AI Companies Spend Less on Sales? 12 S-1s Say It Is Mostly Classification

Eight classic SaaS companies spent a median 66% of revenue on sales and marketing in the year their S-1 covered. CoreWeave spent 1%. That gap looks like proof that AI-era companies grow without a sales machine. Measured line by line, sales and marketing as percent of revenue mostly tracks how a company sells and what it books under the label, not whether it is efficient.

The short answer, for anyone who needs it in one paragraph: AI infrastructure companies report very low sales and marketing ratios because they sell capacity to a handful of very large buyers. AI-adjacent software companies such as Figma and Netskope still report 50% to 63% on a GAAP basis, close to the classic SaaS range. The real difference shows up only after you strip out stock pay and reclassified costs.

Sales and marketing as percent of revenue in 12 S-1 filings

The claim being tested is simple. Classic SaaS companies went public spending 40% or more of revenue to acquire customers. AI-era companies supposedly grow on product pull, word of mouth and platform distribution, so they should spend far less.

To test it, this post builds two groups from primary filings. Neither group is a statistical sample. Both are chosen because they are the companies a go-to-market leader is most likely to be benchmarked against.

How the cohort was built

The classic group is 8 software IPOs from 2017 to 2021. For each, the ratio uses the latest full fiscal year in the S-1, taken from the comparative figures in the company's first annual report via the SEC's XBRL company facts. Okta and MongoDB later recast those years slightly, and the table uses the recast pair so that revenue and spend share one basis.

The 2025 to 2026 group is 4 companies with a public registration statement and a reported sales and marketing line: Figma, Netskope, Cerebras and CoreWeave. Anthropic is discussed separately, because its prospectus is confidential and no sales and marketing figure has been reported from it. Nscale, which filed in September 2026, presents a combined selling, general and administrative line, so it cannot be placed in the table at all.

All figures are GAAP and include stock-based compensation. Classic SaaS rows come from SEC XBRL comparatives in each company's first 10-K. Figma, Netskope and CoreWeave rows are checked against their filings. The Cerebras row comes from its amended S-1 summary table and had no XBRL cross-check at the time of writing.

The classic SaaS median is 66%, not 40%

The common shorthand says SaaS companies spent "40% or more" on sales and marketing at IPO. In this group, 40% would have been the most frugal company by a margin. Datadog was lowest at 44.9%, and 2 of the 8 spent more than their entire revenue.

My reading is that the 40% figure describes mature public SaaS, not the S-1 year. Companies go public at peak acquisition spend, because the IPO story is growth and growth is bought. Any comparison against "40%" flatters the classic cohort's efficiency and makes the new cohort look less exceptional than it might be.

CoreWeave's 1%: a sales line shaped by one customer

CoreWeave is the number most often cited as proof that AI companies need no sales machine. Its S-1 reported $18.4 million of sales and marketing in 2024 against $1.92 billion of revenue. That is about $104 of revenue for every dollar of go-to-market spend.

The same filing explains why the ratio says little about efficiency. Microsoft accounted for 35% of CoreWeave's revenue in 2023 and 62% in 2024, according to the S-1 risk factors. The 2025 annual report says Microsoft supplied about 67% of 2025 revenue, and that the top 2 customers supplied about 77% in 2024.

A company that sells multi-year capacity contracts to 2 buyers does not need a field sales force. It needs a small team, a few executives, and a balance sheet that can finance the build. The cost of winning those contracts shows up in debt and in data centre commitments, which is why the breakdown of CoreWeave's debt stack is a better guide to its customer acquisition cost than its sales and marketing line.

2025 shows what diversification costs

CoreWeave's own filing gives the counter-evidence. Sales and marketing rose 700% in 2025 to $144 million, driven mainly by personnel and by advertising and sponsorship spending. The 10-K states the company expects the line to keep rising as it expands go-to-market capabilities, builds its brand and diversifies its customer footprint.

