←2026-08-30 · VM Genius
分析
AXT Inc. doesn’t make GPUs. It doesn’t make switching ASICs. It doesn’t build 800G or 1.6T optical modules. Optical communications chips aren’t even its core business.
And yet, as the AI optical networking trade keeps broadening, AXTI is showing up more and more frequently on investors’ screens.
There is really one reason:
InP — indium phosphide.
More specifically, AXT sells InP wafer substrates.
It sounds obscure. That is exactly why the setup is interesting.
While investors are busy chasing optical transceivers, silicon photonics, and co-packaged optics, one of the least glamorous components sitting at the very top of the supply chain may be turning into one of the hardest bottlenecks to solve.
The real AXTI trade is not simply “AI.”
The real question is:
If 1.6T optics, silicon photonics, and CPO all ramp at the same time, will there be enough InP supply?
That is the core of the thesis.
For the past two years, the central bottleneck in AI infrastructure was compute. Everyone needed more GPUs.
But once GPU clusters scale from thousands of accelerators to tens of thousands, and eventually hundreds of thousands, the bottleneck changes.
It doesn’t matter how many GPUs you own if those GPUs cannot move data between each other fast enough. Cluster efficiency starts becoming a networking problem.
That is why AI infrastructure is gradually shifting from a compute race to a connectivity race.
Network speeds are moving rapidly:
100G → 400G → 800G → 1.6T → eventually 3.2T
The faster the network gets, the more traditional copper interconnects run into problems with distance, power consumption, and signal integrity.
The natural consequence is that more data transmission has to move from electrons to photons.
That is why 800G, 1.6T, silicon photonics, and CPO have become major new branches of the AI infrastructure trade.
But investors tend to focus on the optical module.
They often forget to ask the more fundamental upstream question:
Where does the light actually come from?
One major answer is the InP-based laser.
And before you can manufacture an InP laser, you first need a high-quality InP wafer.
That is where AXT sits.
The supply chain can be simplified like this:
Raw materials
→ InP single crystal
→ InP wafer substrate
→ epitaxy
→ DFB / EML lasers / photodetectors
→ optical engines / silicon photonics PICs
→ 800G / 1.6T optical modules
→ AI data centers
AXT operates near the very top of that chain, supplying InP wafer substrates.
That means if downstream companies such as Lumentum and Coherent want to expand laser and optical-device capacity, they first need enough wafers upstream.
That is what makes AXTI interesting.
It is not directly monetizing end-market AI spending.
It is monetizing the upstream material demand created by the entire industry’s capacity expansion.
More GPUs mean more data.
More data means more optical links.
More optical links mean more lasers.
More lasers mean more InP wafer consumption.
So stripped of all the AI branding, AXTI is actually a very old-school investment thesis:
an upstream supply-demand mismatch.
A common question when investors look at AXTI is:
If silicon photonics is the future, why do we still need InP?
Because silicon photonics does not mean everything is made from silicon.
Silicon is excellent for waveguides, modulators, and large-scale photonic integration.
But silicon has one fundamental weakness:
It is not very good at generating light efficiently.
That means many silicon photonics architectures still require external lasers, and those lasers are frequently built on InP-based material systems.
So the future may not be:
Silicon replaces InP.
It may be:
InP generates the light. Silicon processes the light.
That distinction matters.
Higher silicon photonics penetration does not automatically mean lower InP demand.
In some architectures, it could actually mean demand for more high-performance lasers.
The same logic applies to CPO.
Co-packaged optics is designed to solve the problem of high-speed electrical signals traveling too far across a PCB. By moving the optical engine next to the switching ASIC, the system converts electrical signals into optical signals much earlier.
But the light source does not magically disappear.
A future architecture may still look something like:
Switching ASIC + Silicon Photonics + Laser
And InP remains one of the most important material platforms for high-performance laser sources.
One of the biggest potential catalysts for AXTI is 6-inch InP wafers.
Historically, InP manufacturing has been concentrated in 2-inch, 3-inch, and 4-inch wafers.
If 6-inch wafers can reach commercial maturity, the economics can improve materially.
A 6-inch wafer has roughly 2.25 times the surface area of a 4-inch wafer.
In theory, that means more devices per wafer, greater manufacturing automation, and potentially lower unit costs.
But there is a catch:
InP is not silicon.
Silicon manufacturing has been scaled successfully to 12-inch wafers.
InP is a compound semiconductor, and scaling wafer diameter makes crystal quality much harder to control. Defect density, dislocations, uniformity, and yield all become increasingly difficult as wafer size increases.
So moving to 6-inch InP is not simply a matter of making the wafer bigger.
It is a genuine technology and manufacturing barrier.
Whoever can stabilize 6-inch production first could gain a meaningful cost and supply advantage during the next optical communications capacity cycle.
