Snowflake adds $230m to its full-year product guide on a $70m quarterly beat

The raise is the part that repays arithmetic. Going from $5.84bn to $6.07bn of full-year product revenue ^[6] is a $230m increase ^[1] built on a July quarter that beat product consensus by $70m ^[2], which is a guide moving 3.3 times the size of the result underneath it ^[3]. Roughly $87m of the increase is already visible in the October guide, whose $1.5905bn midpoint clears the $1.504bn consensus by 5.8% ^[5][4], leaving something like $73m attached to a quarter that has not been guided at all ^[5]. The prior guide and the sell-side number sat $10m apart ^[6], so mixing those bases costs about 4% of the calculation ^[6], which is a rounding error carrying an asterisk rather than a reason to ignore the shape.

The shape is this: revenue guided up, unit economics guided down. Adjusted product gross margin came in at 74.7% against 75.8% a year earlier ^[7], a 110 basis point give-back ^[7], and the full-year guide of 74% ^[9] sits 70 basis points below the quarter just reported ^[8]. That is what happens when the incremental workload is inference rather than a dashboard. The $6bn, five-year Graviton agreement signed with AWS in May ^[14], which Ramaswamy says fixes the price paid for storage and compute ^[15], works out to $1.2bn a year if it is spread evenly, and the source does not say it is ^[9]: call it a fifth of guided product revenue and about three-quarters of the product cost of revenue implied by that 74% guide ^[10]. Snowflake has chosen to rent compute at a locked price rather than build toward its own, and $6bn is the size of the choice.

The pull-up case reads well on the customer file. Net revenue retention of 126% and 828 accounts above $1m of trailing product revenue, up 27% ^[12], plus 9,100 accounts on the CoCo coding agent and 5,800 on CoWork since February ^[16], describe existing customers consuming more, not a model vendor going around them ^[19]. Adjusted operating margin of 15.3%, 410 basis points wider than a year ago ^[8], says the extra consumption is not being bought with discounts.

What would prove it wrong is sitting in the same release. Remaining performance obligations of $9.0bn grew 30% ^[13] while product revenue grew 37% ^[3], seven points behind ^[12], and the book now covers about 1.48 years of guided product revenue ^[11]. Consumption running ahead of contracting is fine for a quarter and is a drawdown if it holds for four. Cash is the other tell: a 23% adjusted free cash flow margin guide ^[9] implies roughly $1.4bn on $6.07bn of product revenue, against $83.8m collected in July, which itself came in under the $104m expected ^[13][11], even as non-GAAP operating income of $237m beat the $187m estimate ^[10].

This is probably wrong, but the more interesting version of the thesis is that AI compute is buying data gravity rather than eating it, and the toll for participating is roughly a point of gross margin a year plus a fixed compute commitment that does not shrink when consumption does.