$LITE CEO just gave the most direct answer to the “is this a bubble” question that anyone in the optics space has given on camera. Michael Hurlston, five quarters into the job at Lumentum, laid it out plainly on @sourceryy. The board’s own forecast when he joined turned out to be roughly 4x too conservative, revenue has tripled since, and the stock ran up 10-12x, briefly clearing $1,000, before pulling back on broad AI-sentiment fears over the past couple weeks. His actual bubble answer is the interesting part. Even if data center construction volume slowed entirely, he argues, there’s still a structural conversion happening inside existing data centers. Copper physically can’t carry rising bandwidth over distance without heat and resistance killing the signal, so electrical connections keep getting replaced with optical ones regardless of the build cycle. The next leg of that, “optical scale-up,” moves fiber into the rack itself for server-to-server links, a shift he says pushes volumes from the thousands Lumentum used to ship into the hundreds of millions. He also confirmed another bottleneck: indium phosphide, a niche material completely separate from the CMOS everything else runs on, requiring years to build a fab, which is why Lumentum runs its own, calling it “the $TSMC of the optical industry.” $AAOI and $COHR sit in the exact same lane, vertically integrated InP producers benefiting from the identical scale-out-to-scale-up transition, and Coherent got the same $NVDA equity backing Lumentum did. Long opticals🚀

Everyone is crowding into $MU and $SNDK for the memory shortage. But most people are overlooking $PENG. Same tailwind, way less attention. → AI moved from training to inference, and inference is memory-bound, not compute-bound. The GPU's onboard HBM becomes the wall. → DRAM prices doubled in Q1 and Reuters reported they could climb another ~63% this quarter. HSBC's Kazunori Ito sees little prospect of a quick drop. → The crowd buys the chipmakers. $MU, $SNDK, the obvious picks. But the shortage also reprices the system integrators who package that memory into AI factories. That's $PENG. Why $PENG is still a good play: → Memory segment grew +63% YoY in Q2 FY2026 on AI demand and pricing. → The MemoryAI KV Cache server (industry-first, March 2026) adds up to 11 TB of CXL 2.0 memory per node ≈ 10x faster than NVMe caching, 3.8x faster than RDMA. → Built on 30 years of SMART Modular memory IP + an SK Hynix DRAM collaboration competitors can't easily copy. The rerate is already starting: → $16 (52-wk low) → ~$71 recent high. Rosenblatt and Stifel both hiked targets into the $60s, Street high sits at $108. → The market is re-rating $PENG from a commodity hardware multiple to an AI-infrastructure multiple. While everyone fights over $MU, that gap is where the alpha lives. I'm in at $30, holding, and $100 is my minimum target. Not financial advice. DYOR.

Where does $SIVE sit in the AI supply chain? This is the part most people miss. Sivers doesn't sell directly to Nvidia or Amazon. They sit upstream of the companies that sit upstream of those hyperscalers. The chain looks like this: Sivers ships InP CW-DFB laser arrays → to companies like POET, Ayar Labs, O-Net, and Enablence, which package those lasers into External Light Source (ELS) modules → those ELS modules feed optical engines built by Celestial AI (just bought by Marvell), Ayar's TeraPHY chiplet, and Lightmatter's Passage → those optical engines plug into ASICs designed and packaged by Marvell, Broadcom, and Alchip → which become AWS Trainium, MSFT Maia, META MTAI, Google TPU, and Nvidia Rubin clusters Four layers up the chain from the GPU. That's why nobody saw them. And that's exactly why they matter. The laser is the input that nothing else in the stack works without. You can swap optical engine vendors. You can swap packaging vendors. You cannot swap the laser source without a 2–3 year requalification cycle that nobody has time for in a 2027–2028 ramp window. That is what supply chain analysts call a chokepoint. @aleabitoreddit calls it "the kingmaker for CPO." The structural position is real. The only debate is whether Sivers specifically captures the volume, or whether Lumentum and Coherent eat the share from above. More on that later. $NVDA $META $MSFT $AMZN $GOOG $AVGO $LITE $COHR $MRVL $POET

$SIVE / $SIVEF - A 75-year-old Swedish chip company with 130 employees just became one of the most important suppliers in the entire AI infrastructure stack. The AI data center buildout is hitting a physical wall. Copper interconnects between GPUs are running out of bandwidth, dissipating too much heat, and burning too much power. Every hyperscaler roadmap from 2027 onward solves this by moving the optical-to-electrical conversion INTO the chip package itself. This is called Co-Packaged Optics (CPO). CPO needs three things: 1. An optical engine (the modulator chiplet) 2. A packaging platform (the interposer) 3. An external light source feeding the engine That third component is the bottleneck. It needs to be a high-power, multi-wavelength, continuously emitting laser array - specifically built on Indium Phosphide (InP). There are maybe 4 companies in the world that can make these at hyperscaler-grade quality. $SIVE is one of them. They're the only Western microcap pure-play. They ship from a 700m² cleanroom in Glasgow that can run over 2 million laser tests per month, with a second-source qualification underway at WIN Semiconductor in Taiwan to unlock real volume. I am going to give you a 5-post breakdown of where they sit, who's already using them, and why this could be the next $LITE - follow if you are interested. This will be posted through the next 2 days