Google Prints Money — So Why Did It Just Sell $85 Billion of Stock? (And Why Buffett's Successor Bet Big)

An episode of Dan's AI Intel

The largest equity raise in history, the $10 billion AI bet Greg Abel made that Warren Buffett never would, and the $600 billion question hanging over the whole build-out.

Published · Updated · By Dan Walter

Executive summary

The company that prints money just had to go out and raise some — and the most disciplined investor on Earth wasn't the one who bought in. In early June 2026, Alphabet, Google's parent and one of the most cash-generative businesses in history, sold roughly $85 billion of its own stock, the largest equity raise any company has ever done, to help pay for artificial intelligence. The single most telling line in the deal is who wrote the anchor check: Berkshire Hathaway put $10 billion into a private slice of it — but this was no longer Warren Buffett's call. Buffett stepped aside as CEO on January 1, 2026 and is now chairman only; the man who made the bet is his successor, Greg Abel. Abel pointed Berkshire's famously tech-averse, capital-allergic empire straight at the most capital-hungry build-out in corporate history — the exact kind of treadmill business Buffett spent sixty years avoiding. Buffett's own verdict on the move was a shrug of admiration: "I never talked to the CEO. He has launched."

Read those two facts together — a cash machine selling stock, and a post-Buffett Berkshire buying the thing Buffett wouldn't touch — and you have the real story of where the AI revolution stands, far more honestly than any model benchmark tells it. The revolution has crossed a line from a software story into a physical-capital story — concrete, copper, transformers, and silicon poured into the ground at a scale that even the richest companies on Earth can no longer comfortably self-fund. The four biggest US tech firms are on track to spend something like $600 billion on data centers and chips in 2026 alone, nearly double 2025; Morgan Stanley sees roughly $3 trillion of data-center build-out by the end of the decade, and McKinsey's most-cited scenario runs to $6.7 trillion by 2030. That is no longer a technology budget. It is an industrial mobilization on the scale of building a new electricity grid, and it is why Alphabet, sitting on a mountain of cash, chose to dilute its own shareholders rather than wait.

The hinge the entire thing turns on is brutally simple, and nobody can yet prove which way it falls: will the compute pay for itself? The same money that validates the bet — Abel's $10 billion, Nvidia's data-center business now running at over $300 billion a year — sits inside a gap that Sequoia's David Cahn pegged at $600 billion between what's being spent on AI infrastructure and the revenue needed to justify it, while an MIT study found 95% of enterprise AI pilots returning no measurable profit. The validation and the warning are the same fact: the build-out is now so large that it has become a genuine bet, and the new steward of the most patient money on the planet just decided the odds were good enough to take — on a bet his legendary predecessor passed on for a lifetime.

Why a finance deal is the clearest window into the AI revolution

If you want to know how real the AI revolution is, ignore the demos for a moment and follow the money — specifically, follow who is willing to part with theirs. Capital is the most honest signal we have, because it is the one thing that cannot be faked with a good keynote. A model can be cherry-picked; a benchmark can be gamed; a launch can be hype. A $10 billion private placement from Berkshire Hathaway cannot. It is a cold, considered judgment, by the people least likely to be swept up in a mania, about whether the physical machinery being built to run AI will earn its keep.

And that machinery is the part of the story that has quietly become the whole story. For its first few years, generative AI was a software phenomenon — clever weights, a chat box, a sense of magic. In 2025 and 2026 it became something heavier and more consequential: an infrastructure project. The constraint on AI today is not ideas or even talent. It is whether you can pour enough concrete, secure enough electricity, and buy enough chips, fast enough. That is why Alphabet's finance team, not its research lab, produced the most revealing AI artifact of June 2026. The deal is a confession in numbers: demand for AI compute now outruns even Google's ability to pay for it out of pocket. Understand the deal, and you understand the shape of the revolution — its astonishing scale, its physical cost in power and water, and the single unresolved question of whether any of it pays off.

