The Shen family salt depot stood on the south bank; the salt market on the north. The gorge between them was too deep to bridge with stone, so there was only a chain span with narrow planks laid across it—wide enough for one carrying pole at a time. When the wind rose, the bridge swayed like a white cloth hung on a bamboo pole.

The first year, the depot had one porter, A-Qi. Ten trips a day, one load a trip, and at sunset he handed over the tally: ten loads.

The second year, the salt market grew hungry. Shen Yan hired.

Two porters: eighteen loads a day—not twenty, because crossing involved waiting, but she accepted it. Four porters: twenty-nine. Eight: thirty-two. When she reached the twelfth, the tally read twenty-nine—three loads fewer than with eight.

The porters were not lazy. At the bridgehead stood a single official scale, kept by Old Cheng. Before any load left the depot it had to pass over that scale: Old Cheng would read the beam, note the figure, and stamp the salt licence with a vermillion seal. One scale, one seal, for the entire depot. Hire more men and the scale does not sprout a second beam. So of every eight porters, one stood at the scale waiting, and the rest walked the bridge.

Shen Yan drew two lines on the wall: one for hands, one for loads. The hands climbed steadily to the right. The loads followed—then bent, as if someone had pressed a hand on their head.

She worked it out: the time the scale ate could never be chased back by adding men. This slice, later people would call contention. The ceiling contention imposes is a constant: however many seals the scale can stamp in a day, that is the most loads the depot can ship. If contention were all, twelve men would have been enough to creep up under the ceiling. No need to hire another.

But there was a second slice.

The chain bridge took one pole at a time. To keep two porters from meeting at mid-span, the rule was: before setting out, each porter had to check in with every other porter on shift that day—which span will you take, which will I, who goes first. Two men: one sentence. Ten men: forty-five sentences. Twenty men: one hundred ninety sentences. By the end, whole mornings went into nodding at the bridgehead.

This slice was neither constant nor straight—it grew in pairs: double the hands, quadruple the exchanges. Later people would call it coherency, or crosstalk.

So the bend did not merely bend. It began to fall.

The day the sixteenth porter joined, they shipped twenty-four loads. The twentieth: twenty-one. The twenty-fourth: eighteen. Shen Yan called the newest porter to the wall and pointed at the sliding tail of the line: “Look. One more man, one and a half fewer loads. Not because they won’t work—because they are busy asking each other the way.”

That night, on the back of the ledger, she wrote down a formula she would keep for life:

loads shipped = hands ÷ [1 + contention × (hands − 1) + crosstalk × hands × (hands − 1)]

It has three parts. With few hands, the bracket is still nearly one, and loads climb with hands—those are the good years. The contention term holds you down, so you can only creep along under the ceiling; that is what Amdahl said. But the crosstalk term carries “hands times hands,” and it grows crueller the further you go, until one day it outweighs the other two and the loads start walking backwards.

Shen Yan did two things.

First, she moved the scale off the path. No longer would every pole be weighed at the bridgehead; porters stacked loads ten to a pallet in the south warehouse, and Old Cheng went to the pallets three times a day, stamping a whole pallet with one seal. Contention had not vanished—the scale remained, the seal was still one—but each contention was now spread across ten loads. That constant thinned by more than half, and the ceiling rose.

Second, she tore down the pairwise check-in. At the bridgehead she set up a blackboard and hung a signal flag. Each morning she divided the bridge into six spans; whoever took a span wrote one line for himself; when the board was full, the flag went up, and no one was permitted to ask a second person anything. From “every pair exchanges a word” to “every man reads the board once”—double the men, double the writing, no longer fourfold.

The crosstalk term was cut nearly flat.

That autumn, twenty-seven men shipped one hundred twenty-two loads. Shen Yan knew it was not the best year available in the world. It was the best year available in her hands. She also knew span five on that blackboard still jammed most days—a new slice of contention was quietly growing there.

