. The question worth asking — the one that is almost entirely absent from the financial coverage of these announcements — is a governance question: who decided that the world's infrastructure should be rebuilt around artificial intelligence, who is bearing the cost, and what happens if the bet doesn't pay off.
what is actually being built
To understand the scale of what is being decided, it helps to understand what AI infrastructure actually is.
The core unit is a data center — a warehouse-scale facility filled with specialized chips, connected by hundreds of thousands of kilometers of cabling, cooled by industrial-scale liquid systems, and powered by dedicated electrical infrastructure. Each facility requires between 100 and 500 megawatts of power — enough to supply a small city. NVIDIA's current generation GPU rack, the GB300 NVL72, packs 72 processors into a single unit. A modern hyperscaler data center may contain thousands of racks.
A United Nations University report published this week quantified what scaling AI to utility level demands from the physical world. In 2025, data centers consumed 448 terawatt-hours of electricity globally — more than Saudi Arabia uses in an entire year. AI accounted for roughly one-fifth of that total. Data centers also used 4.5 trillion liters of water for cooling and generated 189 million tons of carbon dioxide.
These figures are from 2025. The spending plans now being announced will roughly double the infrastructure that produced them. Goldman Sachs's baseline model projects $765 billion in annual AI capital expenditure in 2026 alone. The physical world — the electricity grids, the water systems, the land, the supply chains for rare earth materials in chips — will absorb the consequences of decisions made in quarterly earnings calls and investor presentations.
the returns question nobody can answer
The financial case for this spending rests on a single premise: that AI will generate returns sufficient to justify the investment. It is a reasonable premise. It is also, at this point in history, unverified.
The Goldman Sachs research team published a notable analysis last year asking whether AI could deliver on its promise. Their finding: AI investment is real and accelerating, but the productivity gains that would justify it at scale have not yet materialized in the economic data. The gap between what is being spent and what is being earned is, for now, enormous.
This is not unusual for transformative infrastructure. The U.S. railroad buildout of the 19th century consumed capital at rates that bankrupted investors repeatedly before the returns arrived. The telecommunications buildout of the late 1990s produced the dot-com collapse when revenue growth failed to keep pace with infrastructure spending. The survivors — the companies that had built real infrastructure rather than just speculative capacity — eventually generated the returns. The investors who had financed the speculation did not.
The current AI buildout differs from the telecom bubble in important ways: the underlying technology demonstrably works, the demand from enterprise customers is documented, and the companies doing the spending are profitable enough to absorb the losses if the returns are delayed. But it resembles the telecom era in one structural way: the assumption that whoever builds the most infrastructure will capture the most value has driven spending to levels that may exceed what the market can absorb, and the cost of being wrong will be distributed broadly — across shareholders, electricity grids, water systems, and labor markets — in ways that the decision-makers will not personally bear.
the governance gap
The question that financial journalism rarely asks about the AI buildout is a political one: who authorized this?
No legislature voted to rebuild the world's electricity infrastructure around AI data centers. No democratic process determined that $7.6 trillion in capital should be directed toward this particular technology over the next five years rather than toward climate adaptation, housing, healthcare, or education. No international body assessed the water consumption implications for communities near data center campuses, or the grid stability implications for regions where hyperscalers are concentrating their facilities.
These decisions are being made by a small number of executives at a small number of companies, ratified by investors whose primary interest is return on capital, and announced in earnings calls that are covered as financial news rather than as infrastructure policy.
Sam Altman said this week at a BlackRock summit that AI will eventually be sold like electricity and water. His analogy is that most people don't generate their own electricity — they connect to a grid and pay for what they use. AI is heading the same way. The problem, as Altman himself acknowledged, is that if OpenAI can't build enough compute to meet demand, "the price gets really high" — which pushes AI access toward the wealthy or forces governments to decide how to distribute limited compute.