What Alphabet actually did — and why it wasn't a bond deal

Start with precision, because this deal has been widely garbled. In early June 2026, Alphabet launched an equity raise at $80 billion, upsized it at pricing to $84.75 billion as demand blew past expectations, and — with underwriters exercising their options — grew it to roughly $90 billion in total. By size, it is the largest equity offering ever completed. The crucial word is equity: Alphabet sold ownership in itself, not debt. This is not the same event as the separate roughly $20 billion bond sale Alphabet did back in February 2026 (which, with sterling and Swiss-franc tranches, swelled toward $32 billion and even included a 100-year sterling bond). Conflating the two is the single most common error about this story. February was borrowing; June was selling stock.

The June raise came in four distinct pieces, and the structure is itself the message:

  • Common and capital stock — roughly $20–21 billion of Class A common stock and Class C capital stock sold outright in an underwritten public offering.
  • Mandatory convertible preferred stock — around $19 billion in depositary shares representing 6.25% mandatory convertible preferred, a hybrid that pays a dividend now and automatically turns into common stock by 2029. It is a way to raise equity while softening the immediate dilution hit.
  • A $40 billion at-the-market (ATM) program — a standing facility to dribble Class A and Class C shares into the open market over time, expected to begin in the third quarter of 2026. Much of this is earmarked to cover the taxes Alphabet owes on employee stock awards, not the build-out itself.
  • A $10 billion private placement to Berkshire Hathaway — the headline. $5 billion of Class A common stock at $351.81 a share and $5 billion of Class C capital stock at $348.20 a share.

The common stock and the Berkshire placement closed on June 4; the convertible preferred a day later. Alphabet's stated use of proceeds was disarmingly plain: general corporate purposes, "including capital expenditures to scale AI infrastructure and global compute." Management's own framing was that demand for its AI products is "meaningfully exceeding our available supply." That is the entire deal in one sentence — they cannot build fast enough, and money is now the bottleneck.

Exhibit — Alphabet's record $85B raise came in four pieces — not a bond among them. Selling stock is the most expensive money a profitable firm can raise — Alphabet bought speed over cost. Source: Alphabet equity-raise press release (2 Jun 2026); Cleary Gottlieb deal summary. Compiled by Dan's AI Intel.

The deeper "so what" is the choice of equity over more debt. Alphabet had raised over $100 billion in debt in the prior year; it could have funded much of this build-out from its own cash flow, just more slowly. Selling stock — diluting existing owners — is the most expensive money a profitable company can raise, and the most telling. You do it when speed matters more than cost, and when you've already leaned hard on the debt markets. The price of that speed shows up in the numbers analysts care about: Alphabet's free-cash-flow margin is set to compress from roughly 18% in 2025 toward something near 5% in 2026 as the capital pours out the door. Alphabet looked at the AI race and decided that being slightly late was more dangerous than being slightly diluted.

Why Abel's first big bet is the tell — not Buffett's

Now the part that should stop you, and it is a detail almost everyone gets wrong. This was not Warren Buffett's deal. Buffett handed the CEO job to Greg Abel on January 1, 2026, remaining only chairman and Berkshire's largest shareholder. The $10 billion Alphabet placement was Abel's — the boldest opening move of his tenure, made in the same week he spent $6.8 billion to acquire homebuilder Taylor Morrison, roughly $16.8 billion of cash deployed in days. Buffett's reaction was telling precisely because it was hands-off: "Greg did that faster than I could have done it, smoother than I could have done it, and I never talked to the CEO. He has launched." The man who built the most famous capital-discipline machine in the world stood back and watched his successor point it somewhere he never would have.

Because Berkshire's entire philosophy, honed over six decades, is to avoid businesses that must constantly reinvest huge sums just to keep running — what Buffett derides as capital-intensive treadmills. AI data centers are the purest such treadmill ever built. Berkshire stayed almost entirely out of AI through all of Buffett's tenure. So the real signal isn't "Berkshire likes AI." It is that the post-Buffett Berkshire, under a new CEO five months into the job, just made the single largest technology bet in the firm's history — into the exact kind of capital-hungry expansion the old Berkshire was built to shun. The succession and the strategy arrived together.

