After the close on 2 September, Snowflake (SNOW) reported Q2 FY27 (fiscal quarter ended 2026-07-31): product revenue of $1,491.9M, up 36.8% year on year, a third consecutive quarter of acceleration; non-GAAP EPS of $0.62, 38% above the $0.45 consensus. The shares closed the next session up 16.5% at $356.47 (all price moves in this note are measured on a closing basis; the intraday high that day was $371.63). On 4 September they gave back 5.41% to close at $337.18, a market capitalisation of roughly $116.9B, as sector money rotated out of pure-play software. In two trading sessions the market repriced this company from "a data warehouse with slowing growth" into "AI Data Cloud re-acceleration".
Our core judgement is this. The acceleration itself survives scrutiny: under consumption-based billing, 37% is usage that customers genuinely burned through in the quarter, which is a higher grade of evidence than the same growth rate delivered under a subscription model. But the line the market is paying for, that "roughly half the acceleration comes from AI", today has only management attribution behind it and no independent dollar disclosure. Reverse-solving the current price gives an implied assumption of a 37.8% revenue CAGR over the next five years with zero deceleration (workings below), which puts the pricing intensity ahead of what has been disclosed. The gap between that narrative and the hard numbers is the spine of this note, and the first reconciliation window is the Q3 print in early December.
What the Three-Quarter Acceleration Is Worth: Consumption Basis, Breadth Moving the Same Way
The five-quarter product revenue sequence runs $1,090.5M to $1,158.4M to $1,226.6M to $1,334.3M to $1,491.9M (10-Q key metrics table), with year-on-year growth climbing quarter by quarter out of the 28-30% band to 36.8%. The breadth metrics point the same way. Net revenue retention is 126%, on a five-quarter track of 125/125/125/126/126, turning back up against a backdrop where management itself expects long-term decline. Customers spending more than $1M number 828, up 27%, on a five-quarter sequence of 654/688/733/780/828. Total customers stand at 14,554 on the 10-Q basis, a net addition of 692 in the quarter, and Global 2000 customers number 829, contributing roughly 41% of first-half revenue, so large-account concentration and breadth expansion are running in parallel. Profitability improved alongside: non-GAAP operating margin of 15%, up 400bps, and a GAAP net loss of $191.7M, 35.7% narrower than the $297.9M a year earlier. Guidance was raised across the board: Q3 product revenue of $1.588-1.593B (+37-38%), full-year FY27 of $6.07B (+36%, raised from $5.84B and +31%), and non-GAAP operating margin guidance lifted from 13.5% to 14.5%.
What deserves attention here is the evidential meaning of consumption billing. Compute and storage are charged on actual usage, so revenue has no amortisation cushion from deferred contracts; 37% is what customers actually consumed in the quarter, and it cannot be watered down. On directional verification, that leg is hard. But the same mechanism also runs in reverse, and the bear section below returns to the point.
The 16.5% Next-Day Gain: Seven Points of Beat, and Price Targets Doing the Real Work
Start by getting the size of the beat right. Pre-print consensus, on the basis as of 2026-09-02, was non-GAAP EPS of $0.45 and product revenue of roughly $1.42B, up 30% year on year, sitting at much the same level as the company's own $1,415-1,420M guidance. The actual outturn was $1,491.9M and +36.8%, roughly 7 percentage points above consensus and about 5% above the guidance midpoint. That is a beat of respectable size, but using it alone to explain a 16.5% single-day re-rating does not add up.
Where the sell side was genuinely bearish going in was price targets. The consensus target was roughly $302 across 51 covering firms, about 8% below the pre-print market price, with JPM pinned at $285. The sell side was in effect saying that the results would be good but the stock had already borrowed against them. Two days after the print, JPM raised to $426 and the consensus target rose to $405.92 (MarketBeat, compiled 2026-09-03), against a rating structure of 34 Buy / 5 Hold / 1 Sell. Targets flipped from 8% below the market price to roughly 20% above it. Our judgement is that the bulk of this re-rating came from a wholesale mark-up of growth-duration assumptions plus underweight money covering; that is an author estimate, a single day's move cannot be attributed precisely, and we are not going to pretend we can decompose it to the single percentage point.
The starting point for sentiment also needs recording. In the session ahead of the print, on 2 September, a sector sell-off pushed expectations going in to the floor: SNOW fell 4.5% that day, while MDB, having delivered an above-consensus beat-and-raise the previous evening, still closed down 12.66%. That fall came from Atlas growth stalling at 29% for a fifth consecutive quarter and implying further deceleration, on top of buy-side expectations that were running high into the print. There was no shortfall in the results; what the market punished was narrative deceleration. The whole SaaS complex took fright at that lesson, which is what gave SNOW's shorts and underweight positions their fuel to cover. After more than $100 of target upgrades in two days, the cushion for error in the other direction is thinning fast.
