A new benchmark started circulating through the industry this week, and if it becomes your new definition of normal, you could end up scaling a cost problem that gets harder to fix as you grow. In this week's feature, find out when the benchmark is useful, when it becomes dangerous, and what your own margin may be hiding.

Pulley told customers it will cease all operations and services Dec. 8 (TechCrunch, Pulley). But anyone considering its preferred migration path to Carta only has until Nov. 30 to opt in for a first-year price match, credit for any prepaid balance and help transferring a standard cap table. Pulley gave no reason for the shutdown.

Existing 409A valuations carry over with the documentation, and Carta bills annually or quarterly, so monthly Pulley customers will move to one of those terms. The FAQ does not say what happens to pricing after the first year.

Customers choosing another provider face a different migration. Pulley's FAQ says the company "is unable to assist a migration to a different provider in any capacity."

Pulley says customer data will remain accessible through Jan. 31, 2027, in a limited format. BPM, an accounting firm, recommends exporting share classes and holders, option grants, vesting schedules, SAFEs and convertible notes, warrants and prior 409A records, then reconciling them against the underlying source documents before choosing a destination.

Founder read: A clean migration can preserve bad data as easily as good data, so reconcile grants, SAFEs, notes and 409As before the new platform makes old discrepancies look authoritative.

Your customers are searching in places your strategy doesn’t reach.

So before your business is buried and left behind, you need to understand the new rules of SEO.

BELAY's SEO in the Age of AI report explains how search is changing, what AI means for your visibility, and the practical steps small businesses like yours can take to stay visible.

BELAY’s U.S.-based Marketing Assistants turn strategy into execution, helping your business stay visible, credible, and competitive in every search.

Download a copy for free

Traditional SaaS has trained founders to expect very high gross margins. SaaS Capital says core SaaS license revenue generally runs at 80% to 85% gross margin across its portfolio and private-company survey data.

Then ICONIQ Growth published its 2026 State of Scaling report.

For its Pacesetters below $10 million in ARR, median gross margin is 55%.

That means 45 cents of every revenue dollar disappears into the cost of delivering the product before sales, marketing, R&D or G&A enter the picture.

Normally, that would look terrible for software.

ICONIQ's companies complicate the diagnosis. Their median year-over-year growth is 900%.

ICONIQ calls the companies in this table "Pacesetters." They have top-quartile revenue growth over the previous three years and are either AI-native or AI-driven.

The broader study draws on quarterly operating and financial data from 137 software companies, including ICONIQ portfolio companies where figures were available and 11 public companies. ICONIQ does not disclose how many companies fall into the under-$10 million Pacesetter group, and figures in the table are rounded where applicable.

Within that group, median net dollar retention is 105% and the median burn multiple is 1.3x. The top quartile reaches 80% gross margin.

So the 55% figure has a very specific provenance. It comes from companies selected partly because they are growing unusually fast.

ICONIQ does not show that weaker gross margins caused that growth. It does not show that these companies deliberately traded margin for growth. And it does not establish 55% as a normal gross margin for young AI companies.

It shows that companies with extraordinary growth can operate there.

That is a useful distinction if your own company is at 55%.

SaaS Capital's 2026 survey offers another view of delivery costs.

More than 1,000 private B2B SaaS companies participated. Among companies with $3 million to $5 million in ARR, median spending was 5% of ARR for hosting, 3% for DevOps, 5% for professional-services cost of goods and 3.5% for other cost of goods.

Those four category medians add to 16.5%.

That calculation has an important limitation. Medians from separate categories cannot be combined into the median cost structure of an actual company. SaaS Capital also does not publish total gross margin in that post.

The categories are more useful than the sum anyway.

They show what a founder needs to inspect when gross margin starts falling: what, exactly, is being charged to cost of revenue?

For an AI company, that might include inference, hosting, implementation, professional services, customer-specific work or infrastructure required to produce the output customers bought.

A single gross-margin percentage hides all of that.

And the reason matters because different costs behave differently.

Implementation expense may fall as onboarding improves. Professional services may remain stubbornly human. Infrastructure may become cheaper with scale. Inference expense may rise directly with usage. Some technical costs can move because engineers change how the product works.

That last category is easier to underestimate than it looks.

On Sept. 17, Tomasz Tunguz wrote about a UC Berkeley study comparing the software systems wrapped around coding agents.

In one comparison, he reports that GPT-5.6 Sol cost 71% less to run inside one runtime than another. Across the paper's 42 comparisons, none showed a statistically significant difference in quality.

There are limits to what you can infer from that. The study covers coding agents on two benchmarks, and this account relies on Tunguz's description rather than the paper itself. A 71% reduction cannot be generalized across AI products.

But the mechanism is hard to ignore.

For at least one class of AI product, a substantial portion of delivery cost changed because of the software surrounding the model. The customer did not receive a worse result. The model itself did not become cheaper. An engineering choice changed what it cost to produce the output.

So "AI is expensive" is not much of a diagnosis.

You need to know which part is expensive.

If your gross margin is 55%, don't start by asking whether ICONIQ makes that acceptable. Pull the last quarter's cost of revenue and find out what is consuming the 45%.

Separate inference and model spend, hosting and infrastructure, implementation, professional services, customer-specific work, and any other material delivery expense. Then identify the two largest categories.

For each one, answer three questions: Does the cost rise as customers use more of the product? Can you materially reduce it without hurting the customer outcome? Does your pricing recover more revenue when that cost rises?

The answers tell you where to work.

High inference costs may send the product team back to model, runtime or architecture choices. Heavy implementation may expose onboarding work that still requires people because the product has not absorbed it. Professional services that grow in step with revenue may mean the business is carrying more labor than the pricing model recognizes.

Then check the growth those costs accompany.

ICONIQ's under-$10 million Pacesetters grow at a median 900% a year. The report does not establish that their spending caused that growth, but it does tell you what kind of companies produced the 55% benchmark.

If your company has a similar margin and far less growth, stop using 55% as reassurance. Pick the largest controllable cost in that 45% and make someone accountable for changing it.

Engagement is easy to mistake for pipeline. On Sep 30, watch how a founder turns LinkedIn content into real outreach. Live. You'll walk away with a repeatable system: what to post, who to reach out to, and how to sequence it. 

Eligible startups also get the LinkedIn-to-Leads Toolkit: ad credits, Apollo, Captions, and HubSpot's Prospecting Agent.

Claim my spot

Thanks for reading. See you next Wednesday.
— Jason

P.S. If the margin looks fine but the cash still feels tight, June's The Profit That Can't Make Payroll follows a profitable month three weeks from missing payroll.