KawzInvests 🦑

@BULLOFBRITAIN @michaelsikand @nvidia

@KawzInvests @michaelsikand @nvidia My quant

NVIDIA GTC is wrapping up. Starting today, @michaelsikand and I are going into full lockdown every hour we have goes into turning what I witnessed this week into the most thorough AI infrastructure breakdown we have published. @nvidia extended an invitation this year. What I saw on-site confirmed several theses I have been building for months and introduced structural setups I had not fully modeled. The report covers all of it: silicon photonics and Co-Packaged Optics, HBM memory demand, inference compute scaling, network scale-up and scale-across architecture, continental and orbital data center buildouts. Every trade thesis. Every structural setup. Where the alpha is heading. Follow and enable notifications. Asymmetrical Bets Substack Report drops soon. $NVDA $AAOI $COHR $MU $LITE $CIEN $SNDK $TSEM

@KarelCapital @michaelsikand @nvidia

@KawzInvests @michaelsikand @nvidia wooow excited for this one

@wolfgangkasper @michaelsikand @nvidia

@KawzInvests @michaelsikand @nvidia BREAK IT DOWN BABY

@michaelsikand They don't want us to win 💯

Institutions have placed a hit on me and @KawzInvests. Hoping we can get this out to retail ASAP 🙏

@gregory_FTA @michaelsikand @nvidia

@KawzInvests @michaelsikand @nvidia Looking forward to it.

@KawzInvests Wow. Thank you for this alpha from GTC!!

Building a memory fab takes 4 years. Building a photonics fab takes 9 months. There is a MASSIVE difference between the build out for Photonics vs Memory A memory fab is a precision lithography operation. You are packing billions of transistors at single-digit nanometer nodes. EUV tools alone take 12-18 months to procure and calibrate. The yield ramp after that takes years. The bottleneck is physics and it cannot be compressed. A photonics fab is an INTEGRATION PROBLEM. You are building devices that manipulate light, not electrons. Indium phosphide. Optical waveguides. Alignment tolerances measured in nanometers of coupling efficiency, not transistor density. No EUV required. The practical timeline difference: Samsung Electronics, $MU, SK Hynix 3 to 5 years from groundbreak to meaningful output. The lithography learning curve is non-negotiable. $AAOI with an existing warehouse 9 months. Not because construction is faster. Because they skip the construction problem entirely. Cleanroom retrofit, tool installation, and process bring-up run in parallel. Most companies do these sequentially. AOI does not. $AAOI has a massive automation advantage AOI runs internal testing systems at 20x the throughput of standard industry equipment. Their product platforms are standardized to the point where each new production line is not a new engineering problem it is a deployment. When they enter a new facility, they are not figuring out the process. They are executing a template they have already optimized across years of production in Taiwan. That is exactly what is happening with their new Texas facility. AOI is not building something new. They are replicating the same factory format, tooling layout, automation systems, and process templates that are already running and yielding in Taiwan. The institutional knowledge, the yield data, the calibration baselines all of it transfers. A semiconductor company standing up a new node from scratch has none of that. AOI walks in with the answer key. Vertical integration across lasers, PCBA, and final assembly means there is no external dependency introducing variance into yield. They own the entire feedback loop from wafer to finished transceiver. That matters because of what the real bottleneck actually is. Most people stop the analysis at fab timelines or InP supply. Both are real constraints. Neither is the hardest part. The hardest part is thermal qualification. A transceiver operating inside a hyperscaler switch runs continuously. These switches need to operate at full load 24 hours a day for the unit economics to justify the infrastructure spend. If the switch is down, the compute behind it is idle. At the scale hyperscalers operate, idle compute is not an inconvenience it is a direct hit to the return on billions of dollars of capex. The failure mode that defines vendor selection is thermal. Transceivers generate heat. Heat degrades the laser. A degraded laser causes signal loss. Signal loss in a switch port takes that segment of the switching fabric offline. Hyperscalers do not tolerate partial switch failures they replace the vendor. This is why qualification cycles are the longest stage of the entire ramp, not manufacturing. Hyperscalers test interoperability, sustained thermal performance, and reliability under continuous full load before committing volume. A vendor that cannot demonstrate 24/7 thermal stability does not get the contract regardless of how fast they built the factory. AOI's vertical integration is a direct solution to this problem. Because they control lasers, PCBA, and assembly in-house, they control the thermal envelope of the finished product end to end. Competitors are integrating components from separate vendors and discovering thermal variance late in qualification. AOI is designing the thermal system, not assembling one from parts. Their automated testing infrastructure means thermal issues surface during production, not during the customer's qualification cycle. That compresses the single longest stage in the entire ramp. And because the Texas facility is a copy of Taiwan, that thermal system arrives pre-validated. They are not learning how to build a thermally stable transceiver in Texas. They already know. They are just doing it closer to the customer. Memory Manufacuturing bottleneck = lithography Photonics Manufacuturing bottleneck = thermal qualification The structural thesis is sound. But there is always a layer of entropy no model accounts for. Execution risk does not disappear because the framework is good.

so… who’s buying the $MU dip? best quarter in terms of the magnitude of beats since those early 2023 $NVDA quarters… stock was $250 in December…maybe it just needs to consolidate even with the explosive numbers? or is the cyclicality narrative just that strong?

@_JoseNajarro What a legend. It was awesome meeting you at GTC!

