The 80% Gross Margin Is Not Coming Back

This spring, in April and May 2026, Kyle Poyar over at Growth Unhinged went out to 230 software companies and asked them a simple question: what gross margin are you actually targeting on your AI products? The median answer came back at about 50%.

That alone should make you pause, but it’s the second number he heard back that matters. Only 12% are still aiming for 80% or more.

Why is 80% relevant? For twenty years, 80% was the line that decided whether you were a software company in an investor’s spreadsheet or something else. It was so consistent nobody bothered to debate it. Now four in five software companies have quietly stopped trying to hit it on the thing they’re building next, and they didn’t make a big announcement about it.

This isn’t margin collapse. It’s a reset that happened while everyone was looking at revenue.

These numbers may not seem shocking on their face but that 80% gold standard in gross margin has, for a long time, been the subtle psychological driver of the tech boom that brought about the FAANG era. The CapEx justifications and investor theory behind why many software companies have been allowed to flaunt conventional operating benchmarks and reap such absurd valuations was, in part, rooted in their potential gross margins. What we’re seeing today is a fundamental shift in just how efficient these companies can be and how fast and far they can grow.

Targets versus reality

Almost every writeup of Poyar’s number makes the same mistake, so let’s separate two things that keep getting collapsed.

Fifty percent is what companies say they want to get to. It is a target, not a result. Companies were asked what they are aiming for. When that gets quoted without the word target, it turns a planned business into a failing one.

What companies actually made looks like this: ICONIQ Growth talked to roughly 305 executives in Q2 of this year for their State of AI report. They put realized AI gross margin at 45% for 2025. Then they asked what comes next. Their respondents said 53% for 2026 and 59% for 2027.

Put those two surveys next to each other and they tell the same story from different angles. Last year: 45% achieved. This year: aiming for around 50%. Next two years: expecting high fifties. That is not a business in freefall. That is a business landing on a new level.

Which means the two hot takes going around are both wrong. One says software margins are collapsing, which the upward slope from 45 to 53 to 59 directly contradicts. The other says inference costs will fall and we’ll be back at 80, which the target data contradicts — if founders believed that, they’d still be targeting 80. They’re not. They are planning to run a different kind of company at roughly 60%.

We love to look at a finished success and pretend it was inevitable. Here we do the opposite: we look at an old success metric and pretend it is a law of nature.

Why 80% existed in the first place

There is a simple reason software got to 80% gross margin: the next customer cost nothing to serve. The ten thousandth user cost the same as the nine thousand nine hundred and ninety-ninth, which was zero.

Because that was true, we built an entire playbook on top of it. Unlimited usage. Unlimited seats. A generous free tier that never worried anyone in finance. Land and expand. Grow first, figure out profitability later. Usage as a sign that revenue was coming, never as a sign that cost was coming.

Inference breaks that assumption in the most direct way possible. It has a marginal cost. Not a rounding error you can hide in hosting. A real cost, priced per token, paid to someone else, every time someone uses your product.

Now run the word unlimited through that new equation. Unlimited was never generous. It was accurate. It was the correct price for a product where incremental usage cost you nothing, and it was great business because more usage built the habit that renewed the contract. Attach a per-token bill to the same promise and your most engaged users become your least profitable. Every funnel software ever built assumed more usage was better. Under this cost structure, more usage can be worse.

The free tier changes the same way. In a zero-marginal-cost world, a free tier is a customer acquisition cost you can forecast. You know what a thousand free users costs you. In a per-token world, a free tier’s cost scales with how good the product is. The better it works, the more it costs you.

That one shift — from zero to something — invalidates habits that went unquestioned for two decades, not because we were careless, but because they were true for so long they stopped looking like assumptions at all.

80% was never a rule about software. It was the result of a cost structure that no longer describes what these companies actually sell.

What it looks like when public companies cross it

Two public stories from this summer show what that crossing looks like from the outside.

