No Matter How Powerful NVIDIA's GPU Is, Without This It's Useless: The Identity of the Hidden Bottleneck

• The Secret of the Magic That Cut Power by 73%: What Is CPO, and Why Did Optics Stocks Explode?
• The Day Copper Surrendered to Physics, Only This Company Smiled: A Full Breakdown of Lumentum's Earnings
• A Billion-Dollar Market Born Because the Laser Ran Too Hot: The Hidden Story of the ELS Module
• The Real Reason CoreWeave, Supermicro, and Lumentum All Erupted on the Very Same Day 

🔥 Lumentum's Earnings Surprise: Digging Past the Numbers to the Real "Why" Behind AI Optics

Whenever people talk about AI data centers, they almost always think of NVIDIA GPUs first. But no matter how powerful tens of thousands of chips are, if they can't talk to each other, they're just expensive scrap metal. The connection between chips, the interconnect, is the hidden bottleneck of the AI era. On August 12, 2026, the companies that make the core components of this bottleneck all reported explosive results at once, and the market shook once again.

What this article sets out to do is clear. Rather than stopping at "earnings were good," it traces the full chain of cause and effect, showing which product, against the backdrop of which physical laws and industry structures, produced these numbers at precisely this moment. It's written so that newcomers can follow along and experts can still nod in recognition. 📈

💡 First, the Results: Just How Strong Was Lumentum's Quarter

Lumentum (ticker $LITE) posted revenue of $1.01 billion this quarter, comfortably clearing the roughly $984.6 million the market had penciled in. That's up 24.5% from the previous quarter and a striking 109.3% from a year ago. Adjusted earnings per share came in at $3.23, beating the $2.95 estimate, while the non-GAAP gross margin was 50.4% and operating margin 36.6%.

A quick note on what non-GAAP means: it's an adjusted figure that strips out items far removed from the core business, such as one-time expenses and stock compensation, to reveal the company's pure operating strength. The fact that margins thickened even as revenue swelled is a decisive hint that the "high-margin new product" I'll explain shortly is taking up a bigger share.

Even more impressive was the next-quarter guidance, the outlook the company set for itself. Revenue was guided to between $1.225 billion and $1.275 billion, well ahead of the $1.16 billion estimate, with adjusted EPS of $4.05 to $4.35 (versus $3.63 expected) and operating margin of 39.5% to 40.5%. This trajectory, with revenue climbing like a staircase from $808 million to $1.01 billion and on to $1.25 billion, is the backbone of today's story. But why now, and why such a sudden jump? That's where the real story begins.

🔬 Back to the Root: Copper Hit a Physical Wall

The network inside an AI data center is easiest to grasp in three layers. Scale-up bundles many chips within a single rack into one giant brain; scale-out links several rack clusters through a central switch; and scale-across hands signals off to another data center entirely once a single building reaches its power limit.

Until now, the inside of the rack, the scale-up domain, belonged entirely to copper wiring. Signals only had to travel very short distances, so copper was enough. The trouble starts with NVIDIA's next-generation Rubin platform. As the speed demanded inside the rack soars to unimaginable levels like 14.4 Tbit/s, copper slams head-on into the wall of physics. Signals smear, power leaks away, and heat becomes unmanageable. Here's the insight worth holding onto: this isn't a shortage of engineering skill, it's a ceiling drawn by the laws of physics. No amount of time will let copper break through it. So the entire industry has no choice but to shift to light, and this forced transition is the most fundamental force propping up Lumentum's results.

⚡ The First Star, CPO: The Principle Behind Cutting Power by 73%

This is where CPO (Co-Packaged Optics) takes the stage. Let me explain the principle using the "local geography" of a circuit board.

Traditionally, GPUs and switch chips held only the compute components, while the optical transceiver that converts electricity into light sat off to the side at the faceplate on the very front edge of the server rack. The electrical signal the GPU generated had to travel several inches across the board to reach that transceiver. A few inches may not sound like much, but for an ultra-high-speed signal it's a distant journey. Because the signal degrades badly along the way, every transceiver needed a power-hungry chip called a DSP (digital signal processor) to restore the signal to a clean state. A single DSP swallows nearly half of a transceiver's total power. Incidentally, this DSP market is what grew companies like Marvell into tens-of-billions-of-dollars businesses.

CPO solves this by relocation. Instead of leaving the optical engine at the edge, it moves the engine onto the same package substrate as the GPU or switch ASIC, just a few millimeters from the chip. With no reason for the electrical signal to cross the board, the signal no longer degrades, and so the DSP becomes unnecessary.

The effect shows up immediately in the numbers. Where a conventional 800G DSP transceiver draws 16 to 17 watts, a CPO optical engine of the same performance needs only 4 to 5 watts even including the external laser, roughly a 73% reduction. This is physics' answer to "why CPO." On top of that, when Meta and Broadcom actually measured it, CPO switches ran without a single failure for over 15 million hours, showing reliability about five times higher than conventional transceivers. It captures both performance and stability at once.

