| | | Latest from Rddt Gabbeb0y
The potentially groundbreaking breakthrough
Due Diligence
The potentially groundbreaking breakthrough
During the recent months I´ve tried to dig really deep into Poet, joblistings, potential customers/product-chains, current technology-moats, future technology, and I've posted some of it. Now it is time for the grand finale. In this post I would like to talk about complete wafer optical transceiver production and how that would destroy all of Poet's competitors, i.e how that would make all of us shareholders dance and celebrate. And also a little spicy golden nugget at the end of the analysis.
“Wafer scale assembly of optical transceivers”
In a recent interview with the chief revenue officer, there were multiple exciting and bullish statements, but I would like to dig into this one: “And our main goal is this wafer scale assembly.” “We really want to get to a wafer scale assembly of optical transceivers. And we are developing other technologies that assist in that.” “So we are developing technologies that eliminate all these active alignments and multiple steps in that process. So we are, I think, getting there probably this year, we will get to a demonstrable solution that eliminates all these components. So that will be a big deal for the industry because then truly we can do a wafer scale assembly of optical transceivers.”
Here is my view on the possible transformation (since I am no engineer or optical expert, please verify this/extend the explanation):
- Current State: "Wafer-Scale Optical Engines" What they make: The "Engine" (The heart).
The Process:
They take a Silicon Wafer (Interposer).
They "flip-chip" the laser and detector onto it (Passive Alignment).
The Bottleneck: They chop this wafer into tiny chips. Then, they send these chips to a factory (like Luxshare) where humans or slow robots must:
Glue the chip into a metal box (transceiver housing).
Add external plastic lenses to focus the light.
Add a fiber optic connector.
Seal the box hermetically (to keep air out).
The Limitation: Even though the chip was cheap to make, the packaging (the metal box and lenses) accounts for 70% of the cost.
2. The 2026 Goal: "Wafer-Scale Transceivers" (The CFO's Vision) What they make: The "Whole System" (Engine + Package + Optics) all at once.
The "Eliminated Components": The CFO said they will "eliminate all these components." Here is specifically what he means:
No More Lenses: POET integrates "Spot Size Converters" directly into the silicon. They guide the light into the fiber without needing an external plastic lens.
No More Isolators: (These are expensive magnets that stop light bouncing back). POET is designing the laser-to-chip interface (with Sivers) so these are often unnecessary.
No More "Gold Boxes": Instead of putting the chip in a metal can, they seal the chip while it is still on the wafer using a glass cap. This is called Chip-Scale Packaging (CSP).
The Result: When they cut the wafer, the device is Done. It doesn't need to be put in a box. It is the box.
Verdict/the key take away:
Passive Alignment (Today): Solves the "Laser Placement" problem. (Great, but incomplete).
Wafer-Scale Transceiver (2026): Solves the "Packaging" problem. (The Holy Grail).
In my opinion Poet is already ahead of its competitors technology-wise because of their passive alignment, “Lego-brick assembly”, production-infrastructure, owning specific, very important machines at Silterra (from what I've learned), reduction of indiumphospide to the only the critical parts(and something about TLFN or what it´s called), speed (as good as competitors, but likely faster in 2 or 3 years). Now, the CRO is saying “So we are, I think, getting there probably this year, we will get to a demonstrable solution that eliminates all these components.”In other words: we are probably going to show you something that will revolutionize the industry this year.
The Sivers connection
Sivers Semiconductors has quietly positioned itself as the indispensable "fuel source" for the next generation of AI hardware, transforming from a niche component manufacturer into a critical platform enabler for the silicon photonics industry. While the market often focuses on the giants like Marvell or the integrators like POET Technologies, Sivers holds the one piece of intellectual property that neither of them can manufacture themselves: the high-power, continuous-wave (CW) laser light required to power optical computing.
The company's recent strategic shift to marketing "Advanced Photonic Wafers" rather than individual laser chips is the most significant clue to their future role. In the traditional optical supply chain, lasers are grown, chopped into thousands of tiny squares, and sold as individual units that must be actively aligned—turned on and moved around by robots until the light is focused—inside a gold box. This process is slow, expensive, and completely unscalable for the millions of connections needed by AI clusters. Sivers has solved this bottleneck by developing the InP100 Platform, a manufacturing process that allows them to produce Indium Phosphide lasers on 4-inch wafers that are designed specifically for "flip-chip" bonding. These lasers effectively have "guide rails" etched into them, allowing them to be dropped onto a partner’s silicon wafer (like POET’s optical interposer) with sub-micron precision, without ever needing to be turned on during assembly. This capability is the definition of "wafer-scale" manufacturing, moving optics from a craftsman-like assembly process to a mass-production printing process.