The AI Disclosure: From Nothing to Account Counts, Still One Step Short of Dollars
The most substantive disclosure progress this quarter is in the AI product line. CoCo (the AI coding agent) has an account count for the first time: more than 9,100, with net additions of more than 2,000 in the quarter. CoWork stands at 5,800. The CFO described the acceleration as doubly driven, by "a substantive step up in the core data platform and in AI revenue", and the CEO went further, attributing "roughly half of this round of acceleration" to AI products. Management also disclosed a mechanism: accounts using CoCo consume more on the core platform. The role of AI products here is as an entry point into core consumption, and that offers the most plausible explanation currently available for that one percentage point of upturn in NRR.
We label that attribution explicitly as inference. In the previous set of results, AI disclosure was zero; this quarter there are account counts and an attribution, which is substantive progress. But the "half" proportion cannot be independently verified, the source materials contain no dollar disclosure of AI revenue, and we will not construct an estimate here. Dollar-level disclosure is the switch that would confirm the claim, management has been questioned on it for several quarters running, and Q3 is the natural window.
The cost of all this is already written into guidance. Third-party cloud infrastructure expense, including AI inference and GPUs, accounts for roughly 75% of cost of product revenue, against 70% a year earlier, and FY27 product gross margin guidance has been revised down to 74.0% as a result. The company is simultaneously locking in supply for AI inference: a three-year AI service provider agreement with total commitments of $390M running to 2028-11, alongside a five-year $6.0B cloud infrastructure agreement running to 2031-03. The higher the AI share, the lower the blended gross margin, and management itself has already conceded the point.
What the Market Has Signed Up For: A 37.8% Five-Year CAGR, Zero Deceleration
Reverse-solving from the 4 September close (reverse DCF: WACC of 9.5%, perpetuity growth of 3%, a ten-year competitive advantage period), the current price implies a 37.8% revenue CAGR over the next five years and a steady-state free cash flow margin of 23.18%. The demanding part is that 37.8% sits above the FY27 guidance the company has only just raised (+36%), which amounts to requiring five years without deceleration. Solve the same price along a path of gentle deceleration instead and the steady-state margin has to expand to 35.7%, above anything in the company's history and pointing the opposite way to rising infrastructure intensity. Shift a third axis and you need a fifteen-year competitive advantage period. Whichever of the three axes you solve from, what the current price is pricing is growth duration.
Method footnote (v1 simplification; read the outputs as orders of magnitude). The reverse-solve base is TTM total revenue of $5.03B, through 2026-04-30 and therefore excluding the latest quarter, which is a different basis from product revenue guidance, so the comparison of 37.8% against +36% should be read as an order-of-magnitude reference. The 23.18% steady-state margin is taken on a GAAP TTM free cash flow basis, close in value to the company's 23% adjusted FCF guidance but not the same source, since the adjusted basis adds back withholding-tax items and in Q2 the two differed by 0.6 percentage points. The valuation target treats EV as approximately equal to market capitalisation and does not deduct the several billion dollars of net cash on the balance sheet; deducting it would bring the implied growth rate down slightly, which directionally favours this note's cautious conclusion. The path assumes margins ramp linearly to steady state over five years, with growth decaying linearly to the 3% perpetuity rate in years six through ten.
For a sense of scale on that compound rate: 37.8% compounded over five years corresponds to FY31 revenue of roughly $25B, broadly twice the combined current run-rate of SNOW and Databricks together. On multiples, the current price corresponds to a forward P/S of about 18.5x and a forward PE of 113x, at the top end of the range for high-quality consumption-based SaaS.
Set that against the company's own numbers: FY27 guidance of +36%, management stating explicitly in the 10-Q that it expects NRR to decline over the long term, and adjusted FCF margin guidance of 23% with no expansion. In our view, the difference between the market's pricing and the company's guidance is the option value of two things that have not yet been disclosed: AI revenue expressed in dollars, and the agentic seat. The market has set the exercise date on that option itself, at the Q3 print.
The Strongest Bear Combination: No Cushion in Consumption, Layered on Zero-Tolerance Pricing
The consumption model cuts both ways. It is highly elastic on the way up, and on the way down it has none of the deferred cushion of a subscription model. In the 2023 optimisation cycle, SNOW growth fell all the way from +70% to +28% and NRR dropped from 178% to 125%, which was the previous run-through of the same mechanism. Two less comfortable readings this quarter are worth recording. RPO is $9.00B, up 29.9% year on year, but down 2.2% sequentially from $9,205.2M in April, which the company attributes to seasonality in orders concentrated in Q4. And deferred revenue fell by a net $807.2M in the first half, with recognition running ahead of billings. Cash flow needs an honest entry too: Q2 free cash flow was $83.8M at a 5.4% margin, on a company pattern of high Q1/Q4 and low Q2/Q3, with the TTM figure at 22.0% (GAAP FCF, not the same source as the 23% adjusted guidance, offered only as a directional reference). Taken individually each has its explanation; taken together they need Q3 to provide counter-evidence.
