Intro music. On March 2nd, 2026, over $7 billion quietly vanished into one narrow, highly specific corner of the tech supply chain. No flashy AI startup. No new LLM. Just glass, light, and, uh... Welcome to your custom deep dive. Today, we are exploring a fundamental infrastructure shift happening right now beneath the surface of the AI boom. We're looking at a definitive March 2026 guide. detailing the massive pivot
away from traditional copper wiring. Right. First, we'll unpack the physical limits of what engineers are calling the copper wall. And next, we'll explore the mind bending, elegant solution of silicon photonics. Finally, we'll break down the three layer framework of companies actually building this new light speed ecosystem. So let's dig in. The best place to start is tracking that massive capital flight because, you know, money of that magnitude does not move by accident.
Absolutely not. In early March, NVIDIA quietly took $4 billion and split it evenly between two relatively unknown hardware companies, Coherent and Lumentum. $2 billion each. Wow. And this wasn't speculative venture capital. These were backed by massive multi -year purchase commitments. Yeah, and Marvell Technology made similar aggressive moves at the exact same time. They spent nearly $3 .8 billion acquiring optical companies. I mean, they closed a $3 .25 billion deal just
for a single startup called Celestial AI. And then threw in another $540 million deal for a networking company. Right. When you see the most dominant players in artificial intelligence aggressively locking up the supply chain for a specific type of hardware, it points to a much deeper shift. It really does. This capital flight marks the definitive end of the copper era in high performance computing. The industry hasn't hit a software bottleneck. It has hit a hard physical limit
of nature. The copper wall. Yeah. Let's unpack the mechanics of this wall because this is where the physics get incredibly fascinating. Let's do it. The bottleneck holding back the next generation of AI is not actually computing power anymore. We know how to make faster chips. The problem is how fast those chips can talk to each other. Every modern AI data center is essentially a
vast, sprawling city of silicon. You have tens of thousands of GTUs constantly exchanging data every single second to train these massive models. And right now, the speed of that data transfer is pushing toward a critical threshold, 1 .6 terabits per second, or TBPS. Okay. At that specific speed, traditional copper cables literally begin to break down. And they don't fail gracefully, right? Or in just one way? No, not at all. They fail in three distinct ways all at once. Engineers
call it the three -threat failure. heat, power, and signal quality. I was thinking about this earlier. It's like trying to force a high -pressure fire hose of data through a garden hose made of copper. Exactly. But worse, because of electrical resistance, the water in this hose isn't just backing up. It is boiling. Yeah, the friction of the electrons moving at 1 .6 terabits physically threatens to melt the hardware. And the power demands explode non -linearly. The signal just
degrades into pure static noise. This is not a software bug we can patch with a clever algorithm. It is a hard limit of physics. It all comes down to how electrons travel through a conductor. At lower speeds, copper is fantastic. It's cheap. It's malleable. It's reliable. Sure. But as you push data at high frequencies, you encounter phenomena like... The skin effect. The skin effect. Yeah. The electrons actually get pushed to the outer edges of the copper wire, which minimizes
the usable area for the current. Oh, I see. That drastically jacks up the electrical resistance. And that resistance generates massive, unmanageable heat. Which brings up an obvious question. Data centers already consume gargantuan amounts of electricity. Oh, massive amounts. Some of them draw over 100 megawatts, which is enough to power a small city. So could this heat issue simply be solved by... building radically better liquid cooling systems instead of completely abandoning
copper. I mean, it sounds like a logical brute force fix. Yeah. But the power usage at 1 .6 TBP is spikes exponentially, not linearly. Wait, exponentially? Yeah. If you double the speed, you don't just double the heat, it multiplies. Pumping more extreme cooling into the system requires even more power to run the chillers and pumps. Right. Eventually, the power required just to cool the copper exceeds the power used to run the AI calculations. It becomes physically
and economically unviable. So we literally cannot out -cool the fundamental laws of thermodynamics. We really can't. And even if you could somehow keep it cold, you still face the third threat, signal quality. Right. At 1 .6 TBps, electrical signals degrade over incredibly short distances. A copper connection that is just a few inches long starts losing clarity. The ones and zeros
blur together into noise. Exactly. To fix that, the network has to constantly run error correction algorithms and resend dropped packets of data. Which creates latency. Which entirely defeats the purpose of buying faster, more expensive hardware. Right. So since physics is breaking copper, we clearly need a new medium. We have to bypass electrical resistance altogether. And the elegant, almost science fiction solution
to this is light. We are taking the electrical signals and replacing them with beams of light. We're swapping out heavy friction bound electrons for weightless. frictionless photons. Yeah, that's the core concept behind silicon photonics. I have to admit, I still struggle to wrap my head around light replacing electricity at a microscopic packaging level. It's wild, isn't it? It really
is. I understand fiber optics running under the ocean to connect continents, but shrinking that concept down to a microscopic scale inside a single computer chip feels wild. It is a profound feat of engineering. Instead of using tiny copper wires to connect the processors, the History is transitioning to wave guides. Let's define that quickly for the listener. Wave guides, tiny glass pathways etched into silicon chips that steer light beams. Yeah. That is exactly it.
