AI Startups Post a Better Burn Multiple Than SaaS. Gross Margin Tells a Different Story.

Every fundraising deck comparing AI companies to traditional SaaS companies is citing the same burn multiple this year. AI startups posted a median of 0.79x in 2026 against 0.89x for the rest of software, a real, sourced number, and a lower burn multiple is supposed to mean better capital discipline. Read the same year’s data on gross margin and the story reverses: AI products are running 50 to 60 percent gross margin, well under the 60 to 80 percent that defined SaaS for two decades. Both numbers are true. The trouble is that most people quoting the first one have not checked the second.

The metric everyone is citing

Start with what a burn multiple actually is. It is net burn divided by net new ARR: how many dollars a company spends to add one dollar of annual recurring revenue. A 1.0x burn multiple means a dollar of cash bought a dollar of new ARR. A 2.0x means it took two. Under 1.5x is considered tight in 2026; above 2.0x heading into a raise, most later-stage investors will pass before finishing the deck, per a 2026 benchmark report from the finance platform Runway, built on survey data from the analytics firm Benchmarkit.

The AI-versus-SaaS comparison, 0.79x against 0.89x, comes from that same body of benchmark work. On its face, it says AI companies are buying ARR growth more cheaply than the software companies that came before them.

What burn multiple does not measure

Net burn is a single number: total spend minus total revenue over a period. The formula does not separate cost of goods sold from operating expense, and it says nothing about how much of each new revenue dollar turns into gross profit. Two companies can post an identical burn multiple while one keeps 80 cents of every dollar and the other keeps 50.

Where AI’s margin actually goes

Inference is the reason. When an AI feature answers a support ticket or drafts a document, the model call is a per-use cost that scales with usage the way a support agent’s salary never did. Most finance teams now book that cost in COGS, the same line as hosting and payment processing, because it is triggered directly by the customer and runs in production. ICONIQ’s January 2026 benchmark survey put average AI product gross margin at 52 percent, up from 41 percent in 2024 and 45 percent in 2025, and still well short of the 70-plus percent SaaS buyers came to expect. A 2026 analysis by the cost-monitoring firm CloudZero estimated that roughly $230,000 of every $1 million in AI product revenue leaves as inference cost before a single salesperson or engineer gets paid.

Run the arithmetic on a single feature and the gap is easy to see. A SaaS subscription that keeps 80 cents of gross profit per dollar, with a bolt-on AI feature costing 15 cents per dollar in direct compute, ends up keeping 65 cents. Same customer, same price, worse economics, and that shift happens before the company changes anything about how it spends on sales or engineering.

How a shrinking margin and an improving burn multiple can both be true

This is the part the aggregate number hides. Burn multiple only tracks the rate of ARR growth relative to burn; it has no view on what that ARR is worth once delivered. An AI company adding new logos and expansion revenue fast enough can post a better burn multiple than a SaaS company growing more slowly, even while giving up nearly twice as much of every dollar to cost of delivery. The ratio rewards velocity. It has no opinion on quality.

That is not a flaw unique to AI companies; it is a property of the metric. It just happens that 2026 is the first year the gap between fast, thin-margin growth and slower, thick-margin growth has been this wide across an entire category.

There is a second wrinkle underneath the first: how the compute itself gets financed. A company that leases GPUs or commits to reserved cloud capacity can capitalize part of that spend and depreciate it over several years, keeping most of it off the current period’s operating expense. A company paying per-token to a model provider expenses the full cost the moment a customer triggers the feature. Two companies running an identical workload can post different burn multiples in the same quarter based on that choice alone, before either one changes anything about the product or the customer.

The benchmark inside the benchmark

Stage and ARR size already explain a large share of the spread that gets attributed to “AI versus SaaS.” Benchmarkit’s 2026 data put the median burn multiple at 1.3x for companies between $5 million and $20 million in ARR, falling to 0.5x for companies between $50 million and $100 million, a threefold difference driven mostly by where a company sits on its growth curve, not by what it sells. Series A companies specifically posted a median of 1.6x, with the top quartile already down at 1.0-1.2x, and that gap between median and top quartile is widening year over year.

AI companies, as a group, skew younger and sit further back on the ARR curve than the SaaS companies they are being benchmarked against. Some of the 0.79x-versus-0.89x gap is genuinely about capital efficiency. Some of it is two different distributions of company age being read as if they were one comparison.

Three questions before trusting a burn multiple comparison

• What counts as net new ARR? Expansion revenue, usage-based revenue, and one-time services fees get bundled into “ARR” with wildly different consistency between companies.
• What is the gross margin trend on the underlying product, not just the topline burn multiple?
• What stage and ARR tier is each company in? An $8 million ARR company and a $70 million ARR company should not be read off the same benchmark line.
• Is the underlying compute owned or leased and capitalized, or paid per-use and expensed immediately? That accounting choice changes how burn shows up in a given period, even when the product has not changed at all.

What the number is actually good for

Burn multiple is still a useful check. A company posting 3.0x while claiming capital efficiency is telling on itself regardless of what it sells. What it cannot do is stand in for a margin analysis, and 2026 is the year that distinction started to matter, because for the first time the companies with the best-looking burn multiples are, on average, keeping the least of each revenue dollar they bring in.