That is a description of a public utility problem being solved by private capital allocation. When electricity became a public utility in the 20th century, the process involved regulatory bodies, rate commissions, public hearings, and legislative authorization. The decision about who controlled the grid and at what price was treated as a matter of democratic governance because it was understood to affect everyone.
The AI infrastructure buildout is proceeding on the assumption that private capital allocation is sufficient — that the market will sort out access, pricing, and distribution. That assumption has not been tested at utility scale, and the governance frameworks that would test it do not yet exist.
what Berkshire's $10 billion means
One data point from this week's announcement deserves particular attention: Berkshire Hathaway agreed to invest $10 billion in Alphabet as part of the equity raise.
Warren Buffett's successor, Greg Abel, authorized the largest single technology investment in Berkshire's history to anchor a share offering for a company he apparently concluded needed the credibility of Berkshire's participation to get the deal done. The fact that Alphabet — one of the most profitable companies in human history — needed Warren Buffett's imprimatur to raise $80 billion without sending its stock into freefall is itself a signal worth reading carefully.
Berkshire's investment thesis has historically been simple: buy companies with durable competitive advantages, predictable cash flows, and honest management, and hold them. The AI infrastructure bet is not that thesis. It is a bet on a technology transition whose timeline, economics, and competitive structure are genuinely uncertain. The fact that Berkshire made it anyway suggests either that Abel has concluded the transition is more certain than the market currently prices, or that the risk of not participating in the infrastructure buildout is greater than the risk of the investment itself.
The second interpretation is the more interesting one. It implies that the AI buildout has reached a scale where opting out is no longer a conservative option — that the infrastructure is now so central to the future of the economy that a company whose entire philosophy is capital preservation has concluded it cannot afford to stand aside.
the physical world's invoice
The financial questions around AI infrastructure are the ones that dominate coverage. The physical questions are the ones that will matter most to the most people.
Data centers require water — large amounts of it, for cooling systems that run continuously. A single large data center can use between one and five million gallons of water per day. The regions being targeted for data center development — the American Southwest, parts of Europe, Southeast Asia — are not uniformly water-rich. In Arizona, where multiple hyperscalers are building major facilities, water rights conflicts between data centers, agriculture, and residential users are already active legal disputes.
Data centers require electricity — and the scale of new demand is straining grids that were not built for it. In Virginia, which hosts the largest concentration of data centers in the world, Dominion Energy has warned that new data center connections are creating grid stability risks. In the UK, National Grid has flagged similar concerns. The electricity demand from the AI buildout is one of the primary reasons global carbon emissions from the electricity sector, which had been declining, are now projected to rise again through the end of the decade.
These are not arguments against AI infrastructure. They are descriptions of costs that are real, measurable, and being borne by people who did not participate in the decision to impose them. The communities near data center campuses, the ratepayers whose electricity bills reflect grid upgrades, the farmers competing for water rights — they are the external costs of a capital allocation decision made in a boardroom.
the question the earnings calls don’t ask
There is a version of this story in which the AI infrastructure buildout is the most consequential and beneficial investment in human history — the construction of a new kind of utility that will compound economic productivity for generations, reduce the cost of expertise and access to knowledge, and ultimately justify every dollar spent on it many times over.
There is another version in which it is the largest speculative capital misallocation since the dot-com era — driven by competitive fear rather than verified demand, sustained by investor momentum rather than fundamental economics, and destined to produce a correction that will leave overbuilt data centers, strained electricity grids, depleted water tables, and diluted shareholders as its legacy.
The honest answer is that nobody knows which version is correct, including the people making the decisions. What is known is that the scale of the bet is extraordinary, the governance framework for making it is essentially nonexistent, and the costs of being wrong will be distributed far beyond the companies doing the spending.
Alphabet raised $80 billion this week to build infrastructure for a technology whose returns are still being established, the market sold the stock, and Berkshire bought in anyway.
The world is being rewired. The question of who decided to rewire it, and on whose behalf, is not one that quarterly earnings calls are designed to answer.
