Anatoly Adamov

A Seismic Shift - Photonics Is Becoming AI Infrastructure For years, AI progress has been framed around compute: GPUs, accelerators, and software. But something is fundamentally changing. As AI clusters scale to hundreds of thousands—and soon millions—of processors, data movement is becoming the real bottleneck.  Electrical interconnects are hitting power and bandwidth limits. The next leap in AI infrastructure will rely on optical connectivity and co-packaged optics (CPO). @NVIDIA’s recent multi-billion-dollar commitments to photonics suppliers @like Coherent and @Lumentum has sent a strong market signal. This isn’t just a supply chain commitment. It confirms that photonics is now strategic AI infrastructure. Large hyperscale commitments validate four things simultaneously: • Optical interconnects are essential for next-generation AI systems • Demand for photonics will scale dramatically • Supply is constrained • Long-term demand visibility is now clear When hyperscalers secure photonics capacity, a new question emerges: Where will the next generation of photonic manufacturing come from?This is especially important for the United States, where domestic semiconductor and photonics supply chains are increasingly strategic. Many critical optical components rely on compound semiconductor materials such as indium phosphide (InP)—key technologies enabling high-speed optical transmitters and scalable photonic integration. As AI systems scale, the future of compute will depend not only on silicon—but also on optics, materials, and the manufacturing platforms that enable them.  Photonics is no longer a niche component category. It is becoming core AI infrastructure. Nvidia's investment in Coherent and Lumentum is a significant step forward for Photonics in general and specifically InP. Having worked with InP since 1998, I have always known it's utility. As for Nvidia, it's an extremely strategic move! It provides a secure supply chain while limiting access to other buyers of this technology. I also think that it highlights the fact that moving forward, getting III/V components will become challenging, at least in the short term. There will definitely need to be other sources of supply, especially US based in order to meet current and future meet demand.#

17

1 Comment

“What does NVidia’s effective acquisition tell us about the semiconductor industry? Data movement is the problem !” There is a lot to like about Groq’s LPU, but the most interesting thing in my opinion is Groq’s choice to use SRAM instead of DRAM. Why ? It turns out that the limiting factors - both in energy and speed - is moving data from CMOS to HBM. A majority of heat is generated through this data transfer. Even the SRAM is way bigger than DRAM and more expensive, it appears to be worth it because of the higher throughput and lower power consumption. It also turns out that packaging with HBM is very complex too and very difficult to test. So by using SRAM only, there is a good chance that we will see significant packaging reduction costs, even though SRAM is much expensive. Another note - HBM GPUs are still better for training (simply because there is lots of memory), but because adding more parameters to LLMs does not necessarily improve performance, its not really that necessary. What’s really going to make the difference in training in the future is synthetic data generation, which LPUs with SRAM appears to also have the advantage. My prediction is that future training clusters will have large percentage of compute move from doing SGD to synthetic data generation. I’m quite interested in the dataflow SDKs that LPUs are building too. (Note Cerebras has a similar architecture). I think this is going to be a major theme in 2026 and beyond. What does that mean ? I’m long on TSMC and not so bullish on Samsung and Hynix. Also, I think there is going to be a huge push in the future to develop RRAM / MRAM alternative to SRAM. This is going to be VERY interesting. If anyone know any MRAM groups that are building AI accelerators, please put me in touch ! Groq NVIDIA #LPU #GPU

25

2 Comments

On-device AI becomes powerful only when it works reliably on real hardware. In this Melange Walkthrough, Yeonseok Kim and Sun-Q Kim break down how the full on-device pipeline works, from model selection and quantization to real-device benchmarking and SDK deployment. Select. Benchmark. Deploy. This is how production-ready on-device AI should operate. Watch the full walkthrough below. #OnDeviceAI #EdgeAI #MobileAI #AIInfrastructure #ZETIC #Melange

26

The bandwidth requirements of studio-scale creativity are rapidly outpacing traditional connectivity. As 8K+ resolutions and AI-driven processes intensify, the move to Thunderbolt 5 represents a significant shift for centralized SAN and NAS environments. The doubled throughput offers a path toward real-time collaborative editing and faster data ingest that was previously more constrained. Integrating Thunderbolt 5-equipped desktops and portable workstations into existing Fibre Channel or Ethernet networks—via ATTO ThunderLink adapters—allows for a more unified approach to high-performance compute. What this looks like in practice: · Avid, Adobe, and Autodesk workflows gain lower-latency access to enterprise-grade storage. · Portable workstations achieve performance parity with traditional desktop systems in hybrid environments. · High-speed data movement scales alongside rising file sizes and metadata demands. This level of connectivity is quickly becoming the standard for maintaining efficient, high-reliability studio infrastructure. https://lnkd.in/gGdNu6hj #ITInfrastructure #StudioCompute #Thunderbolt5 #EnterpriseStorage #PostProduction #fibrechannel #ethernet #networking #mediaandentertainment

27

AI’s best-kept secret: the smartest models are hopelessly hardware-obsessed. 🔹 Speed → attention kernels hand-tuned for specific silicon 🔹 Moat → vertical stacks (TPU, Dojo, H100) that outsiders can’t replicate 🔹 Risk → betting on generic GPUs while rivals optimize end-to-end I break down why “tight coupling” is reshuffling the AI power map—and how open-source can close the gap. Read the full analysis ⤵️⤵️ https://lnkd.in/dTbrvh68

