The AI buildout is starting to outspend Big Tech's cash

Information Technology 4 min 80 sources

Hyperscalers are on track to spend more on AI than they earn, and investors have stopped cheering the buildout and started asking what the money buys.

Key takeaways

Big Tech's AI spending is on track to top the cash these companies generate by 2027 — a first — and investors have stopped giving the buildout a free pass.
The tell that returns are thinning: new flagship models now match older ones at half the cost, so the race has moved from "who's smartest" to "who's cheapest."
The buildout's real-world bill is electricity — US data centers may use a fifth of the nation's power by 2035, and those costs reach ordinary ratepayers, not just shareholders.

This week’s Big Tech earnings turned on one number: capital spending, or capex — the money a company sinks into buildings, chips, and data centers. Alphabet, Google’s parent, told Wall Street it will spend even more on AI than planned [1]. The market’s reaction showed a shift in mood. Investors who once cheered the buildout now want to see what it earns [8].

Alphabet spends more, and its cash flow turns red

Alphabet raised its 2026 capex forecast by $15 billion, to a range of $190–205 billion, and said spending will climb again in 2027 [11]. The heavier investment pushed its second-quarter free cash flow — the cash left after running the business and paying for capex — to negative $5.8 billion [11]. It was the first negative quarterly reading in the company’s history [11]. The quarter was otherwise strong: Google Cloud revenue jumped 82% from a year earlier [11]. Jim Cramer, the CNBC host, said he was not comfortable with the level of spending [11].

The whole industry is crossing the same line

Alphabet is not alone. The five biggest US “hyperscalers” are the giant cloud operators: Microsoft, Alphabet, Amazon, Meta, and Oracle. Together they are on track to spend more on capex than they generate in free cash flow by 2027, a Reuters analysis found [1]. The gap is stark. Those firms are expected to add about $340 billion in annual operating cash flow between 2025 and 2027, but to lift capex by roughly $534 billion [1]. That is about $1.57 of new investment for every $1 of new cash flow [1]. Oracle’s capex in its 2026 fiscal year already reached 174% of its operating cash flow [1].

The bond market gets nervous

The unease is showing up where these companies borrow. Yields rose this week after Alphabet lifted its forecast, and bond investors are demanding more reward to lend to Google, Amazon, and Meta — their credit spreads are widening [13]. Oracle’s five-year credit default swap is a form of insurance against the company failing to pay its debts. It is trading at a multi-year high, and has become the market’s go-to gauge of AI-debt fear [13]. Analysts at the bank Mizuho wrote that the spending is testing investor limits, and that firms “once seen as capital fortresses” now face a sharp rise in AI-tied costs [13]. Some of the debt is hard to see. Filings from the five giants revealed off-balance-sheet debt of $1.65 trillion — up eightfold in four years — as they race to build data centers, Nikkei reported [51]. Morgan Stanley and Moody’s flagged the same issue [51].

The frontier itself is flattening

Here is the deeper reason the returns question bites now: the models are converging. Anthropic released Claude Opus 5 this week, which it says delivers nearly all the capability of its top model, Fable 5, at half the cost [80]. The company was explicit that Opus 5 is not its smartest model, and that the race is shifting “from raw capability to the economics of daily use” [80]. Microsoft launched its own in-house models and claimed they cut costs up to 89% versus OpenAI’s, arguing it no longer needs to lean on the frontier [46]. The Register put the industry’s quiet truth plainly: open models, Chinese or not, are competitive now [28]. When the best model is only a little ahead of the cheap one, “best” stops being worth what it cost.

The counter-case: Apple

Not everyone is on the treadmill. Apple has spent far less on AI infrastructure. Its shares led the “Mag7” — the seven biggest US tech stocks — this week, and some investors now treat it as a hedge against hyperscaler spending [34][35]. The investor Lyn Alden summed up the mood: AI demand is real, but profit is the biggest open question [74].