🌡️ The Second Star, the ELS Module: A Market Born from Heat

This is exactly where Lumentum's true gold vein reveals itself. The "initial order for ELS modules" mentioned in the earnings release is the protagonist here.

It all starts with heat. The laser, the light source, is physically brutally sensitive to heat. As temperature rises, performance sags and lifespan plummets. Yet CPO places the optical engine right next to a Blackwell GPU hot enough to boil water. Put a heat-sensitive laser alongside it and the laser can't survive. So the compromise the industry reached is this: CPO attaches the optical engine but not the laser. Only the signal-modulating components move onto the GPU package, while the laser gets pushed out to an air-cooled box at the front of the rack. That box is the ELS (External Laser Source).

This physical constraint of having to move the laser far away paradoxically gave birth to an entirely new market. Because the laser is now distant, the light has to run an obstacle course of optical fibers, connectors, and splitters to reach the GPU. An ordinary 70 mW laser would fade away on this journey and end up far too dim. To overcome the losses and light up multiple optical engines at once, you need an ultra-high-power (UHP) laser pumping out 300 to 400 mW per diode. This is the reality behind the "rising demand for ultra-high-power CPO lasers" in the earnings release.

🏆 Why Lumentum and Not Someone Else: The Secret of the 50% Margin

Here lies the answer to "why is it Lumentum in particular that's doing so well." Ultra-high-power lasers aren't something just any company can make.

Push a tiny chip up to 400 mW and you generate a fearsome heat density. If the crystal structure is even slightly imperfect, a phenomenon called catastrophic optical damage (COD) occurs, in which the laser melts its own mirror and self-destructs. Preventing this requires flawless crystal growth, that is, epitaxy, and a thick, intricate crystal layer, achievable only in a specialized reactor called MOCVD. It's a manufacturing challenge incomparably harder than a standard laser.

Right now, the company that can mass-produce these ultra-high-power lasers in volume, and meet the brutal reliability standard (GR-468) that AI clusters demand, is essentially Lumentum alone. Its rival Coherent ($COHR) leads on production volume but trails by roughly two years in ultra-high-power laser development and still struggles to stabilize the heat of high-power chips. There's one decisive piece of evidence: Coherent, a laser company itself, still buys the EML (electro-absorption modulated laser) for its own transceivers from Lumentum. If it can't even fully supply standard lasers, the reasoning goes, how could it make the far trickier ultra-high-power ones?

This is precisely where the secret of the 50.4% gross margin hides. A common low-power commercial CW laser costs only $5 to $6 apiece, but an ultra-high-power laser sells for several times that, and an ELS module packing eight of them carries an average selling price of around $1,800 with an estimated gross margin of about 60%. Because each switch sold drags along an order for 18 ELS modules for Lumentum, this multiplier effect pushes revenue up the staircase. Add to that the transition to 1.6T transceivers, which sends 200G EML demand soaring; since a 200G EML occupies the same wafer area as a 100G but sells for twice the price, this shift in product mix piles up directly as margin. The earlier riddle of "revenue rising while margins also climb" is solved right here.

🧩 So This Is How the Earnings Commentary Reads

Now the sentences in the earnings release land completely differently. The author noted that "the growing discussion around NPO is the first sign that optics are starting to penetrate in-rack connectivity, which significantly expands the optical TAM." NPO (Near-Packaged Optics) is a cousin of CPO that places optics close to the chip, and TAM is the total market a company can reach. Until now optics were used mainly in the long scale-out runs outside the rack, but now, driven by the laws of physics, they've begun pushing into the scale-up domain inside the rack, a declaration that the market Lumentum can sell into has widened wholesale. What's more, this scale-up market has far more connection links by nature, so its potential dwarfs that of scale-out.

In the same vein, it becomes crystal clear why Applied Optoelectronics ($AAOI) was singled out as an exception. AAOI's management stated, "We just can't make enough CPO lasers to be involved in their current first-generation deployments, because we have to prioritize making lasers for our own transceivers first." As we saw, ultra-high-power laser manufacturing is so brutally difficult that only the handful who have secured capacity can ride this wave, and AAOI, lacking that capacity, stayed in the spectator's seat. Even within the same industry, in the end it's "who can actually make it" that separates winners from losers.

🌐 CoreWeave and Supermicro: Why They Erupted on the Same Day

On that same day, CoreWeave (CRWV) announced a revenue backlog of roughly $104 billion, and that figure didn't even include the $25 billion of new customer commitments added early in the third quarter. Supermicro (SMCI) unveiled a striking revenue guidance of $65 to $72 billion along with more than $60 billion in new orders. A backlog refers to volume that has already been contracted and will show up as revenue down the road.