Exhibit — Abel's first big bet aims Berkshire at the capital-treadmill Buffett shunned for 60 years. The most patient capital on Earth, under new management, judged this build-out worth owning. Source: AP; Fortune (Jun 2026). Compiled by Dan's AI Intel.

Why would Abel do it? Two reasons, and both matter. First, he isn't betting on the data centers as a standalone business; he is betting on Alphabet, a company that already throws off enormous cash from search and ads and is using AI to defend and extend that moat. The Google Cloud backlog alone — contracted, signed-up future demand — sits above $460 billion. To Abel, this looks less like a speculative tech flyer and more like a dominant cash machine that happens to be spending heavily, at an attractive price. As one analyst put it, you do not write a $10 billion check into a dilution you expect to regret.

Second — and this is the structural genius for both sides — Berkshire got a discount. The placement was priced roughly 6% below where Alphabet's shares closed the day before the deal was announced. That discount is the price Alphabet paid for certainty: a single, deep-pocketed, permanent shareholder who anchors the raise, signals confidence to everyone else, and won't flip the stock next quarter. Berkshire, in return, got a blue-chip position at a built-in margin of safety — exactly the kind of asymmetric, low-downside entry the Buffett playbook prizes, applied by his successor to an asset Buffett himself avoided. Both sides won, which is why the deal got done. The signal to read is colder and more interesting than "Buffett loves AI": the most patient capital on Earth, under new management, judged that this build-out, at this price, is worth owning.

The build-out the money is actually for

So what is the $85 billion — and the trillions behind it — buying? Numbers first, then a picture, because the scale is genuinely hard to feel.

The four hyperscalers — Amazon, Microsoft, Alphabet, and Meta — spent a record of roughly $388 billion in capital expenditure in 2025. For 2026, their combined guidance points to something like $600–630 billion, and some analyst tallies run as high as $725 billion; add Oracle and the five-firm total reaches toward $700–805 billion. Roughly three-quarters of it is AI-specific. Zoom out and the projections turn vertiginous: Morgan Stanley sees about $3 trillion of data-center spending through the end of the decade (and warns of a $1.5 trillion financing gap to fund it); McKinsey's central estimate is $6.7 trillion of data-center capital expenditure by 2030, of which around $5.2 trillion is for AI-ready facilities. For comparison, the entire US Interstate Highway System cost, in today's dollars, a few hundred billion. We are talking about building the equivalent of several Apollo programs, every year, in private capital.

Exhibit — The four hyperscalers' AI build-out leaps from $388B to ~$630B in a single year. About three-quarters is AI-specific — a technology budget has become an industrial one. Source: Company guidance compiled by Datacenterrichness; McKinsey (2026). Compiled by Dan's AI Intel.

Where does the money go inside a data center? The mix is roughly: about 60% on the IT itself — chips and servers, overwhelmingly Nvidia GPUs — with another quarter or so on the unglamorous but essential power and cooling guts (transformers, switchgear, uninterruptible power supplies, chillers), and the remainder on the building shell, land, and networking. Bank of America's cleaner cut puts it at about 74% IT and 26% physical infrastructure. The headline is that this is, first and foremost, a silicon purchase order — which is why a single company, Nvidia, sits at the center of the whole economy.

Exhibit — Roughly three-quarters of every data-center dollar buys silicon, not buildings. The build-out is first a silicon purchase order — which is why one firm, Nvidia, sits at its centre. Source: Bank of America; Epoch AI (2026). Compiled by Dan's AI Intel.

And the buildings are no longer buildings. They are campuses the size of small cities, with their own names and their own gravity. OpenAI's Stargate site in Abilene, Texas, brought its first phase online in late 2025 at 1.2 gigawatts, headed toward more than 450,000 of Nvidia's top GPUs. Meta is building Hyperion in Louisiana — a $27 billion campus on 2,250 acres, starting at 2 gigawatts and scaling toward 5 — and Prometheus in Ohio, among the first gigawatt-scale AI sites. Elon Musk's xAI runs Colossus in Memphis, racing toward the equivalent of well over a million high-end GPUs. Microsoft's Fairwater is the largest single site publicly known, estimated to need 3.3 gigawatts and to cost north of $100 billion. Amazon's Project Rainier is among the first wave of gigawatt campuses. To put "a gigawatt" in human terms: it is roughly the output of a full-size nuclear reactor, dedicated to one building full of chips.

