Demystifying the Semiconductor Value Chain: Interconnect

In April 2017, the Kansas City Chiefs traded up to draft Patrick Mahomes with the tenth overall pick. Everyone knew that Mahomes was meant to be the future of the franchise. The coaches knew, the fans knew, and so did the person he was drafted to replace.

And then he sat his first season. The coach chose Alex Smith as the starting quarterback. Around the league, Smith was viewed as the archetypal "game manager", a label scouts often used dismissively. He didn't have an overpowering arm, preferred the safe checkdown over the risky throw, and rarely forced the ball into tight windows. Smith made up for it by taking what the defense gave him and keeping mistakes off the scoreboard. His reliability earned him the trust to start every Sunday.

The AI build-out has an Alex Smith of its own. It is a piece of copper wire. Everyone knows optics will move data faster and farther than copper's physics will ever allow, and every industry roadmap has already penciled in the handoff. Yet when hyperscalers assemble the 100K+ GPU clusters driving today's AI boom, the highest-bandwidth links in the machine still run on copper. What it lacks in flash, it makes up for in reliability, the one thing a production deployment cares about most. However, the technology handoff and the stock-market handoff are not the same bet.

The Scouting Report on the Phenom

A 2017 draft profile described watching Patrick Mahomes play quarterback as "a bit like watching someone slice a banana with a playing card — you're not sure how it happened, but it looks cool as hell." In college, he threw for more than 5,000 yards while routinely attempting passes that most quarterbacks wouldn't even consider. The only knock on him was that his team didn't win. Mahomes once put up 819 yards of total offense in a single game, an FBS record that still stands, and lost 66–59. College football's most dazzling player couldn't consistently produce the one stat that mattered; saddled with a 13–16 record as a starter for Texas Tech, he was in the Heisman conversation in October and forgotten by December.

For over a decade, silicon photonics demos looked like Patrick Mahomes's college highlight reel. The demos showcased trillions of bits flowing through a chip no larger than a fingernail, but the wins that mattered never came. The prototypes skipped nearly every problem standing between a lab demo and a shipping product. Tiny ring-shaped modulators that encode data were so sensitive that just a few nanometers of manufacturing variation or a single degree of temperature drift could knock them off their wavelength. Lasers posed another problem. A failed laser in a pluggable module just meant swapping the module, but a laser sealed inside a chip that cost tens of thousands of dollars turned the same failure into an expensive loss. And even if reliability had been solved, mass production remained out of reach. Attaching fibers to wafers and testing optics at volume remained brutally difficult, and no one had a financial reason to get them right for servers running traditional create, read, update, and delete workloads.

The Game Manager's Secret

So why is copper reliable? The answer is simple. A passive copper link is just a piece of metal. There is no laser to degrade, no modulator to drift, nothing inside the cable to wear out or overheat. Its failure modes are predictable and easy for the engineer to catch while building the server rack. And like most game managers, copper comes without a premium. Moving a bit over copper costs a fraction of the power and money required to move it over optics, because the data never has to change form from electrons to photons and back. Copper gets no fanfare for making the short checkdowns that move the ball, but the data keeps moving and nothing gets dropped.

An optical link is complicated. At each end of every fiber-optic cable sits a small plug-in module that uses a laser to turn data into light. But first, the signal has to cross several inches of circuit board to reach the module. The trip leaves it smeared and faint, so a chip inside rebuilds every bit before handing it to the laser. That cleanup chip alone can burn nearly half the module's power. And the lasers themselves wear out, drawing more and more current to hold the same brightness, until one day they cannot.

Nvidia initially planned an optical 256-GPU scale-up back in 2022 and quietly shelved it due to the cost, power, and latency of running high-bandwidth scale-up over pluggable optics. Scale-up is the tight, ultra-fast mesh that binds a group of GPUs into what behaves like a single machine. Scale-out is the ordinary datacenter network that connects those machines by the thousands. A GPU's scale-up links carry roughly nine times the bandwidth of its scale-out network, and delivering that over transceivers would have been prohibitively expensive in both dollars and watts. That is before even asking how many plugs could fit on the front of the server.