That is why AXT’s progress on 6-inch InP deserves to be monitored separately.
When I analyze upstream materials companies, I do not care very much when management says on an earnings call that “AI demand is strong.”
The more useful signal is much simpler:
Are customers willing to put real money down to secure future capacity?
That is where AXTI becomes significantly more interesting.
Coherent has already been working with AXT around 6-inch InP and has provided financial support tied to capacity expansion.
Lumentum has also entered into long-term InP wafer supply and capacity reservation arrangements.
That matters.
When customers are willing to lock up capacity years in advance, they usually are not primarily worried about price anymore.
They are worried about one thing:
Will the material even be available?
That is when an industry narrative starts shifting from:
“Demand is strong”
to:
“Supply may be tight.”
And that is often the stage where upstream materials stocks generate their biggest operating leverage.
Once customers stop negotiating for the lowest possible price and start competing for available capacity, bargaining power can begin migrating upstream.
InP is not a commodity silicon wafer market.
The market itself is relatively small, but manufacturing is difficult, customer qualification cycles are long, and high-quality supply has historically been concentrated among a handful of vendors.
The market is effectively dominated by three major suppliers:
Sumitomo Electric: 40.2%
Beijing Tongmei / AXT: 38.3%
JX Metals: 13.1%
Together, they represent approximately 91.6% of the market.
That creates a very specific supply structure.
Historically, the market was not large enough to justify massive capacity additions.
Then AI suddenly accelerated demand for optical connectivity.
Downstream customers began reserving capacity in advance.
But upstream supply cannot ramp overnight.
That is where the setup becomes interesting.
The biggest risk in a materials market is not simply high demand.
It is:
Demand accelerating faster than supply can respond.
This type of capacity cannot be created just by throwing money at a factory.
Crystal growth, wafer slicing, polishing, defect control, yield optimization, and customer qualification all take time.
So AXTI’s upside may not ultimately come from the absolute size of the InP market.
It may come from something more powerful:
extremely low supply elasticity.
A small market can still create enormous earnings leverage if a sudden supply deficit develops.
Strip away the narrative and there are really only five things investors need to watch.
First, can InP revenue continue accelerating?
Second, does capacity expansion stay on schedule?
Third, what happens with 6-inch InP yields, customer qualification, and volume production?
Fourth, do major customers such as Lumentum and Coherent continue increasing purchases, prepayments, or long-term capacity commitments?
Fifth, does anything change around China-related supply-chain exposure and export licensing risk?
And if you are wondering where AXT’s upstream production exposure sits, take a closer look at Beijing Tongmei.
There is another layer of the story there.
The first four variables determine the upside.
The fifth determines the tail risk.
The real reason AXTI deserves attention is not whether the market can slap an “AI stock” label on it.
The real questions are:
How much InP will the industry need if 800G, 1.6T, silicon photonics, and CPO all scale simultaneously?
And:
Can the existing supplier base expand high-quality capacity fast enough to meet that demand?
If demand grows normally, AXT is simply another upstream materials company.
But if AI data centers push high-speed optical interconnects into an entirely new volume regime, while InP supply fails to keep pace, then the market is no longer trading an “AI theme.”
It is trading one of the oldest setups in the book:
Demand accelerates. Supply lags. Customers lock in capacity early. Pricing power migrates upstream.
That is the real reason investors are starting to take another look at AXTI.
Not because the company suddenly learned how to tell a better AI story.
But because the supply chain itself is starting to send a very different message — backed by actual capital commitments:
InP supply may be a lot tighter than the market assumes.
AXT’s latest operating results are where the thesis starts becoming much harder to ignore.
For Q2 2026, revenue surged 77% sequentially to $47.6 million, setting a company record.
InP revenue also reached an all-time high.
GAAP gross margin expanded to 44.9%, demonstrating just how powerful the operating leverage can become when higher-value InP volumes move through the model.
Non-GAAP net income reached $11.9 million, compared with a non-GAAP loss of $0.15 per share in the year-ago period.
GAAP net income came in at $11.1 million.
So yes, the valuation is already aggressive.
AXTI currently trades at roughly 77.7x FY1 non-GAAP earnings.
That is not cheap by any definition.
But the multiple falls sharply further out:
FY2 non-GAAP P/E: ~29.3x
FY3 non-GAAP P/E: ~19.8x
That is ultimately what the market is underwriting.
If AI infrastructure continues driving strong InP demand, if AXT can execute on its capacity expansion, and if 6-inch InP scales successfully, today’s elevated valuation can theoretically be absorbed by earnings growth.
But that also tells you exactly where the risk is.
At nearly 78 times forward earnings, AXTI is no longer priced for a merely “good” InP cycle.
It is increasingly priced for the thesis to work.
And that makes the stock much more interesting — but also much less forgiving.