차세대 AI 병목 현상은 칩이 아니라... 빛이다

The Next AI Bottleneck Isn’t Chips... It’s Light THE DAILY --- H.E.A.T. Torching Wall Street’s Obsolete Playbook. Delivered straight to your inbox 5 days a week. Subscribe Home Posts The Next AI Bottleneck Isn’t Chips... It’s Light Mar 4, 2026 I’ve been a trader and investor for 44 years. I left Wall Street long ago—once I understood that their obsolete advice is designed to profit them, not you. Today, my firm manages around $4 billion in ETFs, and I don’t answer to anybody. I tell the truth because trying to fool investors doesn’t help them, or me. In Daily --- H.E.A.T., I show you how to Hedge against disaster, find your Edge, exploit Asymmetric opportunities, and ride major Themes before Wall Street catches on. Table of Contents --- H.E.A.T. --- The AI Boom Just Hit the Speed of Light Why photonics is suddenly “the main character” What Nvidia is really doing with these deals --- Winners and losers: the “Photons > Electrons” trade The obvious winners: the companies Nvidia just blessed The “next layer” winners: the photonics stack and AI networking ecosystem The underrated winner: anyone who can deliver reliability at scale Who gets hurt: the losers in a photonics-led AI buildout • 1) “Copper forever” thinking • 2) The “AI is only a GPU story” crowd • 3) The frothiest optics high-flyers if capex hiccups --- The bottom line --- --- News vs. Noise: What’s Moving Markets Today --- Macro regime: “rates up + oil up” is a tax on long-duration growth If this persists I’d stay away from anything related to the consumer. --- A Stock I’m Watching --- In Case You Missed It Talking with Josh Brown about European Digital Sovereignty… --- --- H.E.A.T. The AI Boom Just Hit the Speed of Light For the last three years, the AI story has been told like this: chips are the bottleneck, Nvidia is the toll booth, and the world will keep paying. But the next phase of AI isn’t about how fast you can compute… it’s about how fast you can move the data between all those expensive GPUs without melting the power budget. That’s why Nvidia just did something that should make investors sit up straight: it agreed to invest $2 billion into Coherent and $2 billion into Lumentum, alongside multibillion-dollar purchase commitments and future capacity access rights for advanced laser/optics components. And the agreements are nonexclusive—which is a key tell. Nvidia isn’t “buying a supplier.” It’s doing something far more important: securing scarce photonics capacity for the next generation of AI infrastructure. Why photonics is suddenly “the main character” AI is building “factories,” not “data centers.” When you’re wiring up tens of thousands (or hundreds of thousands) of GPUs, the plumbing becomes destiny. The more compute you add, the more bandwidth you need to keep it fed. Copper works within short distances. But as speed requirements climb and distances stretch across racks and rows, copper runs into physics: signal loss, heat, power draw, and latency. Optics solves that by moving data as light—photons instead of electrons. And the really big step-change is co-packaged optics (CPO): instead of plugging optical modules into a switch like Lego bricks, you push the optical conversion into (or right next to) the switch silicon, cutting power and improving bandwidth density. In plain English: less conversion, less waste, more speed. Here’s the part most people miss: at AI scale, networking optics isn’t a rounding error. Industry commentary around these architectures has highlighted that in a very large “AI factory,” optical networking power can become a meaningful slice of total compute power, and CPO is pursued specifically because it can materially reduce the networking power burden versus traditional pluggable optics. What Nvidia is really doing with these deals Nvidia didn’t just “make an investment.” It effectively announced: “We’re not letting the next bottleneck be someone else’s production schedule.” From Nvidia’s own language, these investments are about accelerating optical interconnects and advanced packaging/integration as foundational tech for the next phase of AI infrastructure—because that’s what unlocks “ultrahigh-bandwidth, energy-efficient connectivity” across AI factories. And note what else is embedded in the announcements: • Capacity access rights: Nvidia wants priority and visibility into future output. • U.S. manufacturing buildout support: the deals explicitly point toward expanding domestic capability, which matters in a world where supply-chain risk is now a board-level issue, not an analyst footnote. • Nonexclusive: Nvidia’s still going to have multiple paths—and it expects everyone else (hyperscalers, other switch vendors) to chase the same scarce parts. This is the “picks-and-shovels” phase of AI infrastructure—except the shovel is a laser, and the pickaxe is a wafer-scale optical supply chain. --- Winners and losers: the “Photons > Electrons” trade The obvious winners: the companies Nvidia just blessed (and de-risked) These are the “capacity + complexity” names—high engineering difficulty, hard to replicate, and now explicitly supported by Nvidia commitments. • Lumentum (LITE) – lasers and photonics components tied directly to scaling AI networking optics. Nvidia investment + purchase commitments + capacity rights. • Coherent (COHR) – same structure: Nvidia investment + purchase commitments + capacity rights for advanced laser products. Why these matter: If the market is shifting toward inference, agents, and distributed AI workloads, the interconnect load doesn’t shrink—it often grows. More “always-on” query traffic means more network pressure. Photonics is how you keep the factory from choking. The “next layer” winners: the photonics stack and AI