Four of these are distances. One is a direction.

Somebody sends you a slide. It shows six phrases in a row, as though they were rungs of a ladder: scale-inside, scale-up, scale-out, scale-across, scale-in, scale-beyond. The arrow at the bottom says more scale, and it points right.

Try to put them on one axis, and the axis does not hold. Of the six phrases:

• you can place two with confidence.
• you saw one used in two incompatible ways.
• you found one in the material of only one vendor.

By the time you know this, the meeting is on the next topic. This happens constantly, and the cause is not a slow audience. The slide combines several vocabularies that happen to share a prefix.

So it is worth doing properly once.

The ladder that does exist

Four of the six really are distances, and they really do nest. Each one contains the one before it. Each boundary is where some physical fact becomes impossible to ignore.

Start at the bottom, with the newest of the four. Marvell uses scale-inside for “linking chips within server or switch trays.” That sounds like the easy case, and it is not. Marvell’s own demonstration is 2.5 meters of cumulative copper with 48 dB of end-to-end bump loss. Eight lanes hold an error rate of one in a hundred billion.

Two and a half meters, inside a tray. A distance that you could measure with a school ruler now needs a 48 dB budget. That is the entire reason the tiers above it are under pressure.

Scale-up is the coherent accelerator domain. Here the boundary is not really distance at all. It is semantics. NVIDIA calls the fifth generation of NVLink “a scale-up interconnect,” and the physical unit is a rack: seventy-two Blackwell GPUs in one NVLink domain, 1.8 TB/s per GPU, 130 TB/s aggregate. The current generation moves to 3.6 TB/s per GPU and 260 TB/s across a Vera Rubin NVL72 rack. NVIDIA puts this at more than fourteen times the bandwidth of PCIe Gen6.

That multiple is the argument for the tier. Inside a scale-up domain, the design expects the machines to behave as one machine. This means that every accelerator can talk to every other at close to full rate. Designers do not build a general-purpose bus for that.

Scale-out is the oldest word on the list and the least interesting. This is worth saying plainly, because people treat it as the new thing. Scaling out meant adding machines instead of buying a larger one, and it is decades older than AI. In this context it is the cluster network across racks and rows. NVIDIA’s own pairing is Quantum InfiniBand and Spectrum-X Ethernet. At the boundary under it, you abandon one-machine semantics and use message passing, which is a software change as much as a wiring change.

Scale-across is the newest rung at the far end. It exists for a blunt reason that is not about networking: at single sites, power became the limit before demand did. NVIDIA describes Spectrum-XGS as a way to connect distributed data centers into one AI factory. NVIDIA claims up to 1.9 times the collective-library bandwidth of off-the-shelf Ethernet. The interesting thing about this tier is not the technology. Its existence is an admission about electricity.

The vendors do not agree where the lines are

Here is a detail worth keeping. It makes a taxonomy feel unstable while you learn it. The instability is real, and it is not your misunderstanding.

NVIDIA separates scale-up and scale-out sharply. They are different products, different fabrics, different silicon. Marvell writes about the same physical territory but combines them. For Marvell, scale-inside is chips within a tray, and “scale-up or scale-out” is the single bucket for “networks linking trays within a rack.”

Both are defensible. A company that sells an accelerator domain has every reason to draw a hard line around it, because that line is the product. A company that sells the SerDes (the serializer-deserializer that drives the copper) cares about the electrical problem. At the electrical layer, the tray-to-tray case looks like one problem, whatever you call the network on top of it. The words describe commercial boundaries at least as much as physical ones. A reader who expects a clean physical taxonomy will continue to find it confusing, because the taxonomy is partly a market map.

The one that is a direction

Scale-in is the source of most of the actual confusion, and the cause is not subtle once you see it. It first belonged to cloud elasticity. There the pair is scale-out and scale-in, and the axis is capacity, not geography. Microsoft states it directly: elastic scaling is “the ability to use as much capacity as you need, scaling out as load increases, and scaling in when the extra capacity isn’t needed.” Its guidance has a section with the heading design for scale in. That section tells a running application how to continue when the platform removes instances from under it.

AWS uses the words the same way throughout its autoscaling documentation. There, a scale-in activity terminates instances. So in its native vocabulary, scale-in is the opposite of scale-out. It means fewer machines, not shorter distance.

Now put it on a slide next to scale-up and scale-across, and see what happens. The reader just learned that every other prefix names a place. So the obvious inference is that scale-in names the innermost place. But that place already has a name: scale-inside. The two words are one letter apart and mean entirely unrelated things.

In 2026 NVIDIA added a second meaning. It calls scale-in the fifth pillar of its AI networking: the layer that secures, manages and operates an AI factory. It runs on BlueField-4, a data processing unit (DPU) that takes infrastructure work off the host CPUs. NVIDIA presents it as the next form of the north-south network, which carries users, storage and security services. SDxCentral reported that the first description of it was thin.

So scale-in now has two vendor meanings, and neither is a distance. One is a direction along a capacity axis. The other is a layer of services around the compute.

The practical rule is this. Use the conversation to find the meaning:

• autoscaling, capacity or cost: scale-in means removing machines.
• DPUs, security or storage access in an AI factory: it means NVIDIA’s services layer.
• trays, packages or SerDes: neither meaning applies, so ask whether the speaker means scale-inside.

The one without a definition

Scale-beyond is a recent word. Nokia’s material for the OCP Global Summit in October 2026 lists scale up, scale out, scale across and scale beyond, as ways to connect distributed compute locations. It gives no reach and no definition that separates scale-beyond from scale-across.

Each rung on the ladder is the reach of one fabric under one design. A word without a stated reach cannot take a rung. So this page keeps scale-beyond off the ladder, next to scale-in, but for a different reason.

In the material we checked, only Nokia uses the word. The NVIDIA networking, NVLink and NVL72 pages do not use it, and Marvell’s blog does not use it. Until other vendors adopt it, treat it as one company’s term.

Why the list keeps getting longer

Look at the whole list and the pattern is clean. A new prefix appears when a boundary that was previously uninteresting becomes expensive.

Nobody needed a word for the inside of a tray while a few inches of copper were free. They need one now that it costs 48 dB. Nobody needed a word for spanning sites while one building could hold the cluster. They need one now that a site’s limit is its substation. Each new rung is a place where the old solution no longer scaled. The vocabulary is a record of where the difficulty moved.

The pattern also tells you where to expect the next one. The rungs that grow are the two ends, not the middle. Both ends grow for reasons that the network did not choose: electrical loss at the bottom, electrical supply at the top. Scale-beyond, if it settles into a meaning, will sit at the top end. The middle of the ladder is the part people usually mean when they say “the network.” That part is stable, and it was stable for years.

What to take away

Four words name nested distances:

• inside a tray.
• across a coherent accelerator domain.
• across a hall.
• across separate data centers.

Learn those four, and the ladder is genuinely useful. Each boundary marks a real change in what the interconnect has to do.

The fifth word, scale-in, is not a distance in either of its vocabularies. In the cloud it means removing capacity. At NVIDIA it means the services layer around the compute. When you hear it, identify which vocabulary the speaker uses before you agree with them.

The sixth, scale-beyond, is one vendor’s word with no published definition yet. It is not a shared term.

None of this is pedantry about labels. Each rung is a different engineering problem with a different supplier list, and that is the reason to order the ladder correctly. A conversation can move between rungs without anyone noticing. In that conversation, two people agree on a sentence and disagree about the thing.