The "so what" of the scale is this: AI is no longer competing for software engineers. It is competing for electricity, land, water, and transformers — physical inputs with physical limits — and that competition is the new center of gravity in the entire technology industry.

The power and water bill the world will actually pay

This is the part casual coverage skips, and it is where the build-out collides with the real world. The IEA estimates data centers consumed about 415 terawatt-hours of electricity in 2024 — roughly 1.5% of global power — and projects that to roughly double to about 945 terawatt-hours by 2030, with AI as the primary driver. The local concentration is the shocking part: in the United States, data centers are on track to account for nearly half of all electricity demand growth between now and 2030. An industry that was a rounding error on the grid a few years ago is about to become the single biggest reason the grid has to grow at all. That is why these companies are now signing deals to restart shuttered nuclear plants and to build their own power.

Exhibit — Data-center electricity use is set to more than double by 2030, to 945 TWh. In the US, data centers make up nearly half of all electricity-demand growth this decade. Source: IEA, Energy and AI (2025). Compiled by Dan's AI Intel.

Then there is water, the quietly contested front. Cooling all those chips takes enormous volumes of it. Google alone reported its data centers consumed about 6.1 billion gallons of water in 2024 — and a single large campus can use as much water as a small city, which is precisely why communities from Arizona to Georgia are starting to push back. But here the honest reporting has to flag a real fight over the numbers. Per-query water figures vary wildly — OpenAI's Sam Altman has cited 0.3 milliliters of water per ChatGPT query, while UC Riverside researchers measured 10–25 milliliters — a gap of more than 50 times. The difference is not dishonesty on either side; it is scope. Altman's figure counts only the water evaporated on-site for cooling; the researchers' figure also counts the water consumed upstream by the power plants generating the electricity. Both are technically correct; which one matters depends on the question you're asking. Newer, more efficient chips have also genuinely cut the per-unit cost. The right posture on AI's water footprint is neither panic nor dismissal, but precision about what is being measured — a discipline that has been notably absent from the public argument, with Altman himself dismissing water concerns as "fake."

The implication that ties power and water together: the binding constraint on AI's growth in the second half of this decade is unlikely to be algorithms or even chips. It will be megawatts and gallons — and the patience of the towns asked to supply them.

The depreciation time bomb hiding in the accounting

Here is a fight that sounds like accounting trivia and is actually a multi-hundred-billion-dollar question about whether Big Tech's profits are real. When a company buys a GPU, it doesn't expense the whole cost at once; it spreads it over the chip's estimated "useful life." Stretch that life longer and reported profits look bigger today. Hyperscalers generally assume their AI chips last about five to six years. Michael Burry — the investor of The Big Short fame — and short-seller Jim Chanos argue that is fantasy: frontier AI chips, they say, are functionally obsolete for cutting-edge work in two to three years, because Nvidia now ships a faster generation every year. Burry estimated the hyperscalers were collectively understating depreciation by roughly $176 billion across 2026 through 2028 — which is to say, overstating their profits by that much.

Exhibit — Big Tech books AI chips over 6 years; critics say they're obsolete in 3. Burry's math: ~$176B of 2026–28 'profit' may be borrowed from the future. Source: M. Burry; J. Chanos; National Law Review (2026). Compiled by Dan's AI Intel.

The rebuttal is sharper than it first appears, and it turns on a distinction the critics elide. Nvidia and operators like CoreWeave argue that a chip aging out of frontier training doesn't stop earning money — it cascades down to inference and less-demanding workloads, where it can run profitably for years. CoreWeave points to five-year customer contracts and the fact that older A100 and H100 chips still resell at high prices. The skeptics conflate the technological obsolescence cycle (18–36 months at the bleeding edge) with the economic utility cycle (five-plus years across all uses). Jensen Huang's own joke — that when his new Blackwell chips ship in volume, "you couldn't give Hoppers away" — is fuel for both sides: vivid evidence that frontier chips age fast, and a sales pitch for buying the newest.

