Snowflake's Frontier Engineering Costs Likely Land Outside Its 74% Margin Guide
Snowflake's new Frontier Engineering offer puts its engineers inside customer teams. Its 74% margin guide covers product revenue only, and professional services ran at a GAAP gross loss in Q2.
TL;DR
- Snowflake launched Frontier Engineering on November 3, putting its AI, data and industry experts inside client teams on outcome-based projects with metrics set at kickoff.
- We see it as a cost of winning consumption. Because the 74% target covers product revenue only, the expense likely lands in a services line that's already losing money.
- Snowflake's fiscal Q3 report will show that line and the 14.5% operating margin guide, the first look at the program's cost.
Snowflake launched Frontier Engineering on November 3. The service puts its AI, data and industry experts inside customer teams to build production AI in the customer's own Snowflake environment.
The Inference read the launch as a test of Snowflake's 74% gross margin guide and reported that the stock fell 2.1% on November 4. Snowflake's own definitions undercut that premise, because the 74% is a non-GAAP product gross margin and product revenue excludes professional services.
Whatever these engineers cost most likely lands in the services line and in the operating margin guide. Frontier Engineering reads best as a customer-acquisition and expansion cost dressed as a services offer. Snowflake is betting that engineers it pays for will move signed AI contracts into consumption faster than they eat into a 14.5% operating margin guide.
The 74% leaves out services
The Q2 fiscal 2027 release is direct about it: "Product revenue excludes our professional services and other revenue." Non-GAAP product gross margin is product gross profit divided by product revenue.
For fiscal 2027, Snowflake guided to a 74.0% non-GAAP product gross margin on $6.07 billion of product revenue, up 36%, and a 14.5% non-GAAP operating margin, up from 13.5%.
Q2 came in at $1.49 billion of product revenue, up 37%, with a 74.7% product gross margin and 126% net revenue retention. Morningstar's copy of the Business Wire release includes the release's definitions of product revenue, net revenue retention and the non-GAAP measures.
The cost of an engineer inside a customer's team would only hit that 74% if Snowflake booked it against product revenue, and the likelier home is the line that already holds services. Snowflake hasn't said which it will use.
By the release's tables, professional services and other revenue was $54.9 million in Q2, under 4% of revenue and up 0.8% year over year, at a GAAP gross margin of minus 38%, or minus 8% on a non-GAAP basis. Snowflake already runs services at a loss, and adding embedded engineers would widen that loss on purpose.
So the test moves to operating margin. Snowflake raised that guide by a point while preparing to staff a new embedded offer, and the fiscal Q3 report will show whether the two fit together. GAAP operating margin was minus 17.0% in Q2, so the non-GAAP figure is the one the company steers by, and the one to watch.
Why Snowflake is paying for this
Snowflake earns product revenue when customers consume, which is the model behind that 126% retention figure. A signed AI contract whose pilot never reaches production burns little. Frontier Engineering goes after that gap.
Snowflake's Frontier Engineering page says every engagement is outcome-based, with success metrics and deliverables set at kickoff. Each one runs through four steps: scope, build, activate and transfer. It ends with the customer running the system on its own.
Madhan Arumugam is Snowflake's SVP of Frontier Engineering. In a blog post, the first in a series on the service, Arumugam writes that most enterprises have promising AI pilots that stall before production.
Arumugam's post lists enterprises with stalled pilots as one audience, alongside regulated industries and teams short on AI experience, and says "field-tested innovations are funneled back into the core platform." The use cases on the page include work a revenue leader would recognize, such as sales prep, coaching and pipeline management, alongside finance work.
The growth picture explains why Snowflake would pay for this. On the September 2 call, management credited roughly half of the quarter's acceleration to its agent products, CoCo and CoWork, according to The Inference. The release counted more than 9,100 CoCo accounts and 5,800 CoWork accounts.
Every one of those accounts that stalls short of production is consumption Snowflake planned for and didn't get. If an engineer gets a regulated customer's agent into daily use and the customer's credits follow, that engineer covers a lot of services loss.
Snowflake's customer examples point the same way, and all of them come from Snowflake. Veradigm built a custom model in seven weeks that it says performed "up to nine times better."
Snowflake cites a pharma customer with $100 million in projected revenue growth, a forecast and Snowflake's own figure. The page also cites a Futurum report, published with Snowflake, that puts lack of skilled personnel at 52%, so the research is vendor-linked and recommends what the vendor sells.
The brand and the outcome-based offer are new as of November 3. Embedded engineers aren't new to software, and The Inference lists comparables at AWS, OpenAI and Anthropic. Snowflake's version comes from a usage-priced vendor, and its margin definitions show where the cost goes.
The handoff limits the payoff
Snowflake's own page cuts against the lock-in case The Inference leaned on. Every engagement ends in a transfer step, where the customer runs the system independently with documented code, so an engineer who hands over the keys builds little lock-in. The services line is also small and barely growing, so nothing in the filings yet shows a services push at scale, or a payoff from one.
Neither point touches consumption. The system Snowflake transfers still runs in the customer's Snowflake environment. After the handover, the customer maintains it and Snowflake keeps billing the usage. The program can lose money on delivery and still earn its keep if the transferred systems keep running.
Snowflake hasn't said how Frontier Engineering is priced, whether customers pay for it at all or how many engineers it employs. The customer results are company-stated, and the pharma figure is a forecast. Until the fiscal Q3 report, the cost shows up only indirectly, through the services line and the operating margin guide. Snowflake may not report either in a way that isolates the program.
Field-tested innovations are funneled back into the core platform.
Snowflake's fiscal Q3 report, with its services line and 14.5% operating margin guide, is the first place Frontier Engineering's cost can show.
