Alphabet's second quarter brought 119.8 billion dollars in revenue and 82 percent growth at Google Cloud. It also brought 44.9 billion dollars of capital spending in ninety days, a raised full-year budget of as much as 205 billion, and free cash flow that went negative. The CFO told analysts Google is now renting other companies' data centers to keep up.

Somewhere in Alphabet's second-quarter slide deck, past the advertising charts that Wall Street reads first, sits the number that actually describes what changed at Google this year.

Its first-party model APIs served roughly 22 billion tokens per minute during the quarter. Three months earlier the figure was 16 billion.

That is a jump of roughly 38% in a single quarter of raw inference demand, measured not in users or subscriptions but in the thing engineers actually pay for. It is the clearest public signal yet of how fast production AI workloads are scaling inside one company.

The next slide showed what serving them costs.

The Bill Came to $44.9 Billion in Ninety Days

Alphabet reported second-quarter results on July 22. Revenue reached 119.8 billion dollars, up 24% year over year and roughly 2.8 billion ahead of consensus. Google Cloud grew 82% to 24.8 billion dollars, beating estimates by more than two billion, and its operating margin expanded to 35.6% from 20.7% a year earlier.

Then came capital expenditure: $44.9 billion in one quarter, more than double the year-ago figure.

Management told analysts that roughly 60% of technical infrastructure spending went to servers, meaning AI accelerators including Google's own TPUs, with the remaining 40% going to data center construction and networking gear.

The company raised full-year 2026 capital expenditure guidance to a range of 195 billion to 205 billion dollars, up from the 180 billion to 190 billion range it gave last quarter. First-half spending already totaled 78.6 billion, which means the back half has to average roughly 58 billion to 63 billion dollars per quarter to land inside the new range.

Management also signaled that 2027 capex will increase significantly from there.

Free Cash Flow Went Negative

Metric	Q2 2026	Comparison
Revenue	119.8 billion dollars	Up 24% year over year
Google Cloud revenue	24.8 billion dollars	Up 82% year over year
Cloud operating margin	35.6%	20.7% a year earlier
Cloud backlog	514 billion dollars	Up more than 50 billion sequentially
Capital expenditure	44.9 billion dollars	More than double year-ago level
Free cash flow	Negative 5.9 billion dollars	Positive 10.4 billion a year earlier
First-party API throughput	22 billion tokens per minute	16 billion in the prior quarter

Free cash flow for the quarter came in at negative $5.9 billion, against positive 10.4 billion a year earlier. On a trailing twelve-month basis it fell 20% to 53.3 billion dollars.

Alphabet is not short of money. It closed the quarter with 242.5 billion dollars in cash, cash equivalents and short-term marketable securities. It has also been raising more: long-term debt climbed to roughly 98 billion dollars after new note issuances, and a June offering of Class A, Class C and mandatory convertible preferred stock brought in 49.6 billion dollars of net proceeds.

The point is not solvency. The point is that the most cash-generative advertising business ever built now spends more on infrastructure in a quarter than it produces in operating cash flow.

Investors noticed. Shares fell in after-hours trading on July 22 and then dropped more than 6% the following day, according to Yahoo Finance, despite revenue clearing consensus by roughly 2.8 billion dollars.

Reported diluted earnings per share came in at $9.11, against the 2.88 or so analysts were modeling. Almost the entire gap came from a 99 billion dollar net gain on equity securities rather than from operations. On an adjusted basis, earnings per share was 2.85 against 2.89 expected, according to LSEG. The headline number is not a measure of this business.

Google Is Renting Other Companies' Data Centers

The most revealing disclosure was not a number at all.

CFO Anat Ashkenazi told analysts that Alphabet is operating in a "supply-constrained environment" and that the company will use third-party data center capacity as a bridge in the third quarter, a move she said will create "modest margin pressure" for Cloud alongside the integration of the recently acquired Wiz security platform.

Google owns one of the largest private compute fleets on the planet. It designs its own accelerators. It is now leasing someone else's racks because it cannot build fast enough.

That constraint has been visible from the outside for a while. Power interconnection queues, not chip supply alone, have become the practical ceiling on how quickly capacity comes online, which is why regulators moved to fast-track grid connections for AI data centers earlier this year. Alphabet also told analysts that most revenue from TPU system sales will land in 2027 rather than 2026, pushing the payback on a meaningful slice of current spending into next year.

What This Means If You Buy Compute

The demand behind that throughput number is concentrated in serious users. Alphabet disclosed that more than 500 Cloud customers have each processed over a trillion tokens in the past year, and more than 2,000 enterprises have consumed at least 100 billion tokens apiece.

Three consequences follow directly for anyone buying compute.

Quota stays tight. A supplier that describes itself as supply-constrained and is renting third-party capacity is not about to loosen GPU and TPU allocation for mid-sized customers.
Cloud margins have a headwind. Rented capacity costs more than owned capacity. Alphabet has told the market to expect that to show up in Q3, which is not usually the quarter a vendor discounts aggressively.
The backlog is the real forecast. Cloud backlog reached 514 billion dollars, growing more than 50 billion sequentially, with roughly half expected to convert to revenue within 24 months. That is demand already under contract, and it is the strongest argument that this spending is not speculative.

The tokens-per-minute figure is the metric to track from here. If it keeps compounding at anything close to 38% a quarter, the capex line stops looking reckless. If it flattens while spending accelerates, the arithmetic turns ugly quickly.

The Other Side

Bulls have the stronger near-term case, and it deserves a fair hearing.

Cloud margin expansion from 20.7% to 35.6% in twelve months is not the profile of a business burning money for market share. It is the profile of one that has found pricing power. The backlog supports that reading. So does the fact that revenue accelerated to 24% growth from 14% in the same quarter last year, which is unusual at Alphabet's scale.

The bear case is about timing and quality of earnings rather than demand. Reported profit was flattered enormously by equity gains that have nothing to do with search or cloud, and the adjusted number underneath them came in below what analysts expected. Funding the build has started to show on the balance sheet too: long-term debt near 98 billion dollars, and 49.6 billion raised in a single June offering. Alphabet is committing to a spending curve management has already said will rise again in 2027, at a moment when it has been struggling to ship its flagship Gemini model on schedule.

The pattern is not unique to Google. Every large cloud provider has reported the same shape this year: revenue beats, then a stock decline driven by the capex line. That dynamic has been building since hyperscaler spending began outrunning operating cash flow at Oracle and elsewhere.

The Bottom Line

Alphabet grew revenue 24%, nearly doubled cloud margins, booked more than half a trillion dollars of contracted cloud demand, and still ended the quarter with less cash than it started.

That combination is the entire AI economy compressed into one earnings report. Demand is real and measurable, right down to 22 billion tokens a minute. The infrastructure required to meet it is expensive enough that even the company with the best cash-generating machine in corporate history has to borrow, rent, and wait until 2027 for a chunk of the return.

Ashkenazi described the environment as supply-constrained. The word carries more weight than it looks like it should. It means the bottleneck on AI in 2026 is no longer ideas or models or talent. It is concrete, transformers, and grid interconnects, and none of those respond to a raised guidance range.