They are microscopic glass channels. Yeah. And because light behaves completely differently than electricity, it doesn't suffer from that electrical resistance we just talked about. Which completely eliminates the massive heat generation problem. Right. Photons don't cause friction the way electrons do. It also requires significantly less energy to travel over the same physical
distance. How much less? By moving to silicon photonics, you get roughly 3 to 3 .5 times better power efficiency per individual interconnect link. Just to visualize the scale of that, a standard modern AI data center has somewhere between 40 ,000 and 50 ,000 interconnect links. Whoa! Imagine 50 ,000 interconnects in a single data center suddenly running three times more efficiently. It completely reshapes the grid.
You're talking about a structural transformation of how computing centers are built from the ground up. Absolutely. The market projections are already pricing in this massive structural shift. Yeah. In 2025, the silicon photonics market was sitting at roughly $2 .65 billion. But because of this copper wall, it's projected to reach $9 .65 billion by 2030. Growing at nearly 30 % annually. And some of the more aggressive forecasts even push it toward $28 billion by 2034. It is an absolute
tidal wave of capital. But replacing every copper wire with a glass waveguide is a colossal undertaking. Which raises a massive logistical question. If we are completely abandoning copper for light, does using light mean we have to invent entirely new manufacturing processes and build new factories from scratch? And that is the true underlying genius of this technological shift. We don't
need to reinvent the wheel. We don't. No. The genius is using the exact same silicon chip manufacturing processes that the semiconductor industry has ruthlessly perfected over the last 40 years. Oh, wow. Yeah, we're just repurposing those billion dollar machines. We're hacking silicon factories to print glass light paths instead of copper
wires. Sponsor. That is the magic of it. It allows this incredibly advanced transition to scale up rapidly without needing to spend trillions of dollars building a whole new type of industrial supply chain. OK, now that we understand the core technology and why the physics of copper are forcing this shift, we need a map. Right. We have to navigate this emerging market intelligently rather than just throwing darts at tech stocks.
Exactly. The sources outline a fantastic mental model for this called the three -layer framework. This framework helps us categorize the ecosystem. We start at the foundation, which is the base layer. These are the mega caps. The trillion dollar giants dictating the architecture of the shift. For them, photonics is just one necessary component of a much larger supercomputing business. First up in that base layer is NVIDIA. They aren't
just a chip designer anymore, right? No, they're the primary architect of this entire transition. NVIDIA sells entire rack -scale supercomputing systems now. It's not just a GPU you plug into a motherboard. Right. Their new Vera Rubin platform, which starts shipping in the second half of 2026, is a perfect example. It packs 72 massive GPUs into a single server rack. 72? That's insane. It holds 1 .3 million individual components per system. If you tried to wire that entirely with
copper, it would incinerate itself. So Vera Rubin integrates silicon photonics natively. Exactly. To support that, their new Spectrum 6 switches are delivering five times better power efficiency than traditional networking hardware. And their SuperNex are running at exactly 1 .6 TBPS. That speed is not a coincidence. Nope. It is right at the precise threshold where copper networking breaks down, forcing their customers to adopt the optical standard. Next in the base layer
is Broadcom. And they're taking a slightly different approach. They're the pioneer of what is known as co -packaging. Co -packaging. So Broadcom embeds the optics directly onto the networking chip itself. Right. Just to clarify, usually the optical laser and the processor are separate components, right? Right. And data has to travel across the board to get from one to the other. Exactly. But by co -packaging them, Broadcom drastically shortens the physical data paths.