$NVDA I GOT TO ASK NVIDIA CEO, JENSEN HUANG, WHAT PRODUCT THE MARKET IS COMPLETELY UNAPPRECIATING, AND THIS WAS HIS ANSWER.

@Midnight_Captl @nvidia @stpitts @GroqInc Great interview, really insightful. It was a pleasure meeting you at GTC

Dion Harris (Sr. Director of HPC @nvidia), and @stpitts, (Head of Product @GroqInc) sat down with me yesterday to talk all things Nvidia + Groq. I learned a lot about how the team is dealing with the relentless asks to deliver on their insane roadmap. 00:00 — Dion & Stuart introduce themselves 01:15 — Nvidia's plan for Groq; did it shift? 03:29 — Why are these systems so complex? 06:00 — Nvidia is a force multiplier in the ecosystem 06:38 — DSX *announced at GTC 2026* - why it's under appreciated 07:46 — How has it felt to go thru the ramp of platform complexity from A100>H100>GB>VR 10:30 — A lot of people at Nvidia don't have to work 11:40 — What should $NVDA shareholders know about @JonathanRoss321? Hope you enjoy!

@joinautopilot @michaelsikand What a legend

Breaking: @michaelsikand 's Photonics Portfolio is now up 24% in the past month After investing in Applied Optoelectronics before it's 90% run $800K is currently Autopiloting the portfolio but here's the 5 stocks he owns & why: 1. Applied Optoelectronics $AAOI - 25% "AAOI is one of the few vertically integrated transceiver makers with in-house laser fabrication. As the smallest market cap of its peer group by a wide margin, $AAOI’s massive 2026 guidance gives it major asymmetry." 2. Coherent $COHR - 20% "NVIDIA’s $2B investment validated Coherent as a cornerstone of next-gen AI networking, with leadership in 800G/1.6T transceivers and a vertically integrated supply chain spanning substrates to modules." 3. Lumentum $LITE - 20% "Another $2B NVIDIA-backed name, Lumentum brings best-in-class indium phosphide laser technology and is a critical supplier as co-packaged optics and 1.6T deployments accelerate through 2026 and beyond." 4. Three month treasury $SGOV - 20% "This cash-equivalent position gives flexibility to add to high-conviction names in a choppy macro." 5. Ciena Corp $CIEN - 15% "Ciena dominates scale across optical networking infrastructure. As hyperscalers build out massive AI clusters across multiple facilities, Ciena’s coherent optical networking equipment is the backbone that links data centers together."

@lisasufanclub Appreciate you sharing this, sent you a DM

if anyone wants Lumentum $LITE management talk at OFC from earlier today, DM me where to email it and I'll send it's amazing: – sold out through 2027 – increasing InP output by 50% (up from 40%). have industry's largest InP capacity ❗️ – demand still growing faster (25% to 30%) than supply even after expanding it by 50% – Lumentum is "one of the very few companies" that's showing 400G per lane optical transceivers❗️ – Lumentum is "in the lead pack" right now delivering 1.6T – Lumentum, which doesn't really want to do transceivers cause it's lower margins compared to other products, is actually able to raise margins for transceivers now❗️ – 1.6T shipments with improved margins starting this summer – yesterday Lumentum closed a new multi-year, multi-billion-dollar agreement for OCS ‼️ – UHP lasers (for CPO scale out) in San Jose fab ramping well and already shipping – earlier today Lumentum completed the acquisition of its fifth InP fab (a Qorvo fab in North Carolina). expect to be shipping out of this fifth fab by 2028 ‼️ – currently the only one supplier CPO scale out to Nvidia – shipping CPO scale up (which is 10x bigger than CPO scale out) starting second half of 2027. this is "pure incremental upside" since it's taking it from copper and much more

@Sam_Badawi @retail_mourinho 😂😂😂

@KawzInvests @retail_mourinho

The baggie mobile 🤣 $NBIS

@TLEROY5 😳

@KawzInvests We are soooo early bro. Genuinely can 10x

Applied Optoelectronics $AAOI released their capacity ramp plan. We are in Q2 2026, which means this timeline has already started. Current monthly revenue: $64 million. By Q4 2026, guidance puts that at $303 million per month. By Q4 2027, $701 million per month. These are monthly figures, not annual. Here is why those numbers are achievable. Transceiver pricing does not compress as speeds scale. It expands. 100G sells at $20 per unit. 400G at $64. 800G at $320. 1.6T at $640. Each speed generation commands a price that far outpaces the throughput increase, because the engineering complexity at each step grows significantly. The revenue mix tells the real story. Today, 800G contributes $44 million per month and 1.6T contributes $6.4 million per month. By Q4 2026, 800G reaches $134 million and 1.6T reaches $147 million per month. The ELSFP module, which is not in the revenue mix today, scales from $3.2 million per month in Q4 2026 to $256 million per month by Q4 2027. This is ASP mix shift, not volume growth. Each product cycle that ships carries materially more revenue per unit than the one before it. The execution risk is the manufacturing facility ramping on schedule. Even at 70% of guidance, the revenue trajectory is not priced into the current valuation $AAOI $COHR $LITE

@JonahLupton 😉

@KawzInvests 😳😳

Covered the full $AAOI thesis in depth on Substack: open.substack.com/pub/asymmetric…

@itsbrucegainz @retail_mourinho 💯

@KawzInvests @retail_mourinho I’m jumping in the front seat