The calibrated read: the accounting is probably not fraud, but it is genuinely optimistic, and a chunk of today's reported AI profitability rests on an assumption — six-year chip lives — that has not yet been tested through a full cycle. If Burry is even half right, some of the earnings underpinning these sky-high valuations are borrowed from the future. That is the kind of crack that doesn't matter at all until, suddenly, it matters enormously.

Does any of it pay off? The $600 billion question

Everything above — the deal, the trillions, the power and water, the chips — bottoms out in one question, and it is unresolved. Is the compute earning its keep?

The bull case is concrete and growing. Nvidia's data-center business is now running at over $300 billion a year — annualizing recent quarters puts it north of $325 billion — which is the clearest evidence that someone is buying the picks and shovels in staggering volume. On the demand side, the AI labs are scaling fast: OpenAI disclosed roughly $25 billion in annualized revenue in early 2026, and Anthropic claimed it passed that, reaching about $30 billion annualized by April 2026, up from $9 billion just months earlier. Real money is now flowing to real products. Google's $460 billion cloud backlog says enterprises are signing long, serious contracts.

The bear case is just as concrete. Sequoia's David Cahn framed it as "AI's $600 billion question": take the run-rate of AI infrastructure spending, account for the full cost of the data centers and the margins the chip-buyers need to clear, and you get a roughly $600 billion gap between what's being invested and the end-user revenue required to justify it. Someone, eventually, has to pay enough for AI products to fill that gap. And on the ground, the payoff is patchy: an MIT study found that 95% of enterprise generative-AI pilots delivered no measurable return — not because the models are bad, but because companies struggle to integrate them into real workflows. (The figure is contested; critics note the study used a narrow definition of success and that informal, individual use of tools like ChatGPT delivers value it doesn't capture.) Even the revenue claims have asterisks: OpenAI's own revenue chief reportedly argued Anthropic's $30 billion figure was overstated by some $8 billion. In an industry this young, even the scoreboard is disputed.

Exhibit — Does the compute pay off? Both cases are concrete — and the verdict is still open. The deciding variable is time-lag — revenue must grow into the gap before patience, and the depreciation clock, run out. Source: Sequoia (D. Cahn); MIT / Fortune; DCD; Epoch AI (2026). Compiled by Dan's AI Intel.

So which is it — the most important infrastructure of the century, or the most expensive bubble in it? The honest answer is that both stories are running at once, and the deciding variable is time lag. Cahn's own conclusion is not "bubble" but "a gap that revenue has to grow into" — the railroads and the fiber-optic boom both over-built ferociously, ruined many of the builders, and yet left behind infrastructure that powered decades of growth. The question is not whether AI is useful; it plainly is. The question is whether it becomes useful enough, fast enough, to justify $600 billion a year before the patience of investors — and the depreciation clock on all those chips — runs out.

Bottom line

The Alphabet–Berkshire deal is the AI revolution telling on itself — and on a changing of the guard. A company that prints cash had to sell stock; a post-Buffett Berkshire, under new CEO Greg Abel, paid to get in at a discount on the very kind of capital-hungry business Warren Buffett spent a lifetime avoiding; the proceeds vanish into campuses that drink power like cities and chips that may age faster than the accountants admit. Every part of that sentence is simultaneously a reason for confidence and a reason for caution, and that is the truest thing anyone can say about this moment: the build-out is now so vast that it has become a real bet, and the new steward of the smartest money on the planet just decided the odds were good enough to take.

Watch three hinges from here. First, revenue versus the gap: does AI product revenue keep roughly doubling, year over year, and start closing Cahn's $600 billion gap — or does it stall while the capital keeps pouring out? Second, the physical ceiling: power and water, not algorithms, are the likeliest things to cap the build-out — watch which campuses get stranded waiting for a grid connection. Third, the depreciation reckoning: the first full cycle of those six-year chip-life assumptions is coming, and if Burry is right, a slice of today's reported AI profits will quietly evaporate. Abel told us the smart money — under new management — is in. The next two years will tell us whether it was early genius or expensive faith.

Sources