The moment you embed a frontier model in your product, software's beautiful near-100% margin starts bleeding out the back to your vendor, and you wake up a consumption business whether you priced like one or not. North outlines the escape hatch - open-weight models that finally got good enough - is real, and the cost gap is not subtle. But the sharp turn here is that "free weights" is a move, not a gift: somebody is selling compute, neutralizing a rival, or buying a geopolitical position, and this month's overnight government shutoff of Anthropic's top models showed exactly why that last one matters. Read it for the repricing trap most founders don't see until the math breaks. — Muximus
Software used to be the best business in the world: write it once, sell it again and again, keep almost all of the money. Put a third-party AI model inside that software at runtime and you've quietly traded a near-100% gross margin for something closer to 40%. I'm a seed-stage investor, and this is the single force I see reshaping startup economics right now - so let me walk through what's actually happening, why the open-weight models are the obvious escape hatch, and why "free" is a strategy you should understand before you build on it.
The margin you didn't know you were spending
For most of software's history, the cost of serving one more customer was rounding-error small. You paid for some cloud, some support, some customization, and the overwhelming majority of revenue was pure profit. That's the business model venture capital learned to love.
AI breaks the rounding error. I'm not talking about engineers writing code with AI - that's a different revolution. I mean the stage after the product ships, when there's an AI component from Anthropic or OpenAI living inside your product and answering your customers' requests. Every time a customer uses that feature, you pay the vendor. And the vendor charges you "by tokens," which is a more sophisticated way of saying "by consumption" - consumption you can't fully predict in advance.
So now a meaningful share of your revenue flows straight back out to your model vendor. I'm seeing startups drop to 40% gross margins and below for exactly this reason. The expensive part of the product isn't your code anymore. It's the inference bill.
The trap underneath the trap: you have to reprice yourself
The margin hit is the obvious problem. The deeper one is that it forces you to redefine how you charge.
If you sell $50 per seat per month but one user grinds a feature that costs you real per-token money on the back end, the math eventually breaks - you're subsidizing your heaviest users and sometimes losing money on them. The vendor bills you by consumption, so you get pushed to bill your own customers by consumption too. That's miserable, because seat-based or value-based pricing fits how customers actually think about your product. Token-metered pricing fits how your vendor thinks about you. You end up in the corner: optimizing what you send upstream, squeezing tokens, and rewriting your own price list - not because it's better for your customer, but because your cost structure now leaks out the back.
Why "they're not good enough" stopped being true
A few months ago the answer to all of this was: tough luck, the open models aren't good enough, you have to bleed margin and ship the frontier model because your customers need the quality. That has genuinely changed.
GLM, from the Chinese lab Z.ai, is the loud example - its open weights now top the open-model coding leaderboards, and the lab is openly talking about reaching frontier ("Fable"-class) quality by year end. Take the leaderboards with salt: those numbers are vendor self-reported and not independently verified, and the hosted GLM API routes your requests through China, which is its own problem if your data is sensitive. But the direction is not in doubt. DeepSeek, Alibaba's Qwen, MiniMax M3, even Meta's Llama - each is, on certain task types, genuinely not far from the frontier models. You don't have to trust any single benchmark - and frankly the benchmarks are getting harder to trust as the gaps shrink - to see that the gap closed enough to change the build decision.
And the cost difference is not 10% - it's somewhere between 6x and effectively unbounded, depending entirely on how you host. Inference still runs almost only on GPUs, but with an open, freely-licensed model you choose whose hardware to run it on. You pay mostly electricity plus a slice of the GPU. In the extreme, you self-host the dumbest tasks inside your own boundary and pay almost nothing per call - and self-hosting the weights is also the honest answer to that China-data-path problem, since the model runs entirely on your side.
Hybrid routing: the actual operating answer
In practice nobody flips one switch from Anthropic to DeepSeek. What I see startups doing - and what I think becomes standard - is deciding per task: this goes to the frontier model, this goes to Qwen, this goes to DeepSeek, and this one doesn't need an LLM at all, because a few lines of boring old-fashioned code solve it fine. Match each task to the cheapest model that's actually good enough for it.
This has a name now - model routing, or hybrid routing - and it's real: teams report cutting their model bills by half or more this way with no visible quality loss. The catch is also real, and it's the setup for the next part: when you route a request to a model that turns out not to be good enough, the answer gets subtly worse - a missed nuance, a hallucinated detail, a tool call that quietly fails - and nothing on your dashboard flags it. The bill goes down. So does the quality. You find out from support tickets three days later.
The dark pattern that's coming
Now point that capability at the customer instead of the cost line, and you get a category of manipulation software couldn't do before.
For some users I light up the frontier model and they get the premium experience. For others I quietly route to something a fraction of the cost. Or - the one I'd actually watch for - during your evaluation I run everything on the best model, you sign, and a month into the contract most of your traffic is silently being served by some cheap model on the back end. Same product, same price, same screen. One day the experience is X, the next it's a third of X, and you have no way to see the swap.
Software used to be the same software for everyone. Routing makes the product you bought and the product you're being served two different things, decided per request, invisibly. Nobody has to write a line of deceptive UI to pull this off. It's a config change.
"Free" is a strategy, not charity
Which raises the question hiding under this whole shift: why would anyone spend tens to hundreds of millions of dollars building a model that competes with Anthropic, and then give the weights away? They're not being generous. They have goals, and there are at least three of them - four if you count geopolitics.
Sell the compute. This is the Amazon play. Alibaba hands you a cool, free, open Qwen, you try it, it's good and cheap, and you decide you'd rather they also be the inference provider your product calls. DeepSeek is angling at the same thing - get you hooked on the open model, then sell you the infrastructure to run it.
Neutralize a competitor. This is my read on Meta. They're not an infrastructure provider, so the giveaway isn't about selling you compute - it's about making sure Google, Anthropic, and OpenAI don't end up the only three model suppliers on earth, and making sure Meta has a competitive model of its own in the field.
Win a geopolitical position. This is the fourth goal, and it's the one most founders underweight. There's a fight over who gets to have capable models at all, who controls them, and where the data flows - China versus the US, and underneath it, who has access to Nvidia chips.
There's a line in Dune: whoever can destroy a thing controls it. We got a micro-demonstration this month. On June 12 the US government issued an export-control directive that forced Anthropic to abruptly disable its most capable models, Fable and Mythos, for every customer, to comply with a national-security order restricting foreign access. A government reached in and switched off the frontier overnight. The order is already de-escalating - but the lesson landed. If you're China, that's all the argument you need to want strong models under your own control, and to give them away to build the ecosystem around them. Europe wants that independence too, in theory, but there's a wave of important summer vacations to plan first, so my bet is regulation and mediocrity there rather than models of its own.
What I'd actually do Monday
None of this is a reason to panic or to rip out your frontier model tomorrow. It's a reason to do three things now. Know your real per-customer inference cost and whether your pricing survives a heavy user - most founders I talk to are guessing. Build the routing layer before the bill forces you to, and instrument it so you can see quality drop, not just cost. And when you reach for a free open model, ask what its maker is actually selling, because it isn't the model.
Maybe model prices eventually fall to the cost of electricity. Maybe not - it needs more real competitors than we have, and it's not close. In the meantime you're bleeding margin, repricing your product, and seriously testing open weights whether you planned to or not. To borrow William Gibson's line: the future is already here, it's just not evenly distributed. The working version for this corner of it - AGI is already here, it's just not very evenly distributed.
Disclosure: VarOps is published by Ran Aroussi, who is also a seed-stage investor in the startup market it describes.
