On a recent All-In, Jason Calacanis put the case against building on someone else's platform about as crudely as it can be put: nobody who partnered with Microsoft in the 80s, Facebook in the 2000s, or the frontier labs now "woke up without their throat slit."
Strip the theatrics and there's a real pattern underneath, one that predates AI by forty years. The enterprise version of this conversation — Palantir, Nvidia, sovereign nation-state AI — doesn't apply to you. The quieter SMB version does.
The pattern is older than the model layer
A platform opens a surface: an operating system, a social graph, a model API. Partners build on top of it. Their success reveals, through usage, exactly where the money is. Then the platform ships a first-party product into that spot, using its partners' own demand as the roadmap.
Windows did it to whole application categories. The social platforms did it to the apps that grew on their reach, then throttled the reach. The incentive now points at the model labs. Every prompt you send is a signal about where value is being created, and the party receiving that signal also sells software.
The clearest case is already here. In April 2026, Anthropic's chief product officer stepped down from Figma's board. Within days, Anthropic shipped Claude Design — a tool that turns prompts into prototypes and slide decks — into Figma's category. Figma's stock fell about 7%.
You don't have to believe any specific lab is acting in bad faith to take the incentive seriously. Incentives don't need permission.
The SMB version isn't "OpenAI will clone my company"
It won't. You're too small to be worth cloning, and cloning was never the real risk anyway.
The risk is quieter and closer to home. The layer where you think your advantage lives might be a layer you're renting. If the thing that makes your agency or your product special is "we wrote clever prompts on top of ChatGPT," that advantage has a landlord. And landlords raise rent, change terms, and sometimes move into the building themselves.
So ask a sharper question than which AI vendor do I trust: which layer of my business am I renting, and what happens the day I'm evicted from it? Four things to get straight.
1. Know which layer is rented
Draw the stack for one of your AI workflows. The model is the bottom layer. Treat it as a commodity you rent: interchangeable, repriceable, not yours. The layer you own is everything above it — your proprietary data, the workflow you've encoded, the judgment about when an output is good enough to send.
Most SMBs get this backwards. They treat the model as the special thing ("we use GPT-whatever") and the workflow as an afterthought. It's exactly inverted. The model is the part anyone can rent. Your edge is everything you wrapped around it.
2. Keep your alpha in a layer you control
Your "alpha" is the stuff that took years to earn and would hurt to hand over. Client lists. Pricing logic. The proprietary process. The prompt library your team tuned across hundreds of real cases. Name it explicitly. Most teams never have.
Here's the gut-check: if your AI vendor shut off your account tomorrow, what could you walk out with? For most teams the honest answer is nothing. The prompts, the tuned instructions, the accumulated context all live inside someone else's console, with no export button. You were renting the thing you thought you owned.
Then stop pasting that alpha into surfaces that learn from it. A consumer tier that trains on your inputs is not where it belongs. The fix is rarely dramatic: move the sensitive slice onto a zero-retention tier or a private deployment, and keep the crown-jewel data in systems you administer. "In the model's context" is a sentence. "In an API call under a no-training, no-retention agreement" is an answer.
3. Treat the model as swappable
If switching providers would cost you a full quarter of migration work, the vendor is holding the leash, not you. Build one thin abstraction between your workflows and whichever model answers the call. Keep it dumb. Its only job is to make the model one line you can change.
This is the cheapest insurance policy in AI right now, and almost nobody buys it, because everything works fine. Right up until the pricing email lands, or the model you depend on gets deprecated with 60 days' notice.
4. Match your sovereignty to your stakes
The enterprise crowd will tell you to self-host open-weight models on your own hardware. For most 20-to-500-person shops that's the wrong end of the ladder. You'd be trading a token bill for a full-time ops problem you aren't staffed to run.
The realistic middle: a zero-retention API for the bulk of your work, and a private or local deployment only for the narrow slice that's genuinely sensitive, or high enough in volume to pay for itself. Treat sovereignty as a dial. Turn it up where the stakes are real, and don't pay for control you'll never use.
The eviction test
None of this says don't build on AI. Build on it. The leverage is real, and the labs ship faster than any of us can. The point is narrower: know which floor you're standing on.
Before you wire a model into anything core, ask one question. If the provider changed the terms, raised the price, or shipped the exact feature I'm building tomorrow, how bad is my week? If the honest answer is "we'd be fine, we'd swap the model and move on," you own your layer. If the answer is "we'd be finished," you don't have a strategy. You have a landlord.
Reference: All-In Podcast — AI Sovereignty Wars, Palantir-Nvidia Deal (the origin of the platform-lock-in framing and the Calacanis quote above).