We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor. We ask that you disable ad blocking while on Silicon
Investor in the best interests of our community. If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.
New contributor: Thoughts pertaining to Celestial AI
Hello All, I have been silently following this group for a while and have now decided to share my thoughts mostly pertaining to Celestical AI.
1. Ayar labs might not be the most feared competitor, even though Nvidia is an investor in Ayar Labs.
2. LightMatter might be there most prominent competitor to Poet/Celestial AI. Their photonic interposer becomes widely available this summer, which is being postioned for connecting several xpu/gpu together. Several other product launches are also identified below, along with their launch video. Their products are infact quite competitive on cost/scaling/energy savings/etc. However, I do believe poet might offer the cost advantage and Ceslestial AI will offer a bandwith advantage. And once we figure out attaching FAUs to the Blazar, I believe our competitive advantage will no douth fiercly balloon.
Key products and timelines include:
Passage M1000 (Photonic Interposer): This 3D photonic superchip, designed for next-generation XPUs and switches, will be available in the summer of 2025.
- 256 optical fibers per chip - 2.3 pj/bit - 114 tbps (package) - supports GPU up to 1500 Watts
Passage L200 (Co-Packaged Optics - CPO): These 3D CPO chips are expected to accelerate time to market and performance for next-generation XPUs and switches and are slated for release in 2026.
Passage L200X (Co-Packaged Optics - CPO): A higher-bandwidth version of the L200 (64 Tbps), also available in 2026
3. Celestial AI / Poet Revenue impact: I think Meta is going to be the driver for Celestial AI products. This is huge because meta just increased capex by around 56% for 2025/2026/2027. And they infact changed their entire data center buildout design and plan to agressively ramp construction. This tells me that Mark is extremely pleased with the photonics fabric and might have validated its' usage. The samples ordered prior might be used in the data center they plan to bring online this summer, because it going to be the first implementation of Mark's new design. This is great because I think we should definetely see a solid PO for the interposer with attached FAUs sometime in Q3.
4. AMD MI400 - Celestial AI or Poet might be incorported into the design. Consider that AMD is an investor in Celestial AI and by now they would have likely view the product and observe the benefits and perhaps were pushed to have AMD connect with Poet for the underlining technology. Just wishful thinking but we'll see how it plays out in the coming years.
5. I also think Lightmatter introducing their products at OFC 2025 is a good thing. It should make investors look into their technology more and try to find a publicly traded company with similar tech. The cost savings of using a phonic fabric compared to optimal modules for each GPU/ASIC connection is insane, so it will be widely adopted in my eyes.
5. Finally, the growith for 1.6T modules will still continue to increase and if poet partners' customers validate the designs in a timely fashion it will definetely drive POs in Q4 2025 and onwards.
"I did not proof read so there might be a few gramatically errors... forgive me"
Lightmatter unveils M1000 and L200 Passage photonic interconnects Company aiming to tackle bandwidth bottlenecks with new offerings April 02, 2025 By Charlotte Trueman
Lightmatter has unveiled two new photonic interconnect solutions designed to tackle the bandwidth challenges associated with increasingly dense AI deployments. Dubbed the Passage M1000 and the Passage L200, both offerings will integrate with next-generation XPUs and switches. The Lightmatter M1000– Lightmatter Boston-based Lightmatter specializes in photonics-enabled technologies that use light instead of electrical signals for computing, meaning its chips don’t experience the same heat or resistance as traditional chip architectures. The new solutions have been built on the company’s Passage interconnect technology, a chip that takes arrays of traditional processors and links them up using a programmable on-chip optical network.
According to Lightmatter, the M1000 enables a “record-breaking 114Tbps total optical bandwidth for the most demanding AI infrastructure applications,” providing connectivity to thousands of GPUs in a single domain. Consisting of an eight-tile 3D interposer with an integrated programmable waveguide network, the M1000 also contains 3D integrated electrical integrated circuits containing a total of 1024 Electrical SerDes (Serializer/Deserializer - technology which is used to reduce the number of inputs and outputs on an ASIC or FPGA), 56Gbps NRZ modulation, and eight wavelength WDM transmission on waveguides and fibers. The offering also has 256 optical fibers edge attached with 448Gbps bandwidth per fiber and delivers 1.5kW of power in an integrated advanced package. Due to be available from summer 2025, the M1000 will be produced by GlobalFoundries and Amkor and utilize GF Fotonix, a silicon photonics platform that integrates photonic components with CMOS logic into a single die.
“Passage M1000 is a breakthrough achievement in photonics and semiconductor packaging for AI infrastructure,” said Nick Harris, founder and CEO of Lightmatter. “We are delivering a cutting-edge photonics roadmap years ahead of industry projections. Shoreline is no longer a limitation for I/O. This is all made possible by our close co-engineering with leading foundry and assembly partners and our supply chain ecosystem.”
Meanwhile, Lightmatter also announced its upcoming 3D photonic L200 interconnect. Described as the world’s “first 3D co-packaged optics (CPO) product,” the L200 3D CPO product portfolio consists of the 34Tbps L200 and the 64 Tbps L200X, which the company claims will enable more than 200Tbps of total I/O bandwidth per chip package, resulting in up to 8x faster training time for advanced AI models. The solutions will also use a standard interoperable UCIe die-to-die (D2D) interface, which Lightmatter said will facilitate “scalable chiplet-based architectures to seamlessly integrate with next-generation XPUs and switches.”
The Passage L200 integrates with Alphawave Semi’s latest chiplet technology portfolio, with Alphawave offering the IP for the accompanying XPU/Switch die royalty-free. Due to be available in 2026, the L200 will be manufactured in partnership with GlobalFoundries, ASE, Amkor, and advanced node CMOS foundries.
Sorry but I forgot to mention that lightmatter last funding raise valued them at about 4 billion. So investing in Poet at this price is amazing. I see us roughly 2 years out from having a larger market cap compared to them. I know this is hard for those to hear that have been invested for going on 15 years, but times have change. AI is the new gold and technologies that increase bandwidth across communications will continue to be highy indemand.