The bill lands on the grid

The spending has a physical form: electricity. US data centers are expected to use about one-fifth of the country’s power by 2035, four times today’s share, according to BloombergNEF [7]. AI compute is pushing that capacity toward 200 gigawatts [7]. That estimate is 83% higher than the same firm’s forecast in December [7]. The strain is already real. In the grid region stretching from Virginia to Illinois, data centers are set to draw 34% of electricity, and one utility has threatened to pull out [7]. The cost of that power is one reason capex keeps rising — and one reason the bill will reach ordinary electricity customers, not only shareholders [13][7].

02 · Lesson · why it matters

Why pouring in more buys less near the top

Fast progress looks like a straight line up — until each new dollar buys less than the last, and the curve quietly bends flat.

A company outspent its own cash

Alphabet just spent so much building AI that, for the first time in its history, its cash flow went negative. And the sharpest question in tech shifted. For three years it was: how good can these machines get? This week it became: what is all this money actually buying? That is the same question, asked from the far side of a hill.

Every technology climbs the same shape

Progress in a new technology almost never runs in a straight line. It traces an S. A slow start while the thing barely works. Then an explosive middle, where each year is a leap. Then a plateau, as the technology bumps against a ceiling and the leaps shrink to inches.

The trap is the middle. From inside the steep part, the curve looks like a rocket — straight up, forever. So everyone extrapolates the straight line and bets on it. But the tell that you have crossed into the plateau is not that progress stops. It is quieter than that: each new unit of effort buys less than the one before it.

The gains are already shrinking

Watch the models. The jump from one generation of AI to the next used to be a leap you could feel. Now the frontier is bunching up. Anthropic’s newest model matches its best one at half the price. Microsoft says its in-house models do the job for a fraction of a rival’s cost. Open models anyone can download are, as one trade outlet put it, competitive now.

When the cheap option does nearly everything the expensive one does, that is not a footnote. That is the shape of the curve telling you where you are. The race has moved from “who is smartest” to “who is cheapest” — and a race about price is a race that has run out of altitude to gain.

Spend climbs with ambition; returns fall with height

Here is the mechanism that makes this dangerous, and it is worth holding onto. Your spending grows with your ambition. You always want the next level, and the next level costs more than the last. But your returns grow with where you already sit on the curve — and near the top, that is almost nothing.

So the ratio inverts. You pay more and more for less and less. The plainest measure of it this week: across the biggest firms, every extra dollar of cash flow is now arriving alongside about $1.57 of new investment to chase it. That is not a rounding error. That is the plateau, written in a balance sheet.

The bet was made far from where the cost lands

It is tempting to file this under “a few CEOs spent too much.” But look at who is actually inside the web. The bond market, demanding more to lend as the debt piles up. The pension funds quietly holding that debt. The power grid, bracing for AI to draw a fifth of the nation’s electricity. The ordinary household whose electricity bill rises as data centers crowd onto the same wires. And the reader, whose retirement fund is heavier in these seven companies than they likely realise.

A decision made in a boardroom does not stay in the boardroom. It travels down the wire to a stranger who was never asked, and the reader is often that stranger.

The law that was never a law

Underneath all of it sits an assumption that came to feel like physics: that adding more computing power reliably buys more capability. It even had a name — scaling laws. But it was never a law of nature. It was a pattern observed during the steep middle of the curve, and mistaken for a promise about the whole climb. Building a $250 billion bet on that pattern treats a temporary slope as a permanent one. Someone chose to read the curve as a rocket. That reading served the people selling the climb.

What the curve won’t tell you

The hard part is this: from inside the curve, no one can see exactly where it bends. A plateau only looks like a wall once you are past it. The people spending two hundred billion dollars cannot see the bend either — they are standing on the same slope, reading the same ambiguous line. Knowing that a curve flattens is not the same as knowing when. The most honest thing anyone on that hill can say is that they are climbing in fog, and the higher they go, the less each step is worth. That is worth remembering the next time a straight line up is offered as a sure thing.

03 · Lab · your turn

The Spend Dial

Rehearse the diminishing return on AI spending -- climb an S-curve and feel each new billion buy less than the last.

04 · Hope · carry this

A plateau is not a dead end — it is the shape of a new thing settling into ordinary life. The moment the frontier stops racing ahead is the moment its gains stop belonging only to the richest, and start reaching everyone.