She drew the curve once more and marked the highest point of the bend: seven men, thirty-three loads. Beyond that—eight men, thirty-two; twelve men, twenty-nine—hands rising, loads falling. The books say that point is

√[(1 − contention) ÷ crosstalk]

Past it, every new hire walks you backwards. Hiring is not adding fuel to a fire; it is adding a tax—and the tax is levied on the square of the headcount.

By now you’ve probably recognized it: this is the Universal Scalability Law.

What it is

The Universal Scalability Law (USL), formulated by Neil J. Gunther, describes relative throughput with N execution units—threads, processes, nodes, instances:

C(N) = N ÷ (1 + α(N − 1) + βN(N − 1))

• α is contention (also called serialization): the small portion of work that must be done exclusively, one at a time. With α alone, the curve can only approach 1/α—exactly the asymptotic ceiling of Amdahl’s Law.
• β is coherency (also called crosstalk): the portion where any two units must exchange information and reconcile state. Its count is N(N − 1)/2, which makes this term quadratic.
• With both present, the curve rises, peaks, then falls—retrograde scaling: one more unit produces less total output.
• The peak sits at N* = √((1 − α) ÷ β). Cross it and adding machines loses money.

Because Amdahl’s Law is only the β = 0 special case of the USL, the USL sits much closer to reality: in real systems, no two units are silent to each other.

Why it matters

1. It explains “more machines, slower service”: teams blame luck, the cloud vendor, a framework bug. The USL says it is arithmetic. Any pairwise synchronization implies a peak, and past the peak is regression. Distributed locks, membership heartbeats, cache-line invalidation, the O(N²) messages of consensus protocols—all of these are β.
2. It splits tuning into two entirely different directions: cutting α (batching, precomputation, shorter critical sections, stamping the pallet instead of the load) raises the ceiling; cutting β (one bulletin board, one-way broadcast, sharding instead of pairwise reconciliation) moves the peak right. Pick the wrong direction and overtime only makes you hit the wall faster.
3. It gives capacity planning a hard boundary: fit α and β from a few measured points and you can compute how many concurrent units a system should ever run, and where expansion must give way to redesign. Far cheaper than “let’s load-test again.”
4. It is the criterion for splitting and cell-based design: β is proportional to “who must agree with whom.” Shrink the scope of strong consistency—partitioning, cells, per-tenant silos—and β falls on its own. That is why every system eventually goes to cells at scale.
5. It explains why autoscaling can hurt at the peak: elastic scaling adds instances in proportion to load. If load sits just right of the peak, extra instances lengthen the coordination chain and intensify crosstalk, turning a peak into a cliff.

Metaphor mapping

• Porter headcount → number of execution units N (threads / nodes / instances)
• Loads shipped in a day → relative throughput C(N)
• The one official scale and vermillion seal at the bridgehead → the exclusive critical section that generates α
• How many seals the scale can stamp in a day → the ceiling 1/α, the Amdahl asymptote
• Every load weighed before leaving → per-operation contention (a serial cost paid each time)
• Ten loads to a pallet, three pallet trips a day → batching, which spreads α thin
• A chain bridge that takes one pole at a time → shared exclusive resource
• Checking in with every other porter before departure → coherency cost β, N(N − 1)/2 exchanges
• Twenty men, one hundred ninety sentences → quadratic growth (double the hands, quadruple the talk)
• Sixteen men and twenty-four loads, twenty-four men and eighteen loads → retrograde scaling
• The marked peak of the bend: seven men, thirty-three loads → N* = √((1 − α) ÷ β)
• One blackboard and one flag, each man writing a line → one-way broadcast / bulletin board, O(N²) collapsed to O(N)
• Six spans, each written by its own porter → sharding / shrinking the scope of consistency
• Span five jamming again → a new contention point born of sharding; α drifts
• A-Qi alone, ten loads → the N = 1 baseline the curve divides by
• “Hiring is not adding fuel, it is adding a tax on the square of the headcount” → the quadratic nature of synchronization cost
• The contention slice that can never be chased back → the non-parallelizable serial fraction
• Nearly four and a half loads per man after both cuts → lowering β moves the peak right, lowering α raises the ceiling