Zero-tolerance pricing amplifies all of that fragility. Should Q3 product revenue land below the $1.588B guidance floor, what gets damaged is the "five years with no deceleration" foundation, and a forward P/S converging from 18.5x towards the sector median of 10-12x corresponds to more than 30% of valuation compression. The control experiment has just run in the same sector, at MDB: an above-consensus beat-and-raise after the close on 1 September, and simply because Atlas growth marked time and implied continued deceleration, on top of elevated buy-side expectations, the shares still closed down 12.66% the following day. There was no shortfall in the results and the market fired anyway; what it punished was narrative deceleration. For a stock priced on five years without deceleration, that precedent is worth reading more than once. With a consensus structure of 34 Buy against 1 Sell, bad news has no buffer zone.
The hidden tax on shareholder returns needs quantifying separately. First-half stock-based compensation was $826.1M, on a cash flow statement basis net of the capitalised portion, or 28.1% of revenue; the $300M buyback broadly only offset withholding taxes on equity settlement, so the share count rose rather than fell. Potential dilution comes to roughly 14.2% of shares outstanding on a nominal basis, adding together options, RSUs and the $2.3B convertible, and the composition needs explaining. The convertible is already deep in the money and the conversion window is open, but the capped call put in place at issuance provides a hedge, and on the company's guidance basis the net impact is only around 10 million shares. A closer reference to reality is the gap between the non-GAAP diluted share count and the basic count, roughly +8.8% (380M against 349M). Read 14.2% as the nominal ceiling before hedges, and 8.8% as the actual pressure today. The market prices on a non-GAAP basis while shareholders bear the dilution per share, and when growth shifts down a gear that scissor gap gets repriced.
The fight over the seat is the medium-term variable. Databricks' self-reported figures (company PR, unaudited) show a $7B run-rate, up 80% year on year, within which the data warehouse product aimed straight at SNOW's heartland is $1.5B and up 100%. Both companies accelerated in the same direction in the same quarter, and we lean towards reading the present as a positive-sum period of AI workload pool expansion rather than a zero-sum share war. The criterion for adjudicating share is set down here in advance: once the two companies' combined growth falls below +40%, if Databricks' warehouse product growth is still running at more than twice SNOW's product revenue growth, then share erosion moves from inference to fact. On Snowflake's claim to the seat, our current formulation is that its presence is confirmed and its exclusivity is not.
The bear list also holds the largest item of all: the three hyperscalers. Microsoft Fabric sells seats bundled into Microsoft agreements enterprises already have, BigQuery and Redshift use bundled pricing to compress gross margin, and AWS is simultaneously the supplier to whom SNOW has just signed a $6.0B five-year commitment, which makes the competitor and the landlord the same set of people. Open table formats, with Iceberg as the leading example, are loosening the lock-in underpinning storage revenue over the long run. Macro is unavoidable too, since a duration asset on a forward PE of 113x is highly sensitive to interest rates. None of this makes it into this note's verification calendar, for the reason that inside the Q3 adjudication window none of it is a first-order variable: seat replacement by Fabric and table-format migration are measured in years, and rates are sector beta. Stretch the holding period and this section moves to the front page of the list.
Verification Calendar
| Timing | Event | How to read it |
|---|---|---|
| Early December | Q3 FY27 results | Product revenue against the $1.588B guidance floor and the 35% year-on-year line; whether AI revenue is disclosed in dollars for the first time (the confirmation switch); whether NRR holds 125; CoCo net additions against the 1,000-per-quarter line; product gross margin against the 72% line |
| Same 10-Q | RPO and deferred revenue | Whether RPO can turn positive sequentially; whether the gap against revenue acceleration narrows |
| Management mid-quarter update | Wording on consumption | Language along the lines of "return of the consumption optimisation cycle", together with net customer additions below 500 in a quarter, moves consumption fragility from a structural risk to a triggered state |
| Ongoing | Databricks self-reported figures | Once the two companies' combined growth falls below +40%, if its warehouse product growth is still more than twice SNOW's product revenue growth, share erosion is established |
To pull this together: three consecutive quarters of acceleration and the upturn in NRR stand up to scrutiny on a consumption basis, so directional verification is complete. "Half the acceleration comes from AI" has attribution but no dollars, while the current price implies five years without deceleration, which puts the pricing intensity ahead of what has been disclosed. That gap will be adjudicated by three readings in the early-December Q3 print: how fully the 37-38% guidance is delivered, whether AI revenue is disclosed in dollars for the first time, and the direction of NRR and CoCo net additions.
The next observation point is that set of results and the accompanying 10-Q. If a management mid-quarter update comes in between, watch the wording on consumption; if Databricks updates its self-reported figures, put both companies' growth rates in the same table and read them together. Whether the intensity gets delivered will answer itself when the date arrives.
This note represents the research views of LunarTulip Research, compiled and analysed from publicly available information, and does not constitute investment advice. Markets carry risk; decisions must be made independently.