AI/Biz

Stable Diffusion Changed the Math on Build Versus Buy

In April, OpenAI opened up DALL-E 2. In July, Midjourney went to open beta. Both are impressive, both are closed, and both are services you rent.

In August, Stability AI released Stable Diffusion with the weights published. It runs on a consumer graphics card. You can download it this afternoon.

The pictures are comparable. The market structure is not, and that second difference is the one worth writing about.

What Actually Shipped

The technical foundation is latent diffusion, from a CompVis paper the previous December. The idea that makes it practical is straightforward once you see it.

Earlier diffusion models did their work directly on pixels, which is enormously expensive because a high-resolution image is a very large array. Latent diffusion first compresses the image into a much smaller representation, does the whole iterative denoising process in that compressed space, and only expands back to pixels at the end.

The computation drops by a large factor. That’s the difference between needing a data centre and needing the graphics card someone already owns for games.

“The research contribution was doing the expensive part in a smaller space. The commercial consequence was that the expensive part stopped requiring somebody else’s data centre.” — Sameer Gupta

Why Open Weights Change the Calculation

I want to be precise about what changes for a business, because “it’s free” is the least interesting part.

Your data stays inside. This is the big one and it’s underdiscussed. Using a hosted service means every prompt, and often every input image, goes to a third party. For anyone working with unreleased products, client material, or anything under a confidentiality obligation, that was frequently a blocker regardless of price. Running locally removes the question entirely.

Cost becomes fixed, not marginal. A hosted API charges per image forever. Your own deployment costs hardware and electricity, and the marginal image is approximately free. At low volume the API wins easily. There’s a crossover, and if you’re generating at scale it arrives faster than people expect.

No vendor risk. No rate limits, no pricing changes, no terms of service update that prohibits your use case, no deprecation. I’ve written repeatedly about the dependency risk in building on somebody else’s API. Open weights are the only real answer to it.

You can modify it. Fine-tune on your product photography, your brand style, your specific domain. You cannot do that with a closed service beyond what its interface allows.

It doesn’t go away. The copy on your disk works in five years whether or not the company that made it still exists.

The Argument Against

I’d be doing you a disservice if I only made one side.

  • Somebody has to run it. Free weights, paid operations. GPUs, deployment, scaling, monitoring, and a person who understands all of it. For a team generating a few hundred images a month, the API is obviously correct and it isn’t close.
  • The hosted versions are still ahead on convenience, and convenience is a real product feature. Prompt quality, consistency, and interface polish are worth paying for.
  • Safety is now your problem. A hosted service filters. An open model does what it’s asked. If you deploy this in a product, you own the moderation problem entirely, and it’s harder than it looks.
  • The legal position is genuinely unsettled. Training data provenance, output copyright, and the licence terms attached to these models are all live questions with no confident answers. Anyone telling you otherwise is guessing.

The Pattern I’d Point At

Step back from images, because I don’t think this is really about images.

We’ve now watched the same sequence twice. A capability appears, initially available only as a service from a well-funded lab. Within a year or two, an open version arrives that’s within touching distance and runs on ordinary hardware.

It happened with vision models after ImageNet. It’s happening with image generation now. I’d expect it with language models next, and probably sooner than the current gap suggests.

“The frontier is rented. The second-hand frontier is owned, and it arrives about eighteen months later at a fraction of the price. For most business problems, eighteen months is not a meaningful delay.” — Sameer Gupta

If that pattern holds, it has a strategic implication for anyone building on top of these capabilities. The right question isn’t “which vendor should we commit to.” It’s “how do we architect so that we can switch when the open version arrives,” because it will.

Abstract the model behind an interface. Don’t build your product logic around one provider’s specific quirks. Keep your prompts and your evaluation sets portable.

Final Thoughts

The demonstrations are what got the attention, and they deserve it. But the durable change from August isn’t that a computer can draw. It’s that a capability which was, for four months, the property of two companies, is now on a lot of people’s hard drives permanently.

That has consequences for pricing, for what gets built, and for who gets to build it. It also has consequences nobody has fully reckoned with, in exactly the way I worried about when adversarial generation first appeared eight years ago.

I’d rather have this capability distributed than held by two firms. I hold that view with less confidence than I’d like.