Datadog reported second-quarter results before Thursday’s open. Revenue was $1.12 billion, up 35.6% year over year and roughly $40 million ahead of a consensus near $1.08 billion. Non-GAAP earnings came in at $0.65 a share against $0.58 expected, an 11% beat. Management then raised the full-year outlook twice over: revenue to $4.45–$4.47 billion from $4.30–$4.34 billion, and non-GAAP EPS to $2.50–$2.54 from $2.36–$2.44.
The stock fell roughly 16% by midday, after trading down as much as 20% before the bell. That puts it on track for the steepest single-session decline in Datadog’s history as a public company — worse than the roughly 17.8% drop it took in March 2020, in records going back to the 2019 IPO. A beat, a raise, and the worst day the company has ever had, all in the same eight hours.
The reconciling detail is one sentence from the call. Datadog’s largest customer — an AI company management would not name, running a nine-figure contract across 17 of Datadog’s products — has told the company it is reducing usage beginning in the third quarter. Executives declined to size the decline or say how long it lasts. The raised guidance already absorbs it. That is the part investors could not price: the guide went up and the largest revenue relationship on the books is shrinking by an unstated amount.
Our take: This is the first clean read on a question the whole AI trade has been deferring — whether revenue from AI companies is annuity revenue or spot revenue. Datadog bills on consumption. When an AI lab scales training and inference, the meter runs and the growth looks structural; when that lab optimizes, renegotiates, or brings observability in-house, the meter slows and there is no contract minimum to catch you. A 35.6% growth rate built partly on one customer’s burn rate is not the same asset as a 35.6% growth rate built on ten thousand seat licenses, and until today the market was paying the same multiple for both. At about $88 billion after the fall, Datadog trades near 19.7 times the midpoint of its own 2026 revenue guide. The repricing isn’t about this quarter. It’s the market discovering that it does not know the customer concentration inside any AI-adjacent software name it owns — and deciding to find out the hard way.
The pattern is three days old and getting worse
Datadog is not alone this week. AppLovin landed inside its own guidance range and lost 21%. HubSpot beat on both lines — $3.26 a share against $3.02 expected, revenue of $911.7 million, up 20% — then cut its full-year revenue outlook to $3.678–$3.686 billion against a $3.71 billion consensus. Piper Sandler and Oppenheimer both pulled their ratings the next morning. HubSpot’s tell was net customer additions: 7,000 against an internal target of 9,000 to 10,000.
Three companies, three disclosures, one identical response. In each case the reported quarter was fine and the forward-looking sentence was the problem — and it is happening while the indexes barely move. The Dow slid 0.6% Thursday, the S&P 500 0.2%, the Nasdaq about 0.1%, with initial jobless claims at 199,000 for the week ended August 1, below the 204,000 consensus. Nothing macro broke. Software just got re-rated in public.
What to watch
- Whether Datadog names the customer. A nine-figure AI account across 17 products is a very short list. If a rival or a hyperscaler claims the work, the story changes from “AI spend is cooling” to “Datadog lost a bake-off.” Those are different stocks.
- Q3 guidance versus the raise. Management lifted the full year while flagging a shrinking top customer. That math only works if the rest of the base is accelerating. The next print tells you whether it is.
- Gross margin. It disappointed this quarter. In a consumption business, margin is the honest signal about pricing power — more honest than the revenue line.
- Concentration language in the next 10-Qs. Every AI-adjacent software vendor has a version of this customer. Watch which ones disclose it before they are forced to.
Datadog did everything a company is supposed to do this quarter, then got punished for honesty about a risk its peers have not quantified. If you want a repeatable way to find the line the headline is hiding, we wrote the six-pass routine for it.