That sentence is the honest version of the story. A 1% ratio is what concentration looks like. When the company tries to sell to the next 100 customers instead of the first 2, it starts building the same machine every software company builds. At 3% of revenue it remains tiny, but the direction is set.

Cerebras follows the same arc. Its amended S-1 shows sales and marketing at 7.2% of revenue in 2024 and 13.9% in 2025. Its 10-Q filings show the ratio at 27.2% for the first half of 2026, as the company moved from a few large buyers toward a broader market.

Figma's S-1 sales efficiency once stock pay is removed

Figma is the closest thing in this cohort to an AI-era software company with a classic seat model. On the face of its income statement, it looks like any other SaaS IPO. Its 10-K for 2025 reports sales and marketing at 40% of revenue in 2023, 63% in 2024 and 55% in 2025.

The 2024 jump has a known cause. In May 2024 Figma modified and released restricted stock units, recognising $801.2 million of stock-based compensation that year across all functions. Stock pay inside sales and marketing alone was $206.8 million in 2024 and $218.8 million in 2025, against almost nothing in 2023.

Remove that stock pay and the picture inverts. Sales and marketing excluding stock compensation was 39.8% of revenue in 2023, 35.4% in 2024 and 33.8% in 2025, by Zan Digital's arithmetic from the same filing. Revenue grew from $504.9 million to $1.06 billion across those 2 years.

What Figma's ratio does and does not prove

Figma grew on product-led adoption: designers invite colleagues, and the sales team expands accounts that already use the product. A cash ratio near 34% is lean for a company growing about 41% a year. It is not AI-native efficiency, because Figma's growth model predates its AI features by a decade.

I would use Figma as the benchmark for product-led software, not for AI. If your company sells seats and has a self-serve funnel, the 34% to 40% band is the realistic target. If your sales motion is enterprise-first with no self-serve, it is not. The mechanics of that motion, and how agents change who signs up, are covered in the analysis of product-led growth when the user is an agent.

The Anthropic S-1 has no sales line you can read yet

Anthropic is the company everyone wants in this table, and it cannot go in. Its draft registration statement was submitted confidentially, nothing is on EDGAR, and none of the reporting on the leaked draft includes a sales and marketing figure. What was reported is more interesting than the number would have been.

The draft's 2025 operating expenses were $12.65 billion, of which $7.33 billion was compute and infrastructure, according to Capital Brief's summary of the Reuters report. Everything else, including research staff, sales, marketing, partner fees and administration, fits in the remaining $5.32 billion. That puts an upper bound on sales and marketing at 116% of revenue, which is too loose to be useful.

47% of sales through two cloud marketplaces

Reuters reported that Anthropic routed 47% of its 2025 sales, about $2.16 billion, through Amazon and Google's cloud marketplaces. That share rose from 11% in 2023 and 32% in 2024. In other words, the hyperscalers have become Anthropic's largest sales channel, and that channel is paid by the transaction rather than by headcount.

Channel fees booked as a marketing cost

Here the classification question becomes concrete. Reuters reports that Anthropic books the full value of marketplace contracts as revenue, because it sets prices and delivers the service. It records the platforms' cut, estimated by Reuters at about $351 million, inside a line called "sales, marketing, and partnerships". OpenAI has argued that this gross treatment inflates Anthropic's revenue, and Anthropic's reply is that it is the principal in the transaction.

Both choices push the ratio in opposite directions. Gross revenue makes the denominator larger. Booking channel fees as marketing makes the numerator larger. That fee alone equals about 7.6% of total 2025 revenue, more than CoreWeave's or Cerebras's entire sales and marketing ratio in their S-1 years.

A company that treated the same marketplace sales on a net basis would report lower revenue and a smaller sales line. Its ratio could move either way, depending on how large the rest of its sales spend is. The gross treatment and its effect on headline growth are covered in the earlier piece on Anthropic's enterprise-heavy revenue, and the operating loss is reconciled in the walk from the $42 billion loss to the $8 billion operating loss.