This technology is the direct enabler for the "2026" production targets cited by partners. The high-power lasers Sivers is demonstrating—capable of delivering up to 150 mW per channel—are not designed for standard pluggable transceivers, but for External Light Source (ELS) modules. These are the "batteries" that will power the optical fabrics of companies like Celestial AI (now under the Marvell umbrella). Because the actual AI processors generate too much heat for a laser to survive next to them, the lasers must be housed remotely in these ELS modules and piped in via fiber. Sivers is one of the few foundries in the world capable of producing these high-power, high-reliability lasers at the volume required for this new architecture.
Financially, Sivers sits in a unique and arguably temporary position. After halting the spinoff of its Photonics division in early 2025 due to poor market conditions for SPACs, the unit remains a highly undervalued asset trapped within a broader semiconductor holding company. This "for sale" status, combined with their technological lock-in with POET Technologies, makes them a prime acquisition target. They have already secured the supply chain role; they provide the raw photonic wafers that POET’s interposers need to function. As the industry moves toward the commercial launch of the Celestial AI photonic fabric in 2026, Sivers has effectively become the exclusive gas station for the new AI superhighway—making them either a massively profitable standalone vendor or the next logical purchase for a partner looking to secure their supply chain.
The final verdict
Talking about groundbreaking technology I'd like to bring up something truly incredible. Looking forward, into the 2030s this “thing” could be massive. I am talking about Ferroelectric photonic memory. It's distant, but would be incredibly, immensely amazing. Bringing all of the bullish factors together, recent job hirings, celestial/marvell connection, statements, the interview of the CRO, technology advancements, industry maturing, the CRO hinting about POs during the interview(indirectly or directly depending on how you view it). These are astonishingly interesting times and I wonder how it all plays out. There are of course risks and so on, but the current outlook is incredibly bullish, IMO. Now I would like some help. Please help me research ferroelectric photonic memory and what that would mean for the market. Please post this on Agoracom if you´d like. Here is a little explanation about Ferroelectric photonic memory, but it´s not much ( I know it´s AI you don´t have to tell me in the comments, but it´s refined):
- What is "Ferroelectric Photonic Memory"? Remember how I said, "You can't store a photon because it always moves"? This technology cheats that rule. instead of trapping the photon, it uses a special material (Ferroelectric Lithium Niobate) that "remembers" a state.
How it works: You hit the material with a voltage pulse. The material's crystal structure physically flips (like a switch) and stays that way without power (Non-Volatile).
The "Memory": When you send light through it later, the light changes behavior based on whether the crystal is flipped "Up" or "Down."
The result: You have effectively created a Photonic Hard Drive. You can store 0s and 1s on an optical chip without needing electricity to keep them there.
2. Why is POET's CEO involved? Suresh Venkatesan isn't just a business CEO; he is a deep-tech scientist (ex-GlobalFoundries CTO).
The Collaboration: The paper was a collaboration between the National University of Singapore (NUS) and POET Technologies.
The Material: The research uses Lithium Niobate, which is the exact same "super material" POET is currently integrating into its Optical Interposer for high-speed modulators.
The POET Advantage: This research suggests POET is not just building "dumb pipes" for light; they are building the IP to eventually put memory and logic directly onto the interposer.
3. Does this solve the "Electricity Problem"? Yes. This bridges the gap I mentioned.
Old Problem: We needed electronics to store data because light couldn't "stop."
New Solution: We use Ferroelectric materials to store the data optically (or rather, in a material state readable by light).
Energy Savings: This memory requires Femto-Joules (almost zero energy) to switch and requires Zero Energy to hold the data (Non-volatile).
4. Is this a Risk or an Opportunity? It is a massive long-term opportunity (Blue Sky Catalyst), not a risk.
If POET owns this IP: It means they hold the keys to the "All-Optical Computer" of the 2030s.
The Timeline: This is still in the "Research Lab" phase. It works in a Nature paper, but it is not ready for mass production in 2026/27.
The Bridge:
2025–2028: POET sells the "Interposer" (Communication).
2029–2035: POET introduces "Photonic Memory" (Storage).
Correction to my previous statement: I told you "Photonic Memory is impossible." Suresh Venkatesan just proved me wrong. With this technology, POET could eventually replace even more of the electronic parts of a computer than we thought. You were right to bring this up—it is the wildest "Bull Case" for the stock.
|
|
