Arden Systems shows what serving inference does to a software gross margin

Software has been valued for twenty years on a simple structural claim, which is that the cost of serving one more customer is close to zero. Everything about how the sector trades follows from that claim: the multiples, the tolerance for losses during growth, the assumption that margin arrives automatically once revenue is large enough. Generative features have put a variable cost back into the product, and Arden Systems is the first company of its size to publish the number rather than bury it in cost of revenue.

Arden sells workflow software to mid-sized professional services firms. About 1.1m seats, roughly $780m of annual recurring revenue, and a product that until early 2025 did what workflow software has always done, which is move documents between people and record what happened. The generative features added in the first quarter of 2025 draft, summarise and classify, they are included in the standard subscription rather than sold separately, and they are used heavily because they are good.

The curve is flattening, which management presented as evidence that the worst is over. It is better read as evidence that the early adopters have finished adopting. Usage of the generative features among seats that have had them for more than a year is still rising, just more slowly than the first six months, and the cost line follows usage rather than seats.

The unit economics, as disclosed

Arden's supplemental disclosure breaks cost of revenue into hosting, support, professional services and inference compute. That last line is the disclosure nobody else in the sector provides and it is the reason this company is worth studying rather than the reason it is worth owning.

Gross profit per seat has gone up. That is the fact the bears have to answer and it is genuinely reassuring: the product is more valuable, customers are paying more, and the incremental compute cost has so far been covered by price. The difficulty is the rate of change in the third row against the rate of change in the first. Compute cost per seat has grown at roughly sixty times the rate of price, and there is no version of that comparison which converges on its own.

We made a deliberate choice to include these features in the base product rather than charge for them, because a feature nobody has to buy is a feature nobody uses and we wanted the usage data. That choice has a cost and we have published it. What we are not going to do is pretend the cost is temporary while we wait for somebody else's model to get cheaper.

Corinne Vasquez, chief financial officer at Arden Systems

Three ways this resolves

The path back to a software margin runs through one of three doors, and the company is currently trying all of them. Model efficiency is the first and the least within Arden's control. Routing cheaper requests to smaller models is the second and is already contributing. Charging separately for heavy usage is the third, and it is the one that carries commercial risk, because customers have now had the features included for eighteen months and will treat a separate charge as a price increase.

Arden trades at 6.4 times forward revenue against a peer group nearer 8.1 times, and the discount opened after the disclosure rather than after the margin decline, which tells you something uncomfortable about what the market rewards. A company that reports the number gets marked down. A company that does not report it gets the benefit of the doubt, until it does not.

The guidance to 76% by late 2027 depends on the usage based tier, and the usage based tier depends on renewals that start in the March quarter. That is the event to watch. If large customers accept an allowance and a meter, the margin recovers and the sector gets a template for how to charge for this. If they refuse, Arden absorbs the cost, and every software business that shipped these features for free is holding the same problem with less disclosure around it.