The most important change in optical communication is here!
By: 美股投资网
At present, despite a clear shortage of DSPs and top-tier clients still scrambling to secure production capacity, the entire industry has already started researching how to reduce DSP usage, or even ultimately phase it out.
This is actually the most significant transformation to track in the next phase of optical communications.
This year’s profits came from capacity expansion; going forward, the gains will come from architectural changes.
Let’s first clarify the current situation.
The 1.6T boom is not over. According to estimates from our US Stock Investment Network research, demand for 1.6T optical modules could reach 60 to 65 million units by 2027, with DSPs remaining a relatively tight link in the supply chain.
Especially as 1.6T DSPs have essentially entered the 3nm era, capacity is not as abundant as with older process nodes. Currently, about 70% of 1.6T DSP orders are locked in by long-term agreements. Leading AI clients like GOOGL, NVDA, MSFT, AMZN, and META, due to large order sizes and closer ties with vendors, naturally have higher purchasing priority.
In other words, over the next one or two years, 1.6T will still ramp up, and DSPs will remain in tight supply.
However, from an investment perspective, this is no longer particularly new information.
If you’re still obsessively tracking “how many more 1.6T orders” or “how much DSP is lacking,” you’re essentially trading on topics that have already been repeatedly discussed over the past year.
Instead, we prefer to look one step further ahead.
Why is the industry trying to eliminate DSPs?
Ultimately, it’s still an old problem—power.
For AI data centers, the most pressing issue now is not just securing enough GPUs, but where the electricity will come from after acquiring GPUs, how to dissipate the heat, and exactly how much power the entire network will consume moving all this data.
There’s an example in our latest research that illustrates this well.
An average 800G optical module consumes about 25W, with the DSP accounting for nearly half. If, due to process differences, a single DSP consumes an extra 3W—this may seem minor, but accumulated over a 128-port switch, that’s nearly 400W.
Now, extend this number to AI clusters composed of hundreds of thousands or even millions of accelerators, and the impact becomes dramatically different.
Moreover, the network continues to upgrade: after 800G comes 1.6T, and eventually 3.2T. The higher the data rate, the longer the distance electrical signals must travel on the PCB, making signal loss and power consumption ever harder to manage.
At this point, simply relying on “deploying a more advanced DSP” to resolve these issues is an increasingly limited solution.
NPO and CPO have emerged to the forefront against this background.
In 2027, my attention will first be on NPO.
You can roughly understand NPO as moving the optical engine, which used to be distant from the switching chip, much closer to it.
By shortening this distance, the electrical signal no longer needs to travel as far on the PCB, lowering loss and naturally reducing the DSP's role and overall power consumption.
According to our estimates in this research, NPO can save about 30%–40% in power consumption compared to traditional solutions.
What’s more noteworthy is 2027.
Everyone now agrees that CPO is the direction, but the problem with CPO is that, while the concept looks promising, mass production isn't so easy. If some CPO switches can’t be shipped as planned next year, NPO will likely step in first as a transitional solution for a while.
This research even estimates that if NPO is widely adopted in these relevant switches, based on the number of ports, the corresponding scale in 2027 could reach around 12 million units.
I won’t take this 12 million too literally, as there are many assumptions here, and it remains to be verified how much CSPs will eventually adopt and whether products can be delivered on time.
But I will remember the change “behind 12 million”: NPO is no longer a pie-in-the-sky concept.
As soon as we begin to see real orders, real switch shipments, and tangible CSP validation next year, market trading logic will shift accordingly.
This is why in 2027 I’ll be paying close attention to NPO.
Looking even further ahead, CPO is still inevitable
NPO is more like a bridge.
What could truly overhaul the current optical interconnect architecture is still CPO.
CPO takes things even further than NPO by bringing the optical engine even closer to the switching chip. The shorter the distance, the lower the power consumption can go.
Our research shows NPO could save 30%–40% power, while CPO, in some tests, can achieve 60%–70% savings.
If future AI clusters actually scale to the million-GPU level or beyond, these differences ultimately translate into electricity bills, cooling requirements, and the overall data center capacity.
But CPO faces a problem: its main difficulties now are not in the design but in manufacturing.
There may be over a thousand optical fibers involved—including how to couple them, improve yields, and control costs, all of which involve intrinsic challenges. Add to that glass substrates, routing, drilling, and advanced packaging—many of these processes are still maturing toward true volume production.
So for now, we’re not declaring “death” for traditional optical modules just because CPO is emerging.
The time window provided by this research is actually quite reasonable: until 2028–2029, DSPs and traditional optical modules will likely still have their place. True, visible technological inflection points may not arrive until around 2029, or even 2030.
There are still a few years to go.
From an investment perspective, this is good news—it means the transition won’t be sudden, but rather a process we can observe and verify continuously.
This also changes our stock-picking mindset. In the past, when evaluating optical communications, our first question was often: Whose optical modules sell the most?
Soon, the question may need to shift: Who controls the architecture of the next-generation AI network?
While these two questions sound alike, the underlying investment logic is completely different.
In traditional architectures, switches profit from switches, DSPs profit from DSPs, and the value is spread between optical chips and modules.
But with NPO—especially CPO—switching chips, silicon photonics, optical engines, and advanced packaging are increasingly intertwined.
Once integration begins, the distribution of value will inevitably be reallocated.
So, in addition to familiar optical communication companies like LITE and COHR, I’ll be focusing more on NVDA, AVGO, and MRVL.
Because their contest is no longer just about a single chip, but about how much they can control within the next-generation AI network.
There’s also another company I’ll be adding to my watchlist: INTC.
If, in the future, AI chips, switching chips, and CPOs all compete for advanced manufacturing and packaging, the entire industry will sooner or later ask: Is there a truly viable second option beyond TSM?
This research has already mentioned that some vendors are assessing INTC’s EMIB and advanced process nodes.
Of course, at this stage we can’t estimate INTC’s revenue just based on this, but if there’s really a shift of advanced packaging capacity in the future, then INTC’s story will be about more than just whether its CPUs can stage a comeback.
From here, I’ll keep my eyes on these three things
When NPO sees truly large-scale orders, when CPO crosses the mass-production threshold, and which among NVDA, AVGO, and MRVL can ultimately claim more of the value in next-generation AI networks.
This year, the optical communications rally was about “volume.”
For the next stage, the focus is on who will capture the “architecture.”
Disclaimer: The content of this article solely reflects the author's opinion and does not represent the platform in any capacity. This article is not intended to serve as a reference for making investment decisions.
Trade now！
You may also like
Crypto prices
More
Become a trader now?A welcome pack worth 6200 USDT for new users!
Sign up now