Classification, not efficiency, explains most of the gap

Put the four filings side by side and the pattern is clear. Each company draws the boundary of "sales and marketing" in a different place, and that boundary moves the ratio more than any difference in go-to-market skill.

Two classification issues matter most. The first is where engineering sits. CoreWeave reports no separate research and development line, and its technology and infrastructure expense was $960.7 million in 2024. Solutions engineers who help a customer design a cluster could plausibly sit there rather than in sales, and the filing does not say.

The second is where channel cost sits. A classic SaaS company paying resellers usually books the discount against revenue or as a commission inside sales and marketing. Anthropic books its marketplace fee as a partnership expense. A company selling through a marketplace on a net basis would make the fee vanish from the expense lines entirely. Three treatments of one economic cost, each producing a different ratio.

My position is that AI-era software companies do not yet spend measurably less on sales than their predecessors once you control for these choices. The two software filers here, Figma and Netskope, report 63% and 52%. The two companies that report almost nothing sell hardware and capacity to concentrated buyers. That is a business model effect that existed long before AI. The more useful efficiency question for an AI vendor is revenue per employee, which the revenue-per-employee comparison of AI-native companies covers.

Where this comparison is weakest

The argument above rests on a small set of filings, and it should be read with three specific gaps in mind.

Twelve filings is not a cohort

Eight classic SaaS companies and four recent filers do not make a statistically meaningful comparison. The classic group is biased toward high-growth infrastructure software, which tends to spend heavily. A group built from vertical SaaS or mid-market tools might show a lower median. Treat 66% as the figure for this group, not for SaaS.

The recent group is also thin on AI-native software. Neither Figma nor Netskope earns most of its revenue from AI products. The company that would settle the question, an AI application vendor with a public S-1, does not yet exist in this data. Anthropic would be the first, and its sales line is not public.

The case that AI companies really are leaner

There is a fair counter-argument. Anthropic's marketplace share rising from 11% to 47% in two years means a growing slice of its revenue arrives with no direct sales cost beyond a fee of roughly 16%. That is genuinely cheaper than a classic SaaS field sales motion, where the S-1 ratio implies the company spends about 66 cents on acquisition for every dollar it books.

If Anthropic's public S-1 shows a sales, marketing and partnerships line well below 40% of revenue, the thesis here weakens. I would still expect most of that line to be partner fees and enterprise account teams. I would change my view if the line came in below 20%, because no classification choice could explain that away.

One more limitation. Every Anthropic figure here is a press report of a confidential draft. The classic SaaS ratios are drawn from XBRL comparatives rather than from each S-1 document, and the Cerebras full-year figures come from a summary table that had no machine-readable cross-check when this was written.

What efficient demand generation spend looks like now

The point of the exercise was a usable benchmark, so here it is. These bands are Zan Digital's reading of the 12 filings above, grouped by sales motion rather than by era. They are a starting point for a board conversation, not a statistical norm.

Read the table against your own sales motion, not your product category. A GTM leader at an AI application company selling enterprise contracts through a field team should benchmark against the first row, whatever the pitch deck says about AI efficiency. A product-led company should look at the second row and at its cash ratio, not its GAAP ratio.

The ratio also changes as companies diversify. CoreWeave and Cerebras both show it rising as the customer base broadens. If your board expects a low ratio because an infrastructure peer reported one, the filing history says that ratio does not survive the shift from 2 buyers to 200. Where sales engineering spend should sit as AI shifts how buyers evaluate vendors is argued in the case for sales engineering as the new marketing.

Frequently asked questions

Where to start with your own ratio

Pull your last 4 quarters and compute sales and marketing as a share of revenue twice: once as reported and once excluding stock pay and any channel fees booked elsewhere. If the 2 numbers differ by more than 10 points, the reported figure is not the one to benchmark.

Then place your company in one row of the benchmark table by sales motion, not by how much AI is in the product. Bring that row, and the reference filers in it, to your next budget review instead of a single industry average.

Related reading