2026年十大主题投资：聚焦AI供应链瓶颈、软体机器人及支付颠覆。

英文原文

2026 Newsletter. Thematic Investments: Evolution, Disruption, and Bottlenecks 1. Soft Robotics - Evolution to $TSLA, $ONDS, Boston Dynamics. 2. SiPh - InP Bottleneck | $AXTI, $LITE, $GOOGL 3. Glass Substrates - Bottleneck | $NVDA, $INTC, $TSM 4. Money Movement - Disruption to $V, Stripe, $BOA 5. AI Cloud Layers - Bottleneck | $NBIS, $IREN, $HUT. 6. LLM Cybersecuirty - Evolution to $CRWD, $CSCO, $MSFT 7. LEO Space Infrastructure | Evolution to $RKLB, SpaceX, $ASTS 8. Consumer Agentic Workflows (50 Step) - Disruption to the Consumer Workforce, from Manus, $PATH Cognition 9. Distributed Computing Latency - Bottleneck | $TSLA, $AMZN, $GOOGL, 10. Copper Interconnect Life Extension - Bottleneck | $NVDA (LPU/Groq), $AMD, $INTC _ This is an light overview of thematic investments I find the most interesting from a public-information synthesis perspective + second/third-order effects from bottlenecks! _ 1. Soft Robotics: The Evolution to Robotics Traditional robotics (Optimus, Boston Dynamics) relies on Inverse Kinematics to rigid joints. Soft robotics changes the math. We've met the point where hardware (Optimus, Boston Dynamics, Figure) met LLMs (Gemini, Grok, Opus), and we're at the beginning of possible widespread commercialization. By using materials inspired by octopus tentacles and human skin, robots are moving away from gears and toward fluidity to handle extremely delicate tasks like handling produce like the human hand, to picking up extremely heavy surfaces adding Octopus-like extensions to $ONDS/Andruil Drones. The evolution is thinking outside the box in terms of what robotics can do. I remember working with some Stanford PHds in this field like 7 years ago, and it just so happens AI is starting to be commercialized after many years of research. So expected, this field to be as well. Possibilities are limitless adding organism-like fluidity to rigid robotics, this is just the natural evolution. Most of these are prob private companies. _ 2. Silicon Photonics - Bottleneck of the AI Infrastructure "InP Chokepoint" Blackwell Ultra Clusters to Google TPUs have hit the upper wall and requires photonics for interconnects | OCS to scale up. The Substrates: $AXTI (via Tongmei) and Sumitomo (Japan) control roughly 60-70% of the world's InP substrate market. The Materials: Companies like Vital Materials (China) and AXT control the refining of the raw Indium itself (78%+ of supply chain). If you are a US tech giant, your entire "AI Growth Story" for 2026 depends on materials controlled by geopolitical rivals. The only scalable solution is engineering around it, either by delivering light-on-chip, while using 90% less InP or companies that use tiny slivers of Indium Phosphide instead of large, expensive wafers. There's opportunities with the bottleneck itself like AXT, Sumitomo. Or companies that help address it like $POET. _ 3. Glass Substrates - Fixing the Bottleneck for CPOs from $NVDA to others. The shift toward glass substrates is essentially the semiconductor industry’s answer to a physical wall they are hitting with current materials. Current chips sit on a substrate made of organic materials (essentially specialized plastic). As chips get larger, like Nvidia's massive GPU packages, plastic substrates warps. So, glass substrates is becoming the industry standard for Co-Packaged Optics (CPO) because they solve the single biggest problem in photonics with alignment. US Government already sees this as a necessity and we've seen huge subsidies funneling down to some of these companies. Companies like $INTC, Samsung Electronics, Absolics (SKC Subsidiary), DNP, and others are the main beneficiaries, especially as MRVL and $AVGO (driving glass for optical switches) move forward with CPO revolution. _ 4. Money Movement - The Disruption to Card Networks, Banking, Exchange, and Payments For decades, moving money has been a "toll road" business. Every time you swiped a card, 2% to 3% of that money vanished into the pockets of the Card Networks (Visa/Mastercard) and Issuing Banks. Or buying/selling crypto from an exchange would be .2-1%. It was the most profitable, "un-killable" business model in history. Until now. The "Genius Act" of 2025 just handed companies like $XRP with Money Transmitter Licenses or Banking Charters the keys to the kingdom. Not really theoretical for me. I happen to be working on this myself at my own startup with some folks who created V / $PYPL's real-time payment networks. But basically companies with existing MTLs or pursuing banking charters leveraging the Genius Act and some other tech can now bypass legacy % fees by doing settlement on top of the Federal Reserve and blockchains, effectively converting percentage-based fees into a few cents. Would 99% companies do it? Probably not since every single margin from across the payment industry would just go to 0. I'd be happy though. But basically Bridge's $1.1B acquisition by Stripe should have been a red-alarm to existing companies that days of 1-Day ACH, interchange models, $25 international transfers, are soon to be over. This extends to many other adjacents from low fee disruptions like $HOOD, Mercury all the way to Stablecoin Neobanks, or companies making their own stablecoins like $SOFI. _ 5. AI Cloud Layers - The Solution to HyperScaler compute Bottleneck While Hyperscalers are stuck in 3-5 year grid interconnection queues, miners like WULF and IREN are sitting on plug-ready GWs today This is the opportunity of a lifetime as hyperscaler funnel their cash cow Cloud revenues down to tiny companies. There's many different layers to this from Fluidstack, Poolside, Fireworks on the GPU orchestration layer, to the bare metal layer that companies