Why these limits last

How do we use these limits to judge new technology?

This mechanism is the only part of the map that will still be true in twenty years. Everything above it — formats, engines, orchestrators, the whole vocabulary of the modern data stack — is a set of temporary answers to questions this mechanism asks permanently. So use a model of the physical substrate to predict: given a new system, you should be able to guess most of its architecture from the four storage answers before reading its documentation. And when a new medium appears, you should be able to guess what will be built on it. That ability to predict is the difference between understanding this material and having read it.

The ratios are diverging

Architecture follows ratios, not absolute numbers. Capacity grew faster than bandwidth, bandwidth faster than latency, and latency is fixed by geometry. So bytes per operation and bytes per FLOP keep falling, which pushes every design towards moving less data and computing where the data already is: pushdown, near-data processing, computational storage, disaggregation (separating compute from storage, connected by a fast network).

The same divergence explains why the industry keeps rediscovering shared-nothing at smaller scales: cluster, socket, core, chiplet. As coordination becomes more expensive relative to computation, the best partition size gets smaller. The USL's β term is the reason. Expect this to continue.

Three limits still moving

1. Energy is becoming the binding limit. Moving a word off a chip costs orders of magnitude more energy than computing with it, and at datacenter scale power and cooling limit capacity more than silicon does. Energy behaves differently from money: it is locally conserved, linked to heat, and appears as throttling (latency variance) rather than as a bill. It belongs here because it is physics, not economics.
2. The layer boundaries themselves are eroding. Every abstraction here (the block device, the page cache, the socket, the filesystem) was drawn where a latency cliff was, and several cliffs have moved. When the kernel path costs more than the device, the abstraction no longer pays for itself. io_uring, SPDK, RDMA, ZNS (Zoned Namespace SSDs, which show flash's block-erase structure to software) and CXL are all this: the same negotiation, done again. An abstraction survives only while its overhead is small compared with what it hides, and when hardware improves 10×, the software layers around it become the bottleneck and get bypassed.
3. This is where you can be certain. Physics limits but does not dictate: above this mechanism the determining forces are economic and organizational, where evidence is weaker and folklore is common. The value of this mechanism is a fixed reference to recognize when a higher-level claim cannot be tested. When someone says a design is faster, ask: which conserved quantity did it economize, and where did the cost go? If there is no answer, it is marketing. This question is worth more than any single number here.