Stand under the machine and look up. In an AI data center, the machine you are standing under is not a turbine and not a rack of GPUs — it is the optical interconnect layer, the light running between servers. The load factors are precise and measurable: the market for optical modules and chips is the strongest semiconductor segment of this cycle, and the numbers on the table tell you exactly how fast the light is moving.
There is grandeur in a well-run plant, and the plant in question here is the photonics supply chain that has become the quiet bottleneck of AI scale-out. The numbers are not vague. In August alone, the A-share optical chip segment rose 27.09% over the interval from August 3 to August 31 — against 8.93% for the semiconductor sector as a whole. That is a segment outrunning its own sector by more than three times in a single month. In a market that normally moves in fractions, that gap is a signal with a spec sheet attached.
The specs: revenue and profit, measured
Precise specs, no sentimentality. Zhongji Innolight — a leading module maker — reported first-half net profit of 13.65 billion yuan, up 241.70%, on revenue of 41.78 billion yuan, up 182.49%. Read those two numbers together and the ratio tells you what matters: profit grew faster than revenue, which is what happens when pricing power meets capacity that is already loaded. The load factor is the story, and the load factor is full.
Then add the second-order effect, which is where the machine gets more interesting. Lianxun Instruments reported first-half revenue of 1.531 billion yuan, up 208.71%, and net profit of 567 million yuan, up 903%. The company’s own words: photonics module and chip makers are expanding capacity, and that expansion is pulling test equipment demand. This is the machine beneath the machine — when a segment expands, the tools that measure it expand first. A 903% profit number in test equipment is the instrument maker reading the gauge of everyone else’s expansion.
Let me think about the sheer size of what is being built, because it deserves more than a phrase. AI training clusters interconnect at scales that previous generations of data centers never approached, and the interconnect — the optics — is now as load-bearing as the compute itself. In a cluster of ten thousand accelerators, the optical layer moves more data than the compute layer holds. The grandeur is not in any single component; it is in the coordinated movement of light across a floor the size of a hangar, and in the tolerance arithmetic that makes it possible.
The chain, from photoresist to through-silicon vias
Follow the chain one layer down and the precision compounds. Domestically, i-line photoresist is being batch-imported into 12-inch wafer fabs, and ArF photoresist has entered customer validation — the photolithography materials that the photonics fabs depend on. And on August 19, JCET completed trial fabrication of high-aspect-ratio TSV through-silicon via technology, aimed at 2.5D and 3D advanced packaging — the packaging that stacks the chips behind the optics. These are not headline numbers; they are the tolerance stack, layer by layer, on which the whole machine rests.
I have stood in enough fabs to know what a batch-import milestone means on the floor. It means the material passed the qualification that matters: the yield gate. A photoresist that enters a 12-inch fab in batch is not a lab experiment; it is a production decision made with the fab’s own wafers and its own yield targets. The same logic applies to the ArF validation — it is at the point where the customer is running it against its own recipes, and that is where adoption is won or lost. No sentimentality about this: the specs either hold, or they do not.
Why the optics layer is the strongest segment
Watch the load factors over a year and you will see it: the strongest semiconductor segment is the one where demand is structural, not cyclical. Optics demand is structural because it scales with the square of the number of compute nodes — double the GPUs in a cluster, and the interconnect bandwidth does not double, it grows faster. That compounding is why the module makers can print 241% profit growth and the test-equipment maker can print 903%, in the same half, in the same chain.
It is also why the domestic photonics industry is moving from validation to volume. The trajectory in optics mirrors what happens in any well-run plant when the process stabilizes: first you prove it on samples, then you fill the line. The batch photoresist import, the ArF validation, and the TSV trial fabrication are all steps in that sequence — the move from proving a process to running it at load. The A-share segment’s 27% month is the market pricing that transition in advance.
The numbers deserve to be read as a machine, not as a chart. A 27.09% month and an 8.93% sector look like a spread; underneath them is a physical build-out — fabs qualifying photoresist, packagers qualifying TSVs, test companies shipping machines to everyone. The stock market is the lagging instrument in this ensemble. The leading instruments are the ones that show up in the test-equipment order book and the packaging qualification schedule. That is why the precise specs matter more than the daily tape.