like IREN are building. Then there's becoming the hyperscaler themselves like NBIS owning the physical locations, the GPU, software orchestration, and then providing simple interfaces for inference. This is the opportunity for a few small companies to become Amazon Web Service or Microsoft Azure over the next year or two, or get acquired (eg. GOOGL buying Intersect for $4.7B) Neoclouds like NBIS, IREN, CRWV, down to colo plays like CIFR, WULF, HUT (and private sectors -> Energy) stand to benefit. _ 6. LLM Cybersecurity - The Evolution to Modern Security and Vulnerability Defense Recent reports (e.g., from Anthropic's Red Team) showed that advanced models like Opus (and future iterations) could autonomously scan open-source smart contracts and identify "Zero-Day" exploits worth millions of dollars in minutes. The Implication: If an AI can find a logic flaw in a immutable Blockchain contract, it can find a flaw in a bank's SWIFT API or a power grid's control software. Same with KYC/AML. Models like Gemini Nano Banana are able to create realistic images/videos of people and people are able to get past a lot of programs. There's tons of things as an unsexy alpha in this field like LLMs automating away SOC2/PCI dss compliance to agents sitting on a server, continuously monitor logs, and auto-generate the evidence needed for auditors. 7. LEO Space Infrastructure | The Evolution to Expanding into the final frontier. Space is the next big thing. This is not anything new. (hope you got the joke). But anywhere from companies like $RKLB, SpaceX. Companies that fix orbital congestion or launch cadence bottlenecks. To companies that commercialize the infrastructure like ASTS or Starlink present many opportunities over the next year. So companies like Impulse, Blue Origin, $ASOZF to RKLB, $ASTS stand to benefit across the entire chain. 8. Consumer Agentic Workflows (50 Step) - Disruption to the Consumer Workforce, from Manus, PATH Cognition This one is simple and needs no explanation. But largely obvious in potential impact on employment + cost saving. How do you automate away business development? How do you automate away marketing? How do you automate away software engineers? This is going past few step ChatGPT answers and directly in to the real world where an AI agent can roam X, find the right people, DM someone, continue conversations, and lead to a sales call in just one workflow. This is the end of the "Chatbot" era and the beginning of the "Action" era replacing everyone previously required in a company. I haven't quite seen this done at scale yet with any company. Public companies like META that own these, don't really present the best exposure. Maybe $PATH for public space. 9. Distributed Computing Latency - Fixing the Bottleneck for AI Compute Capacity Strains Hyperscalers like GOOGL Cloud, MSFT Azure at max capacity. Elon Musk already floated distributed computing as the future of solving this issue (eg. having networks of $TSLA's providing compute for LLMs for inference). The "Tesla Compute Cloud" thesis is fascinating, but the single biggest physical barrier I've identified is: Inference Latency. Too generate "Token B," the model must first finish generating "Token A." It cannot do both at the same time. If you split a massive model (like Grok-3) across 5 different cars to fit it in memory, you have to send data between those cars for every single token generated. So, if your network latency between cars is even 20ms (optimistic for 5G), and you are generating 50 tokens, you just added 1 full second of pure "waiting time" (latency) on top of the compute time. In a data center using NVLink, that wait time is measured in nanoseconds. Same applies to any spare computer, GPU, and others owned by retail users. And there's billions of consumer GPUs (Teslas, iPhones, Gaming PCs) that sit idle 90% of the time. Solving the "distributed latency" problem for inference presents one of the single greatest arbitrage opportunity in the history of computing. Haven't really seen any companies that accomplished this at scale yet. Maybe NVIDIA Dynamo, $AKAM, TSLA, getting a little closer. 10. Copper Interconnect Life Extension - Addressing the Bottlenecks of Nvidia and Others Since we can't have infinite InP, we have to engineer around it with what we have (eg. Copper), so copper cables can do things that physics said it shouldnt like carrying 224G signals across a rack without signal loss. The industry is hitting a hard stop on InP where, US cannot physically cannot mine and refine enough InP to turn every link in a data center into fiber optics. If anything helps, then it's good. EG. NVDA's $20B "Acqui-hire" of Groq's team and IP. LPU is more about inference latency/architecture but it addresses copper life extension as a byproduct. Groq’s entire architecture beat Nvidia on latency because it rejected optics. Groq uses a "deterministic" mesh that relies on direct electrical (copper) connections between chips, avoiding the "jitter" and conversion time of optical switches. Companies like $ALAB, $CRDO, Groq, or anyone who can find ways to engineer around the optical bottleneck with copper will be a winner. _ There are tons of trades from both private sector investments to public! Just wrote up my thoughts on the fly today, but happy to elaborate later. Regardless I believe a lot of these thematic investments from: Investing in InQ Bottleneck Workarounds ( $POET ) or the bottleneck itself ( $AXTI ) to Disruptors ( $CRCL ) in the public sector. To Investing in copper extension bottleneck fixes (Groq), bank charter disruptors (Mercury) to evolutionary companies (Lightmatter, Festo) in the private sector. Present asymmetrical upside in 2026. Happy New Year!

在 X 上查看原推 ↗