Earlier this year, when Anthropic was forced to pull their most capable model "Fable", a frontier model provider was essentially cut off on short notice. A lot of companies discovered that their AI product had a single point of failure nobody had put in a risk register. At the time, we were running most of our stuff on the previous model, Opus, but still: we were shocked and started to think about what would happen if we were cut off entirely.
Here's what we came up with: we built a second door. Hear me out on this.
If your core features run entirely through one vendor's API, you have not bought a tool. You have essentially rented the thinking part of your product. Token price, availability and which model exists next quarter are decided somewhere else, by people who might have never heard of you, and don't care if your business ceases to exist.
The usual answer is a call for independence. For 30+ years, that's how we've always handled these kinds of situations. It's the reason we run a lot of open source software in our data centers, and run our own mobile core network. But when it comes to AI, independence is the wrong goal. Nobody running a mid-sized company will build chips or train a frontier model. What you can build is a second door: open-weight models running on hardware you control, or at a partner you could replace too.
That second door is the actual AI investment. Not the tokens you purchase from the large frontier providers. It's the GPUs, the power, the special data center contracts, the cooling. Yes, that's very expensive! It's an investment the size of a new factory building. We've never seen this kind of investment in our budget. But we've decided to build this second door because it signifies an exit. And the strange thing about an exit is that it pays off even if you never walk through it, because its existence changes every conversation you have with frontier model vendors. Suddenly, you're not negotiating multi-million-dollar token budgets with them. Small plans for the occasional outlier use case that needs a frontier model will suffice.
The next cutoff will come. It always does. The companies with a second door will have a slow afternoon. Everyone else will have a lost quarter and a very long all-hands.
Creating and maintaining a second door requires work and money.