1

👓 The AR Hardware Stack: Atomic Bonding, Metasurface Coatings, and Geometric Perfection The "theory" phase of AR is over. The latest flood of patent filings reveals that the industry has moved decisively into the "hard manufacturing" phase. We are no longer just designing optical paths; we are defining how to fuse glass at the atomic level, how to measure the RI of a hidden prism facet, and how to coat a metasurface to kill the "glowing eye" effect. This week’s deep dive covers the full spectrum of hardware innovation from Meta, Magic Leap, Lumus, Corning, and Sunny Omnilight. Here is the technical breakdown: Lumus Ltd. (Geometric Waveguide Suite) 1. Metrology for Internal Facets: You cannot manufacture what you cannot measure. Lumus reveals a slit-based metrology technique to isolate and inspect individual semi-reflective facets inside the waveguide. 2. Compact Folded Optics: A 2D expansion architecture that uses a partial plane reflector to "fold" the optical path within the waveguide. 3. Interference-Free Coupling: A specific geometric optimization for the input coupling prism. By tuning the length of the partial reflectors in the beam splitter, this design prevents light rays from striking the same surface multiple times in unwanted paths, effectively suppressing interference patterns (banding). Meta (Atomic-Level Assembly) Patent: Fusion Bonded Glass Substrates The Innovation: Adhesives are the enemy of high-performance optics - they introduce haze, thermal drift, and refractive index mismatches. Meta has developed a method for Fusion Bonding waveguide substrates directly to each other without glue. Magic Leap (Artifact Control) 1. Invisible Metasurfaces: Metasurfaces are notoriously reflective, causing the socially awkward "eye glow." Magic Leap is attacking this by applying specific AR coatings directly to the grating or PBOE. The patent claims this reduces incident reflection by over 50. 2. Homogenized Light Engines: A new illumination system using Compound Parabolic Concentrators (CPC) with varying cross-sectional shapes to collect and homogenize light from multiple emitters, ensuring a spatially uniform image without "hot spots." 𝗠𝗮𝘁𝗲𝗿𝗶𝗮𝗹 𝗦𝗰𝗶𝗲𝗻𝗰𝗲 & 𝗧𝘂𝗻𝗶𝗻𝗴 Corning Incorporated: As waveguides get thinner, light leakage is a killer. Corning has derived a precise formula for matching the refractive index of a high-index lightguide with its support substrate based on the projector's pupil diameter and critical angle to ensure total internal reflection. Sunny Optical Technology (Group) Co. Ltd.: A geometric optimization of partial mirrors based on the Aspect Ratio (e.g., 16:9) of the optical engine, ensuring maximum coupling efficiency for specific display formats. 𝗠𝘆 𝗧𝗮𝗸𝗲: The battle is shifting from "which architecture is best" to "who can build it at scale." #AR #VR #SmartGlasses #Optics #AugmentedReality #IntelectuallProperty #Patents 👇 Links to all patent applications in the 1st comment.

46

3 Comments

Just wrapped up day 1 at SPIE Photonics West. The next 3 to 5 years should finally open some valuable every day life use cases: 1. AI models that can train 10x faster without overheating the planet, 2. Self-driving cars that see the world with near-perfect accuracy, even in fog or at night, 3. Unhackable quantum communication networks that enable truly private global connectivity. When we started OVNI Capital three years ago, all these used to be cool lab demos. But now it feels like the integrate in real world applications is imminent. As a Paris/SF-based fund, we’re right in the middle of it. We’ve backed a few companies leading innovation in the space such as NcodiN (photonics interposer) or enabling this revolution like Isentroniq (cabling for quantum computers). Who else is around? #PhotonicsWest #DeepTech #FutureTech

45

1 Comment

Memory is the New AI Bottleneck. That message came through loud and clear on the CES main stage, where NVIDIA's Jensen Huang, put it directly: “Context memory no longer fits in HBM.” NVIDIA’s answer is Dynamo, adding multiple terabytes of memory per GPU to handle AI memory overflow in data centers. But CES 2026 also revealed a growing gap on the other side of the market. This year, Intel and AMD announced more than 300 AI PC designs, most shipping with 16–32GB of memory, and even expandable systems typically capped at 64–96GB. Meanwhile, AI workloads are exploding: ✅Reasoning models generating ~5X more tokens ✅Long-running AI agents operating for hours or days ✅Vision and video inputs demanding 100–1000X more memory than text Vendors are moving AI to local devices for privacy, latency, cost, and control. But when memory fills up, falling back to the cloud undermines all of that. This is the challenge Phison aiDAPTIV+ was built to solve. Phison aiDAPTIV+ extends AI working memory, model weights, context, KV cache, and long-running agent state, so AI can continue running when DRAM hits its limits. It addresses the same fundamental problem NVIDIA is solving in data centers, but for AI PCs, workstations, and edge systems where massive DRAM expansion isn’t practical. My takeaway from CES 2026 is clear: Compute scales. Models scale. Memory must scale too, everywhere AI runs. Phison Electronics USA #Phison #aiDAPTIV #EdgeAI #MemoryExtension

74

7 Comments

After years of leading engineering teams and developing new-to-the-world products in optics, display, and instrumentation, I’ve learned that innovation thrives where collaboration lives. Magic Leap has shown how bringing together diverse expertise can turn complex challenges into real progress. This new direction of becoming an AR accelerator for technology partners is a powerful way to build on that experience. By combining our deep technical knowledge with the strengths of our partners, we can move the AR ecosystem forward faster and with greater impact. The future of augmented reality isn’t built in isolation. It’s built together. This article reflects that evolution and the shared vision driving it forward.

132

1 Comment

https://lnkd.in/eUfEDYeG Why bother with instructions, when you're executing the graph itself? That's what NextSilicon are doing, with the promise of enabling speedups on the most complex software run at scale. I've seen the chip, and it's pretty cool! At Supercomputing, I sat down with their VP of Architecture, Ilan Tayari, to find out more on how this chip works, and how it scales.

27

1 Comment

NVIDIA has realized that space inference chips will be by far their largest market in the not-too-distant future 🚀 Jack Kuhr

305

22 Comments