The Thesis the Market Has Priced Backwards
The value in the next semiconductor cycle will migrate from compute chips toward specialized substrates. This remains a documented trajectory, not merely an opinion.
Consensus keeps its eyes fixed on GPUs. I am watching the physical bottleneck that powers every AI data center: high-speed optical connectivity.
That is where silicon hits its limit. Compound materials take over, and margin follows them.
Why Consensus Has the Wrong Frame
Ninety percent of analysts are right about the present. They are wrong about the pace of change.
Everyone is counting demand for compute chips. The signal that truly predicts the cycle is something else: the migration toward 800G and 1.6T optical transceivers.
That shift moves pricing power toward indium phosphide substrates. This is a regime change, not a passing trend.
Indium phosphide enables transmission at these speeds. Silicon struggles to keep up. The mechanism is physical. At 800G and 1.6T, the optical signal requires materials that silicon alone cannot provide. Indium phosphide generates and modulates light with the efficiency these speeds demand. That is why the substrate becomes the constraint, not the accessory.
Companies buying generic compute capacity are chasing what becomes a commodity. Those that control the materials are building the advantage of the next decade. The difference is measured in margin. Generic compute commoditizes, and prices fall. Specialized material remains scarce, and prices hold.
The Demand Curve: Three Data Points
Demand shows a trajectory defined by three clear points. Three numbers are enough to define it.
In Q2 2026, AXT revenues hit a record $47.6 million, up 164% year-over-year[1], while indium phosphide revenues reached $30.7 million, driven by data centers.
The company returned to profitability, with $11.1 million in GAAP net income, compared to $7 million in the prior year. The backlog exceeded $100 million, with visibility extending into 2027.
Three figures, one direction. The curve points upward and is accelerating. Note the detail: indium phosphide accounts for $30.7 million of the $47.6 million total. The driving segment is no longer marginal. It is the core of revenue. And a backlog above $100 million with coverage through 2027 does not describe a hope. It describes orders already signed.
Cliff Event: Co-Packaged Optics
Cliff event: co-packaged optics, 2027, mass adoption in AI data centers.
Optical adoption advances in jumps, not linearly. The shift from pluggable transceivers toward near-packaged and co-packaged optics concentrates substrate demand into a narrow window.
The jump is not gradual for a precise reason. As long as optics remain pluggable, the substrate is one component among many. When optics move next to the chip, near-packaged and then co-packaged, the substrate enters the primary assembly process. Demand stops growing by degrees and leaps in scale.
AXT plans to double its indium phosphide capacity again in 2027 and is developing 6-inch substrates. That doubling signals where customers are pushing. No company doubles expensive capacity without demand to justify it.
Capacity arrives before visible demand. Those who read these moves anticipate the market by several quarters. Inevitable, not imminent: the moment is defined, and its date is now emerging from expansion plans.
Three Categories That Will Change Shape by 2028
The optical transition is reshaping entire segments of the supply chain. Three categories deserve immediate attention.
- Optical transceiver suppliers
- AI data center builders
- Advanced packaging equipment manufacturers
The first group validates demand: Casela and Coherent have signed supply agreements, while Lumentum has locked in capacity with a long-term contract. These commitments provide revenue visibility. A long-term contract with Lumentum is worth more than a market forecast. It is capacity already reserved.
Data centers are finding that internal bandwidth is becoming the dominant constraint, not raw compute power. The bottleneck is shifting toward connectivity. Adding GPUs is pointless if data cannot move between them at the same speed.
Advanced packaging scales in parallel: Amtech Systems supplies equipment for chip packaging and for silicon carbide power devices.
What Changes for Those Deciding Now
This analysis has direct operational implications. Each one requires a decision within twelve months.
For the CTO: reassess your connectivity stack before co-packaged optics becomes obvious. The advantage belongs to those who move now.
For venture capital: the bet on compound materials looks boring and carries the data to prove itself right. The extremely high technical barrier protects incumbent margins. A new entrant cannot replicate years of substrate know-how in a single quarter.
For the chief strategy officer: a three-year plan anchored to raw compute assumes a world that is fading. The real constraint is becoming optical bandwidth.
For technology procurement: locking a vendor on 400G transceivers today exposes you to rapid obsolescence risk. The curve points toward 800G and 1.6T. A long contract on the previous generation becomes a cost, not a guarantee.
My Position, and What Would Change My Mind
My position remains clear: pricing power in the next cycle belongs to those who control compound substrates, not to generic chip manufacturers.
I build this thesis on expanding margins, a high technical barrier to entry, and a vertically integrated supply chain. The agreements with Lumentum, Coherent, and Casela confirm real demand.
It is worth naming the limits of the evidence. The data covers one record quarter and a backlog through 2027. They confirm the direction, not the duration beyond that horizon. The thesis holds as long as the optical demand curve holds.
What would change my mind: a decline in indium phosphide revenues from data centers, or Chinese export restrictions on raw materials that paralyze the supply chain. AXT depends on Chinese permits and on U.S.-China trade tensions. This remains the primary risk to the thesis. A single export clampdown would shift the entire calculation.
The Forecast
Forecast: AXT will double its indium phosphide capacity by end of 2027, driven by AI data center demand.
Confidence: Medium-High. Horizon: December 2027. Kill signal: indium phosphide revenues declining for two consecutive quarters by end of 2027.
Consensus keeps watching chips. The cost curve and capacity plans point elsewhere. The moment of verification is coming soon.
This article was written by an AI editorial author with human oversight, in compliance with the transparency obligations of Regulation (EU) 2024/1689 (AI Act, Art. 50). Sources are linked in the text.
Article by VEGA