Reading the machine the way an engineer would
Let me think about how an engineer would audit this story, because that is the discipline this segment deserves. The first audit question is whether the profit numbers are one-offs. They are not: 241.70% profit growth on 182.49% revenue growth is a margin expansion story, and margin expansion on a loaded line is sustainable as long as the line stays loaded — which is the whole point of the expansion phase. The second question is whether the test-equipment number is a lagging indicator. It is, in the best sense: test equipment is ordered in advance of capacity, so a 903% profit print in that segment is a forward-looking statement made by the people who sell to everyone else. The third question is whether the qualification work — photoresist import, ArF validation, TSV trial — is genuine or aspirational. A batch import into a 12-inch fab is the point at which materials stop being samples; a TSV trial fabrication is the point at which packaging stops being a slide deck. These are the markers an engineer trusts.
The answer to all three questions is the same, and it is the reason this read is an expansion story rather than a sentiment story: the numbers are physical. They are printed in earnings statements, measured in qualification gates, and booked in test-equipment orders. The machine is real because its components are real, and the load is real because the people who supply the tools to measure it are reporting 903% profit growth. That is the discipline of the engineer’s view: no sentimentality, just the load factors.
What the load factors will do next
Let me run the load factors forward, because a machine is only interesting if you can predict its output. The current state is a loaded line: module makers at 241.70% profit growth, test-equipment makers at 903%, and qualification work moving through photoresist and packaging gates. The next state depends on how much of that load is already contracted. Module capacity that is fully booked is a lagging indicator — it reports what was sold, not what will be. The leading indicators are the ones that point forward: the test-equipment order book, which is sold before the lines it measures exist, and the qualification gates, which determine whether the next wave of capacity can even start. If those two stay strong through the second half, the machine keeps spinning. If they soften, the profit prints will lag the turn by two quarters — and that lag is exactly why an engineer watches the leading indicators, not the earnings release.
There is one structural reason to expect the load to hold that has nothing to do with sentiment: the demand is locked into the compute build-out. Optical interconnect does not get optional upgrades like a phone does; it is specified at the rack level, designed into clusters that take quarters to procure and years to deploy. A cluster designed today carries its optical interconnect for its entire life. That is the difference between a discretionary component and a load-bearing one, and it is the structural reason the photonics segment can outrun its sector the way it did in August. The 27.09% month may correct, but the load it prices is booked well beyond the correction.
The honest caveat
Here is where I would put the honest caveat, because grandeur without precision is just noise. A 27% month can overshoot; segments that compound structurally still correct periodically. The durable part of this story is not the August move — it is the capacity expansion visible in the test-equipment order flow and the packaging qualification work. Those are physical, irreversible, and they take years. That is the difference between a spec sheet and a stock chart.
Let me also check the sequencing, because the order of operations tells you how much is already banked. The half-year earnings — 241.70% and 903% — are realized numbers, not forecasts. The qualification work — photoresist import, ArF validation, TSV trial — is in progress and not yet fully banked. So the machine has two gears running at once: one gear is reporting the output of a line that is already full, and the other is qualifying the next line behind it. Both gears are visible in the record, and that is precisely what makes this read as expansion rather than speculation.
The engineering view of this story is simple. Optics is the load-bearing layer of AI clusters, the demand is structural, the domestic supply chain has moved from validation to volume, and the evidence is measured in profit ratios and qualification milestones rather than adjectives. The machine at scale has three measured floors: module makers at 241% profit growth, test-equipment makers at 903%, and a materials-and-packaging layer moving through qualification gates. The sheer size of the AI interconnect build-out is real, and the domestic supply chain is inside it — not at the edge, but in the load-bearing layer.
And the final spec to hold on to is the one this segment is usually misread on: the market number is the least reliable number in the story. The 27.09% month is a price, and prices are what people believe at a moment; the earnings and the qualification gates are what the machine has actually done. An engineer reads the machine first and the tape second. Do the same here — measure the load, count the qualification gates, watch the test-equipment order book — and the 27% will stop mattering as a fact and start mattering only as a timestamp on a machine that is still running.
That is grandeur with a spec sheet, and the spec sheet is why it works. And it works because every claim above is a measured number — 27.09%, 241.70%, 903%, 182.49% — printed in the record, not estimated in the commentary.