networking ecosystem These names are the “arms dealers” around the photonics buildout. Not all are pure photonics, but they’re directly levered to the bandwidth/power race. Optical modules/transceivers & manufacturing: • Applied Optoelectronics (AAOI) – high-speed optical transceivers; tends to trade like a call option on datacenter optics demand. • Fabrinet (FN) – manufacturing/assembly leverage to optical modules (the “quiet” way to play the plumbing buildout). Switch + interconnect silicon (the brains that drive the light): • Broadcom (AVGO) – switching silicon and ecosystem exposure; also repeatedly shows up whenever “how do we scale bandwidth” becomes the question. • Marvell (MRVL) – networking silicon/datacenter connectivity exposure. • Credo (CRDO) – high-speed connectivity/signal-chain leverage. Network systems that “drink” optics: • Arista (ANET) / Cisco (CSCO) – not photonics manufacturers, but major beneficiaries of data center network upgrades that require more (and faster) optical ports. The underrated winner: anyone who can deliver reliability at scale Photonics isn’t just “faster.” It has to be serviceable and reliable in monster clusters. One of the big debates is whether CPO adoption happens fast or stays gated by operational concerns (swapability, failure modes, yield). The industry is still early, and timelines matter. So the winners aren’t just “who has optics.” It’s “who has optics that hyperscalers trust at 2 a.m. when something breaks.” Who gets hurt: the losers in a photonics-led AI buildout • 1) “Copper forever” thinking Copper isn’t disappearing. But if you’re positioned for a world where copper scales cleanly into the next bandwidth regime, you’re fighting physics. As lane speeds and overall fabric throughput push higher, optics moves closer to the compute, not farther away. Translation: if a company’s story depends on “we’ll just do more of what we did last cycle,” the market will eventually reprice it. • 2) The “AI is only a GPU story” crowd The market spent years treating the AI buildout like a one-variable equation. But the bigger AI gets, the more the bottlenecks multiply: power, cooling, packaging, memory, networking—and now optics explicitly. Nvidia’s move is a signal that interconnect is now strategic. When the platform owner starts locking up lasers, you don’t want to be the investor still modeling AI like it’s 2023. • 3) The frothiest optics high-flyers if capex hiccups Optics can be both: the most important bottleneck, and the most violent trade when sentiment shifts. If hyperscaler spend wobbles, or if photonics timelines slide, the highest-multiple, most crowded optics names can get hit hardest—because they’ve become the market’s favorite “pure play” expression of the AI network thesis. --- The bottom line Nvidia’s $2B + $2B photonics commitments aren’t a side quest. They’re a flashing sign that the AI race is moving from “who has the best GPU” to “who can build the fastest, most power-efficient AI factory.” We’re watching a new kind of arms race: Compute is expensive. Power is scarce. Bandwidth is the new oxygen. And the companies that control the light—lasers, optical engines, and the manufacturing capacity behind them—just got promoted from “supplier” to “strategic asset.” --- --- News vs. Noise: What’s Moving Markets Today Yesterday highlighted some of the difficulties of trading and investing during a war. Monday we saw the traditional playbook—fade the dip. Yesterday we saw the real move based on fear of a prolonged war and higher oil prices (see below). But, we closed way off the lows on a statement from Trump about teaming up with the Kurds. So far this morning the biggest takeaway is crypto and precious metals up at the same time for a change. Wish there was an ETF for that :)…… --- Macro regime: “rates up + oil up” is a tax on long-duration growth The tape they’re describing is classic risk-off duration pressure: futures down hard, global equities weak, front-end yields ripping higher, and the market repricing to <50 bps of Fed easing. In that regime: • high-multiple software is structurally more fragile (even when fundamentals are fine), • “real-economy” cash flows and AI capex beneficiaries with tangible demand tend to hold up better, • anything dependent on easy financial conditions gets punished. If this persists I’d stay away from anything related to the consumer. Bottom line is we are going to have underlying volatility until this is over, and it’s going to be very hard to make moves based on what you are seeing. --- A Stock I’m Watching UBER: I view Uber as one of the cleaner “AI-in-the-real-economy” compounders because AI isn’t a bolt-on feature here—it’s the operating system that prices, matches, routes, fights fraud, and continuously improves utilization across a two-sided network (Mobility + Delivery). In 2025, the core platform kept scaling while profitability tightened: Gross Bookings grew 19% to $179.9B, revenue grew 18% to $52.0B, Uber exited Q4 with 202M MAPCs (+18% y/y), and Adjusted EBITDA rose to $8.73B with Free Cash Flow up 42% to $9.76B. The “asymmetry” angle is that this is already a sizable cash-flow engine and it has credible upside optionality from autonomy (Uber doesn’t need to own AV hardware to monetize demand aggregation, dispatch, payments, and network density)—but you need to watch whether take-rates stay resilient as the ecosystem evolves. Capital return is also a real support: Uber’s board authorized $27B of repurchases (with $19.2B remaining as of 12/31/25), which gives management meaningful flexibility if the stock is weak while fundamentals hold. --- In Case You Missed It Talking with Josh Brown about European Digital Sovereignty and European Defense…. All rights reserved. Keep Reading View more