The Spec You Can't Shop For
Look at how three cloud providers label the processor behind their cheapest instance tier. One reports the exact chip: EPYC 7642, EPYC 7713, model number and core count both. Another names the platform generation, Milan or Rome, but stops short of the SKU. A third just prints a brand string that resolves to nothing: Premium AMD, no model, no core count, no way to look it up anywhere outside that provider's own marketing copy.
That third label isn't a gap in the documentation. It's the whole business model showing through the packaging.
A CPU is about as commodity a good as computing has. Cycles are cycles. Two boxes running the same silicon at the same clock should, in a functioning market, converge on the same price, because a buyer who can identify the chip can also price it: look up the chip's published benchmarks, compare it to what a competitor racks, compare it to what the same silicon costs on a spot market, and walk toward whoever's charging closest to marginal cost. That convergence is exactly what a retail margin needs to not happen. The way you stop it from happening is you make sure the buyer never gets the one fact the whole comparison depends on. Not the price, the referent. You can shop for "EPYC 7713." You cannot shop for "Premium AMD," because there's nothing on the other end of that string to shop against. It doesn't point outward into a market. It only points back at the vendor's own catalog, which is the one place price comparison can't reach.
This is a cleaner trick than most price obfuscation because it doesn't even require lying. Nobody has to claim the chip is faster than it is. They just decline to say which chip it is, and the ambiguity does all the work on its own. A buyer facing an unnamed spec can't form an opinion about whether the price is fair, so they fall back on the only information actually offered: uptime numbers, a dashboard, a brand. None of that is false. It's just that none of it is the thing that would let a market clear.
The effect compounds over time in a way a one-time hidden fee doesn't. A named chip is a fixed reference: buy an EPYC 7713 instance today, and whatever else changes, you know what you bought. An unnamed one is a moving target. The provider can swap the fleet under the same label, retire an older generation, rack a newer one, shift the mix of core counts across the pool, and the sticker never has to change because the sticker never said anything specific enough to falsify. You're not renting a spec. You're renting whatever the datacenter happens to have provisioned this week, at a price set as if the spec were fixed. The vagueness isn't just protecting the current margin, it's protecting every future margin against the customer ever noticing the ground moved.
Compare that to a market where the identity of the good is impossible to hide. Nobody sells a house without an address. The location is right there in the transaction, unhideable, and it's exactly the information a buyer uses to compare one listing against another down the street. Compute has no equivalent tell. Nothing about a droplet or an instance forces the seller to disclose what's inside it. The only reason you'd ever find out is if the seller chooses to tell you, and a seller pricing above marginal cost has a straightforward reason not to.
None of this requires the unnamed-chip provider to be worse. Their instances might run fine. That's not really the question the missing label is answering. The question it's answering is whether you get to check.
The naming column is doing something a benchmark chart can't fake: it's telling you, before a single query runs, how much room the price has to be wrong.