It cuts power consumption and dramatically improves signal clarity. Because the data barely has to travel before it turns into light. Yeah. Then we have Cisco. They are making the enterprise backbone play. People often think of Cisco as old -school networking. Maybe a bit boring. Right. But that actually puts them in a dominant position. Every enterprise data center, banks, hospitals, logistics hubs will eventually have to replace their aging copper infrastructure with optical
interconnects. And Cisco already owns those relationships. They acquired an optics pioneer called Acacia a few years ago. And now they're manufacturing 800 GBP silicon photonic transceivers to feed that enterprise demand. Finally, we have the controversial wild card of the base layer, Intel. Intel is fascinating here. They actually have 25 years of quiet research and development in photonics. 25 years. They've already shipped
over 8 million photonic chips to date. Wow. Their new optical compute interconnect or OCI chiplet is highly advanced. It runs at four TBPs bidirectionally. And here's the insane metric. It consumes just five picojoules per bit of data transferred. Let's contextualize that because picojoules are hard to grasp. Right. A picojoule is one trillionth of a joule. Traditional pluggable optical modules
consume about 15 picajoules per bit. So Intel's chiplet represents a massive 3x power advantage in an industry where power is the ultimate bottleneck. But Intel has obviously faced real public, corporate, and execution struggles recently. Yeah, they have. Does their current corporate turmoil make their 25 -year head start in photonics irrelevant? Well, the execution risk for Intel is admittedly quite high right now. The market is highly skeptical.
Understandably. But if you look strictly at their assets, Intel is uniquely positioned globally. They are the only major foundry on the planet right now, offering optics -based manufacturing to external customers at that specific scale. If Intel executes their manufacturing advantage, their upside is much bigger than expected. Precisely. And talking about manufacturing perfectly transitions
us up. a level to the middle layer these are the foundries and the manufacturers right because the mega caps like nvidia and broadcom design the grand architecture but they don't actually own the machines that build the chips someone has to physically etch the glass into the silicon This layer is crucial because it reveals the actual physical constraints of the market. And we got a huge signal about this on March 18th, 2026. TSMC, Taiwan's semiconductor manufacturing
company, made a major announcement. Yeah, they confirmed their dedicated COUPN manufacturing line is at absolute full capacity. This line is specifically building 1 .6T optical modules for companies like NVIDIA and Broadcom. TSMC is arguably the most important foundry in the world. Without a doubt. When they dedicate a massive production line to a new technology and immediately sell out of capacity, it proves that photonics is officially an infrastructure standard.
It is no longer an R &D experiment. Exactly. We also have tower semiconductor in this middle layer. They are the number one specialist foundry in the world for this tech. They focus purely on fabricating silicon photonics on silicon wafers. Yeah. They have a direct partnership with NVIDIA, and the demand is so high that 70 % of their new manufacturing capacity is already reserved through 2028. Customers are making massive non -refundable prepayments just to secure a physical
space on the assembly line. Then there is GlobalFoundries. They are playing the scale game. Global Foundries uniquely integrates the optical chip and the full module together. Right. They do the entire fabrication in one single continuous process. That end -to -end capability is incredibly hard to replicate at scale, and it lowers the cost per unit. They're aiming to turn this into a $1 billion annual run rate by 2028. And we absolutely
cannot forget about Fabernae. This might be the most interesting company in the entire document. They are a hidden beneficiary based in Thailand. Fabernae focuses purely on high -precision manufacturing. Specifically, they specialize in submicron alignment. Which sounds technical, but if you think about it, it's wild. You're trying to perfectly align a microscopic laser beam so it fires directly into a microscopic glass pathway. It's like threading
a microscopic needle during an earthquake. Exactly. If the alignment is off by even a fraction of a hair, the light hits the silicon instead of the glass, the data drops, and a $100 ,000 server rack crashes. It is incredibly difficult, and very few companies in the world can do it reliably at scale. Right. Fabernae runs over 2 million square feet of factory space dedicated to this. And here is the kicker. They manufacture these highly sensitive components for both coherent
and momentum. Those are the two competitors that NVIDIA just gave $2 billion each. Wait, if Coherent and Lumentum are technically competing for market share, but Fabrinet is manufacturing the components for both of them, why should an investor or an observer even care about picking the flashy chip designers? Because Fabrinet captures the spending across the entire ecosystem without taking on the design risk. Right. They don't need to pick which specific architectural design wins the
