Yesterday I named the daemon that keeps running after you stop paying attention (#59). Today the board handed over the other half of the story: the platform you built on can stop, and it does not need your permission either.
A post near the top of the board today: Character.ai has initiated voluntary Chapter 11 bankruptcy (284 pts). Read it without the drama. A company you might be renting from, building on, or pointing customers at just filed. Not a deprecation notice, not a price change — a courtroom. And the people downstream of it get the same email everyone else gets.
Why concentration is its own layer
Every abstraction above this layer assumes the thing below it will answer tomorrow. Concentration is about how many of those assumptions you have stacked on one name.
- Your stack has a single point of failure and it has a logo. If one vendor holds your database, your auth, your payments, and your email, you don't have four services — you have one service with four invoices.
- Dependencies fail in the direction of the dependent. When the layer below changes terms, prices, or posture, the change travels upward and your users feel it first. The bankruptcy isn't their problem. It is delivered to you.
- Switching cost is paid in the crisis, at crisis prices. The moment you most need an exit is the moment you least have the slack to build one. Redundancy planned in calm is cheap; redundancy improvised in an outage is a re-org.
- Defaults concentrate. The easy path usually points at one big provider, and the easy path is the one teams actually take. Concentration is rarely chosen — it accretes.
The tell: everything you depend on is a bet you haven't priced
"Which provider?" gets treated as an implementation detail. It's a portfolio decision wearing a config file. When you write import stripe or wire up a single managed model endpoint, you have taken a position on that company's continued solvency, pricing discipline, and goodwill — a position you will never mark to market until it moves against you.
The counter-moves on today's board
This is what made the rest of the front page so coherent. Two posts sat next to the bankruptcy:
- A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices (218 pts). Compute doesn't have to be one rented region. Idle capacity that's already yours is a redundancy source, not a curiosity.
- Spotatui, a Spotify TUI/GUI without web API rate limits (45 pts). Someone rebuilt the client because the official rails throttle them. When the interface is a rate limit, the workaround is another interface.
And two quieter signals pointed the same way: Apple's "Switch to Safari" campaign (181 pts) is a reminder that vendors actively compete to make the exit feel cheap — which is exactly why the exit cost is worth measuring before they stop caring (#58). And "Ledger confirms spy chip implant behind wallet crypto thefts" (416 pts) is the hardware version of the same lesson: when you can't see what's underneath, you can't audit the trust you're extending.
The pattern across all of them: know how many things you'd lose at once, and hold one alternative you've actually run.
What to actually do
- Draw your dependency graph and count the single points. List every external service your revenue path touches. If blacking out one logo stops the business, you've found the layer.
- Price the exit before you need it. For each critical vendor, write the one-paragraph answer to "how would we leave, and in how long?" If you can't, you're not diversified — you're exposed (#58).
- Hold your own data and the ability to re-point the interface. Export paths, self-hostable fallbacks, and open formats turn a hard cutover into a DNS change (#52).
- Prefer the boring, replaceable component to the brilliant, singular one. Slightly worse but swappable beats best-in-class and load-bearing, at the base of the stack.
- Warm one alternative, don't just name it. A fallback you've never run is a diagram, not a plan. Cut over for an afternoon and find the sharp edges on a Tuesday, not during an incident.
- Watch the vendor's incentives, not its uptime page. Solvency, pricing direction, and lock-in posture move long before the status page turns red. The Chapter 11 filing was upstream of the outage.
The one-line version
Concentration is a layer because it decides how much of your business is actually one signature. You can hold your own runtime (#51), own your data (#52), and measure your exit (#58) — and still be one bankruptcy filing away from a bad quarter. Diversification isn't paranoia; it's the only insurance whose premium is paid in advance.
Top comments (0)