Figma reported Q2 on August 12. GAAP gross margin was 84%, down from 89% a year earlier. Five points gone, in what the company itself described as its first full quarter of AI credit monetization, and more than 80% of its customers above $10,000 in ARR were consuming AI credits every week. Revenue was up 48% to $370.1 million. Net dollar retention was 136%. This is not a struggling business. This is a healthy business absorbing a new cost.

I need to be careful here about cause and effect, for the same reason I was careful in the last piece. Figma’s release does not say AI caused the five-point decline. The market and analysts made that connection, and the timing — the drop lining up exactly with the first full quarter of high AI credit adoption — makes the circumstantial case strong. But Figma did not say it. If I wrote that they did, I’d be doing exactly what I spent the last article complaining about.

Canva is the clearer lesson, precisely because it surprised everyone. In August, Canva cut its 2026 revenue growth forecast to roughly 20%, down from around 30%. That’s a third of expected growth taken off the table, and the company said it was because of the cost of delivering AI features.

The details under the headline are weirder than the headline itself.

CEO Melanie Perkins said demand for the new AI features had “significantly exceeded” expectations. At the same time, Canva said it had cut cost per task by nearly 90% since launching Canva AI 2.0 in April, in a period where users created three times as many designs.

So unit cost down ~90%, volume up 3x. On those two numbers, total serving cost actually fell quite a lot over the exact window where Canva cut its growth forecast by a third.

That breaks the simple story. This wasn’t inference bills overwhelming a company.

Where the cost really hits

Perkins put it plainly: “Rather than broadly rolling out a product before the underlying economics were ready, we decided to slow the rollout while we rebuilt the architecture, reduced unit costs and strengthened the business model.”

Read that sentence again. Canva didn’t lose margin to AI. Canva traded growth for margin on purpose, and told investors that’s what it was doing.

The hard part isn’t that inference is expensive. Canva’s own 90% reduction shows how quickly that number is moving. The hard part is the window between shipping a feature and having pricing that accounts for what it costs to run. Inside that window, adoption works against you.

A feature no one uses costs nothing. A feature everyone uses, priced before anyone understood the cost curve, becomes a liability that gets bigger the better the product works. Figma’s disclosure — more than 80% of $10K+ ARR customers using credits weekly — is a success metric and a cost forecast in the same sentence.

How long you stay in that window is a choice. Canva closed it by throttling distribution. Figma closed it by launching credits and eating five points while they did. Both are defensible. The failure mode is not realizing you’re in the window at all.

Think of Blockbuster. This is that moment. Not a bad decision. A set of good decisions made inside an old cost model that no longer applies. We just don’t know we’ve written a different contract until we flip the page.

Pricing is not a philosophy. It is arithmetic.

We keep telling the story about moving from per-seat to consumption as if it’s about aligning with value. Sometimes it is. Mostly it’s math.

If what you sell costs more the more it’s used, and your price doesn’t move when it’s used, you’ve written a contract that lets your customer consume an unlimited amount of a variable-cost input. That’s a bet on customer restraint. Per-seat was a perfect structure when usage cost you nothing. It is a poor structure now.

The surveys say the same thing. ICONIQ’s respondents report 42% on consumption-based pricing and 23% on outcome-based, running on average 1.7 models at once instead of cleanly switching. Poyar’s survey has hybrid models jumping from 25% to 37% of companies in the last twelve months, with 29% now offering AI credits and another third planning to add them within a year.

Credits are the giveaway. A credit system is just a meter with a friendlier name, bolted onto a subscription that was never designed to have a meter. Nobody chooses it because it’s elegant. They choose it because it’s the fastest way to make revenue move when cost moves.