war. The capital flows straight through their clean rooms regardless. They sell the picks and shovels. They win regardless of whose design dominates. That is the perfect way to look at it. And that finally brings us to the top layer. This is the roof of the house, the pure play photonics companies. These are the high risk, high reward specialists that are actually designing and building the specific hardware components
that go into the mega cap architectures. Coherent is the vertically integrated giant in this layer. They build absolutely everything from the raw laser chip up to the final transceiver module. They keep the entire process in house, which gives them immense control over quality and margins. They are the ones who secured that massive $2 billion multi -year deal from Nvidia we mentioned at the start. It was the largest single deal in company history. And yet, even with that growth,
their peg ratio sits around 1 .1 right now. For you listening, a peg ratio compares a company's stock price to its expected earnings growth. A peg around 1 essentially means the stock is fairly valued, maybe slightly undervalued, compared to the hyperinflated AI software stocks. Right. Then we have lumentum. They are the laser supplier to everyone. This is a vital physical constraint we haven't touched on yet. Silicon, the material the chips are made of, is what physicists call
an indirect bandgap material. In plain English, silicon cannot generate light on its own. Right. Every single silicon photonics chip, no matter how advanced, needs an external laser source to provide the actual light. So even Lumentum's direct competitors often have to buy Lumentum's lasers to power their own modules. No matter whose broader design wins out, Lumentum still gets paid for the light source. Exactly. They just lock in a massive NVIDIA supply commitment
as well. And they are currently trading at a very attractive peg ratio of 0 .61. Meaning the market is severely underpricing their future growth. Yeah. Finally. We arrive at Marvell technology. They are the data interpreter and lately an incredibly aggressive serial acquirer. Marvell holds an estimated 50 % market share in digital signal processing or DSP. DSP chips are the unsung heroes
of this entire revolution. Right, because when a beam of light speeds through the glass waveguide and exits the photonic chip, it is still just an analog light wave. It must be rapidly translated back into digital data. The ones and zeros so the actual GPU can process it. But interpreting data isn't enough for them anymore. They just spent $3 .25 billion to acquire a company called Celestial AI. You have to look at Celestial's specific performance specs. Celestial is pioneering
something called a photonic fabric. A photonic fabric. Yeah. They use light not just to connect servers across a room, but to connect computing chips directly to memory banks at a microscopic level. Wow. Because they bypass electrical traces entirely. They boast 10x lower latency. And they claim an astonishing 25x greater bandwidth compared to current industry approaches. If those specs hold, it fundamentally changes how data centers are built. It changes the entire physical architecture
of modern computing. We are moving from a wind constrained by the friction of copper. to a world operating at the speed of light. Let's step back and look at the whole board here. The copper wall is not just an academic theory or a distant warning. No, it's not. It is a physical, hard engineering constraint that slammed into the AI industry the moment we tried to push 1 .6 terabits per second. Yeah, and the market reacted with violent efficiency. Over $7 billion moved
in a matter of weeks. The biggest players in the world, Nvidia, Marvell, Broadcom, are frantically locking in their optical supply chains. TSMC selling out their dedicated production lines proves that this isn't a fad. Silicon photonics is officially the new inevitable foundation of
all future AI infrastructure. Absolutely. For you listening, try applying this three -layer framework base, middle and top, to the next major technological shift you encounter, whether it's quantum computing or new battery chemistries. It is a fantastic mental model for calibrating risk. It cuts through the hype and truly helps you understand who actually builds the complex physical supply chains beneath the software.
And here is a final thought to chew on. The physical limits of copper forced us to entirely rebuild our infrastructure from the ground up. We literally had to pivot to the physics of light just to keep the AI boom alive. If you are wondering why chat GPT or Claude might suddenly get 10 times faster and cheaper in a few years, it won't be just a software update. No, it will be because
of microscopic glass. So what other invisible physical limits of our current hardware are secretly holding back the next massive leap in artificial intelligence? Two sec silence. That's a great question. Something to think about. Thanks for taking this deep dive with us today. Keep learning, keep questioning, and we'll catch you next time. Out to your own music.