Snowflake’s Losses Are the Wrong Thing to Focus On
Snowflake booked a net loss of $-1.3 billion under GAAP for the quarter ended July 31, 2026, reported on 2026-09-02, and the stock rose +16.6% that same day. Read only the loss line and the reaction makes no sense.
It has made no sense to a certain kind of investor for six years now, ever since Snowflake went public on September 16, 2020, in what was, at the time, the largest software IPO on record, a $3.4 billion raise that included Warren Buffett’s Berkshire Hathaway buying $250 million of stock at the offer price, the only IPO Berkshire has ever bought into directly. That pedigree bought Snowflake patience from the market. Six years of GAAP losses later, the patience is still being extended, and I think the reason is visible in this quarter’s numbers, not in Buffett’s name on the cap table.
My thesis: the GAAP loss says almost nothing about whether this business is working, because Snowflake runs the cost of every customer’s cloud compute through its own income statement, and what actually explains the stock’s move is product revenue growing 37% to $1.49 billion with net revenue retention at 126%, both from the results Snowflake filed with the SEC on 2026-09-02.
A $1.3 billion loss the market shrugged off
Operating loss for the quarter was $-1.4 billion, an operating margin of -31%, and net margin came in at -28%. None of that improved much against a year earlier; net income was also about $-1.3 billion in the year-ago quarter. What did change is the denominator. Revenue grew under a roughly flat dollar loss, so the loss as a share of revenue is shrinking even where the loss in dollars is not. Wall Street’s own model assumes that keeps happening: consensus puts forward EPS at $-2.18 against trailing EPS of $-3.17, arithmetic that implies the per-share loss narrows by roughly 31% from here. I would treat that as a direction, not a promise. Consensus estimates get revised. Snowflake’s financials page has the quarter-by-quarter detail.
Gross margin, the cleaner signal in a consumption business, moved the right way: 67.2% in the quarter against 66.5% a year earlier. That is the line I would watch before the net loss, because gross margin mostly reflects the cloud bill Snowflake cannot easily negotiate down, while the operating loss reflects spending choices management is actively making on sales and research.
Consumption pricing runs the opposite way from a subscription
A seat-license company signs a contract once a year and books the revenue whether or not the product gets used. Snowflake customers pay for the compute and storage they actually consume, so revenue moves with their workloads inside the quarter, not with a renewal date. That is the mechanical reason product revenue growth, 37% to $1.49 billion, is the figure I would lead with over total revenue, which grew 35% to roughly $1.5 billion over the same three months, according to the earnings release Snowflake filed with the SEC on 2026-09-02. The company also raised its full-year product revenue outlook to $6.07 billion, implying growth accelerates modestly into the back half rather than fading.
Net revenue retention was 126% in the quarter. Strip out the jargon and it means customers who were already paying Snowflake a year ago are now paying about 26% more, before counting a single new logo. Remaining performance obligations, the backlog of contracted future revenue, rose 30% to $9 billion, and customers spending over $1 million in trailing product revenue reached 828, up 27%. Total customer count grew a slower 32%, to 14,554. Put those two growth rates side by side and existing customers, not new logos, are doing most of the work.
Cortex turns into a product line, not a slide
The bull case that replaced “unprofitable software company” is Snowflake as the data layer AI models run against, since most large customers’ data already sits inside it. Cortex, the AI layer built into the platform, now carries usage numbers instead of a roadmap slide: Cortex Code, the coding agent, passed 9,100 accounts and added more than 2,000 of them in the quarter alone, while CoWork, the AI assistant, reached 5,800 accounts. Management said AI products drove roughly half of the quarter’s growth acceleration. I read that as early but real, since it is usage data rather than a projection. One quarter of adoption numbers does not prove Cortex becomes as durable a moat as the compute layer itself, the kind of position I described in Nvidia’s own AI buildout.
Why would an AI coding agent even need Snowflake underneath it? Because the data an enterprise wants an AI agent to reason over, customer records, transaction history, inventory, usually already lives in a Snowflake warehouse, and moving it somewhere else for a model to query is exactly the kind of duplicate infrastructure companies have spent a decade trying to avoid. That is the argument, anyway. It is unproven at scale, and 9,100 accounts out of more than 14,000 customers is adoption, not saturation. I would not pay up for Cortex as if it were already the dominant reason to own the stock; I would treat it as a call option that consumption growth and retention are funding for free right now.