What to actually do Monday morning

This is where time slows down. The future is undefined, and the PRP question is not “is AI a good idea?” but “is now the right moment for this pricing, and are we in a position to capitalize on it if so?” Five moves:

1. Break AI COGS out as its own line before you are forced to. Most software companies still report cost of revenue as one number where hosting, support and inference are indistinguishable. That worked when two were fixed and the third didn’t exist. It doesn’t work now. If your board can’t see inference cost per customer, per feature, and per plan, it can’t tell a pricing problem from an adoption success, and those need opposite responses. Practitioners are starting to call out AI COGS as an earnings risk separate from AI revenue. Your board should see it that way.

2. Price before you roll out, not after you panic. That is the Canva lesson turned into a rule. The pattern in 2025 and 2026 has been ship, measure, panic, reprice. The pattern that works is model the cost curve, set the commercial terms, then scale distribution. It feels slower. It is much cheaper.

3. Stop underwriting new products at 80%. If your launch gate still assumes SaaS-era margins, you’ll kill products that are perfectly good at 55% or approve products on a forecast no one believes. Only 12% of the industry is still aiming at 80. Your model should match the business you are actually running.

4. Make margin a product decision. The biggest lever on AI gross margin is not procurement. It’s product: which model handles which prompt, how much context you send, what you cache, what you route to a smaller model, what the free tier is allowed to do. Those are engineering and design choices that hit gross margin directly, which means gross margin is now something product managers change every week. In most companies, no one has told them that.

5. Expect your comps to break apart. If you sell a traditional seat-based product with an AI feature bolted on, your blended margin will sag as adoption rises, and it will look like decline. If you sell an AI-native product, your margin should improve as you scale and optimize. Those two paths point opposite ways and they will be averaged together in every benchmark you read this year. They are different businesses.

What this does not mean

Software is not turning into a services business. A 55 to 60% gross margin is still a fantastic business, better than most industries ever see, and it still comes with recurring revenue and operating leverage that made software attractive in the first place. ICONIQ’s direction of travel is up, not down — 45 to 53 to 59.

It is also not a temporary dip back to 80. That is the comfortable story, which is why I would push hardest against it. Inference costs will keep falling, but the companies building these products are telling you they don’t plan to bank those savings as margin. They plan to spend them on more capability, which is exactly what companies in competitive markets do when an input gets cheaper.

The number is settling in the high fifties or low sixties. You will land there whether you plan for it or not. The useful thing is to know where you are before your board sees someone else’s quarter and asks you why yours looks different.

Early in my career I was taught we are the sum of the decisions we make. We are also the sum of the contracts we write without realizing we wrote them. An unlimited plan in a zero-marginal-cost world was a smart contract. The same plan in a per-token world is a different contract entirely. The question was never whether 80% was good. It was. The question is whether now is the moment to keep pricing as if it still exists.

---

Sources: Kyle Poyar, Growth Unhinged, “The 2026 State of B2B SaaS and AI Monetization Report,” 230 companies surveyed April to May 2026 (median target AI margin ~50%; 12% targeting 80%+; hybrid pricing 25% to 37% year over year; AI credits offered by 29% with another third planning within a year). ICONIQ Growth, State of AI 2026, ~305 executives surveyed Q2 2026, published July 2026 (AI gross margin 45% actual 2025, 53% projected 2026, 59% projected 2027; AI revenue mix 32% to 42% to 53%; consumption pricing 42%, outcome pricing 23%, average 1.7 models). Figma Q2 2026 results, published 12 August 2026 (GAAP gross margin 84% against 89% a year earlier; non-GAAP 85% against 90%; revenue $370.1M, up 48%; net dollar retention 136%; first full quarter of AI credit monetization; >80% of customers above $10K ARR consuming AI credits weekly). Canva growth forecast revision from ~30% to ~20% and Melanie Perkins quotations via Fortune, 12 August 2026 (demand “significantly exceeded” expectations; cost per task reduced nearly 90% since April Canva AI 2.0; users created 3x designs; quote: “Rather than broadly rolling out a product before the underlying economics were ready, we decided to slow the rollout while we rebuilt the architecture, reduced unit costs and strengthened the business model”).

Featured image generated with AI.