Their erupting on the same day isn't coincidence but a single chain of cause and effect. The hardware that physically builds the data center (Supermicro), the cloud that sells compute on top of it (CoreWeave), and the optics that link the chips with light (Lumentum) are simply different layers of the same AI infrastructure investment cycle. One connection is especially worth watching: the author's model assumes CoreWeave will take roughly half of the 2026 CPO switch volume, thanks to its tendency to adopt the latest NVIDIA networking ahead of rivals. Ultimately CoreWeave's order surge flows straight through to demand for Lumentum's ELS modules. It's a set of interlocking gears, where growth in one layer pulls growth in another.

⚡ NuScale Power's Equity Raise: Why Now, Why This Way

Meanwhile, the small modular reactor (SMR) company NuScale Power ($SMR) announced a $750 million ATM equity offering.

An ATM (At-The-Market) offering is a funding method in which the company sells shares little by little at market price whenever it needs to. Unlike a traditional offering that prints a large block all at once, it spreads out the shock to the share price while raising money flexibly. NuScale explained it will use the proceeds for capital expenditure and research and development, and that this is an additional plan following the full depletion of the funds it raised back in February of this year.

Let me draw out a chain of cause and effect that newcomers easily miss. An equity offering dilutes existing shareholders' stakes, so in the short term it tends to read as bad news. But once you understand the essence of the SMR business, the nuance shifts. An SMR is a next-generation reactor that modularizes a large plant into small units you can add as needed to scale up; it requires large upfront investment over a long period before commercialization and is still in the stage before real revenue arrives. In fact, both revenue and costs shrank this quarter as well. In short, NuScale is still in the "burning before earning" phase, so fundraising is a condition of survival.

And this is where it connects back to the earlier AI story. The demand explosion at CoreWeave and Supermicro translates into enormous power consumption. Although CPO cuts power by 73% as noted, coldly speaking the network consumes only 9 to 15% of an AI cluster's total power, with the rest swallowed by thirsty GPUs. In other words, what AI truly thirsts for is a large-scale power source that is both stable and carbon-free, and SMRs are cited as a leading candidate. NuScale's continued fundraising can be read as positioning for this coming surge in power demand. That said, the long road to commercial operation and the repeated dilution are risks that must be weighed coldly.

🧭 Questions We Rarely Ask: Looking from the Other Side

First, "isn't the real bottleneck for this growth supply rather than demand?" A precise point. Lumentum's InP (indium phosphide) laser fabs are currently running at 100% full utilization, to the point of turning customers away because it can't keep up with flooding demand. So over the coming quarters it's carrying out a 40% capacity expansion, including a switch from 3-inch to 4-inch wafers. Here, InP is a specialized semiconductor substrate for lasers; unlike silicon wafers that have grown to 300 mm, InP remains around 100 mm, making it fundamentally lower in productivity. How many of these wafers can be produced is, in effect, the ceiling on supply. That's why in an industry like this you should watch "how far capacity expansion has progressed" more closely than "the size of the orders." NVIDIA's multi-billion-dollar investments in both Lumentum and Coherent to support the construction of a fifth wafer fab are aimed precisely at breaking this bottleneck.

Second, "aren't these results just counting the same money several times over?" CoreWeave's backlog, Supermicro's orders, and Lumentum's revenue are, in large part, the same AI investment capital flowing through the industry's various layers. The moment the will to invest of the hyperscalers, the large cloud operators at the source of that capital, falters, every layer can shake at once. A mutually supporting structure is powerful on the way up, but when the direction reverses, the shock rolls in together.

Third, "is the CPO transition really arriving right now?" In truth there are clear obstacles. With CPO, if a single component dies it can knock out 64 or more ports at once and render an entire switch unusable, hurting serviceability, and integrating the optical engine into the switch ASIC brings concerns about dependence, as pricing power concentrates in one company such as Broadcom or NVIDIA. That's why even NVIDIA is straining to stretch copper's lifespan as long as possible. Still, these obstacles are temporary factors that time will resolve, whereas copper's physical ceiling is a structural problem, so the transition is ultimately a question of "when."

🎯 In Closing

Compressed into a single sentence, this quarter comes down to this: as copper hit the wall of physics, AI data centers were forced into a transition to light, and Lumentum, which holds a virtual monopoly on the toughest gateway of that transition, the ultra-high-power laser and the ELS module, has begun harvesting the reward as earnings. The physical inevitability of a 73% power cut, the new market born from the constraint of having to move the laser away because of heat, and the 50% margin defended by a manufacturing difficulty no one can easily imitate: these three interlocked to produce staircase-like growth.

That said, the whole picture only comes into focus when you read together the true nature of the supply bottleneck, the circulation structure of capital within the industry, and the details of capacity expansion and orders that management will lay out at the upcoming earnings call.

This article is merely a summary of information and is not a recommendation to buy or sell any particular security. The final judgment and its responsibility rest with the investor.