Pricing this without a trailing P/E
A negative P/E is not a useful number here, so I use price to sales. At around $332.43 as I write this, Snowflake trades near a $117.3 billion market cap, about 21.5 times trailing revenue and 16.3 times the revenue analysts expect over the next year. Its own five-year average trailing multiple is 29.5 times, so today’s multiple sits roughly 27% below where this stock has typically traded against its own revenue. For comparison, Microsoft trades around 27.5 times earnings, a business already GAAP profitable enough that a P/E ratio means something. That gap is the whole reason price to sales, not price to earnings, is the fair yardstick for Snowflake right now; comparing a negative P/E to Microsoft’s would tell you nothing useful.
Consumption pricing cuts the other way in a downturn, and it is worth sitting with that for a moment. A customer locked into an annual seat license cannot cut spending mid-contract even if it wants to. A Snowflake customer can simply run fewer queries next month, and the revenue follows immediately. Net revenue retention above 120% says that has not happened yet. It is also the first place I would look if the broader software spending environment weakens, well before it shows up in a headline revenue number.
Price to book tells a different story: 54.3 times now against a five-year average of 18.6 times. I cannot fully explain that gap from the figures in front of me. I would want a 10-Q footnote on share count and buybacks before leaning on it either way, and I am flagging it here rather than pretending it is settled.
The stock is also 13.6% below its 52-week high of $385 and 181% above its 52-week low of $118, a low set during a much rockier stretch for the shares. That range alone says this is not a quiet stock. I use a similar framework here to the one I laid out judging SpaceX after its own steep drop: let growth and retention data lead, and treat one accounting line as a lagging indicator rather than the verdict.
| Metric | Value | Context |
|---|---|---|
| Price / market cap | around $332.43 / $117.3 billion | 13.6% below the 52-week high of $385 |
| P/S trailing / forward | 21.5x / 16.3x | five-year average 29.5x |
| P/B | 54.3x | five-year average 18.6x |
| Product revenue, fiscal Q2 2027 | $1.49 billion, +37% | third straight quarter of accelerating growth |
| Net revenue retention | 126% | existing customers spending about 26% more than a year ago |
| Analyst targets avg / high / low | $435 / $525 / $325 | 35 analysts, 91% buy-rated |
Where thirty-five analysts actually landed
The average target across 35 analysts is $435, about 31% above the current price; the high estimate is $525, 58% above; the low is $325, essentially in line with where the stock trades now. 91% of that coverage carries a buy-equivalent rating. Options pricing implied an average earnings-day move of 16.7%, and the actual move after the September 2 report was +16.6%, about as close to the option market’s own estimate as this stock has landed in recent quarters. Short interest is a modest 5.2% of shares, not a name bears are pressing against right now. Our own quant grade for Snowflake is a B, unchanged from before the print, which reads to me as solidly ranked rather than either an endorsement or a warning.
The case against my own thesis
Here is where I would be wrong. If net revenue retention slips toward 115% over the next two or three quarters, that is a sign existing customers are pulling back consumption, and no amount of new-logo growth or AI headline numbers fixes a shrinking installed base. Growth decelerating from the high-30s toward the mid-20s while retention also falls would be the specific combination that breaks the “GAAP losses don’t matter” argument. At that point the market would start pricing Snowflake like a maturing vendor rather than a compounder, and a 21.5-times-revenue multiple gets much harder to defend.
Position size matters too, given a stock whose average earnings-day move is 16.7%. I have written before about how I handle sizing when a stock’s beta runs above 2, and Snowflake, with earnings moves like these, belongs in that conversation.
The number I would watch hardest at the next quarterly report is net revenue retention. Anything at or above 124% keeps this thesis intact. A print below 118% would make me take the richer side of that price-to-sales multiple seriously as a risk rather than a feature.
Analysis and opinion only, not investment advice. Figures come from Snowflake’s fiscal second-quarter 2027 results, filed with the SEC on 2026-09-02, and its investor site; valuation multiples are approximate and were checked on 2026-09-18.