When AI models outgrew the 8-GPU server, Nvidia followed the adage, use copper where you can, optics where you must. The GB200 packed all 72 GPUs and their switches into a single liquid-cooled rack, keeping every link inside copper's two-meter reach. Here are the tradeoffs.

Nvidia's decision to go with the game manager paid dividends not just for Nvidia, but for everyone holding the wire. The clearest winner was Amphenol, the connector maker that custom-engineered the NVLink spine cartridge for the GB200 NVL72. By 2025, Amphenol's IT Datacom business accounted for more than a third of company sales, with that segment's fourth-quarter sales growing 110% year over year on the back of AI infrastructure spending. Customers were ordering faster than Amphenol could ship, placing $1.31 of fourth-quarter orders for every dollar it sold. A connector company that spent nearly a century selling unglamorous metal had become one of the most direct ways to invest in the AI build-out.

The Succession Plan

Andy Reid built the Chiefs' offense around Alex Smith's strengths, emphasizing easy completions and yards after the catch. By scheming short, quick throws on first down, Reid shrank the field for the quarterback and let the Chiefs' playmakers generate extra yards to move the chains. Smith then became a mentor to Patrick Mahomes, who spent his rookie season learning Reid's system and absorbing Smith's methodical approach to reading defenses and managing the game. Kansas City also spent to build the roster around him, adding weapons on offense and fortifying its defense, giving Mahomes the supporting cast he did not have at Texas Tech. When Mahomes took over the following season, he immediately turned all of it into an MVP campaign.

Nvidia is now running a similar succession plan to transition from copper-based scale-up to co-packaged optics (CPO). Co-packaging moves the optics from the plug-in module at the front of the box to right beside the chip, and because the signal no longer has to cross the circuit board, the pluggable DSP is removed. The laser remains the exception, staying pluggable in Nvidia's first CPO switches because it is one of the most failure-prone parts of the link and needs to be replaceable outside the package. The savings are dramatic. Where a DSP-based pluggable module burns around 16 watts to deliver 800G of bandwidth, Nvidia's co-packaged version needs 4 to 5 watts, a roughly 73% cut.

At GTC 2025, Nvidia introduced its first CPO products, and tellingly they were scale-out switches, not scale-up systems. In scale-out, Nvidia is rehearsing its co-packaged optics debut the way Reid sat Mahomes. The switches give the supply chain time to ramp, build a track record on the field, and teach operators to live with optical engines that no longer unplug, all in the network that carries one-ninth the bandwidth of scale-up. The real payoff comes later in scale-up, where retiring copper's two-meter reach would finally let the machine grow beyond a single rack.

Along the way, Nvidia has been buying its way into the optics supply chain, and investors hunting for the Amphenol of optics have bet on every name it touches. In March 2026, it took sizeable stakes in Lumentum and Coherent, which are expected to supply the high-power lasers at the heart of co-packaged optics, with SemiAnalysis expecting Lumentum to supply Nvidia's initial CPO switch shipments and Coherent to enter behind it as a second supplier in late 2026. Nvidia also partnered with Corning, the company that draws the glass fiber itself, taking a warrant package worth up to $3.2 billion if fully exercised and backing three new plants in North Carolina and Texas that will grow Corning's US optical connectivity capacity tenfold. The money came with multibillion-dollar purchase commitments and reserved manufacturing capacity, aimed squarely at the flaws that kept silicon photonics stuck in the lab for a decade.

Since March, Nvidia has put around $6.5 billion into the optical supply chain, with warrants that could push the total past $9 billion.

The market is not waiting for the handoff. Experts describe CPO for scale-up as not a matter of if and why, but when and how, and the prices agree. Nvidia paid $695.31 per share for its Lumentum stake in March, valuing the company at roughly $67 billion fully diluted on about $3 billion in revenue, and the stock has traded well above that level since. Investors should notice that the most informed buyer in the room is not paying for a promising rookie. The sticker price far exceeds what the Chiefs paid for Mahomes. Nvidia bought in at 84 times this year's expected non-GAAP earnings, and as of August 7 the public pays over a hundred, prices that both warrant multiple Super Bowls. It is worth doing that arithmetic out loud ...

Paywalled Content

Nvidia's entry price marks Lumentum at $67 billion on a fully diluted count, well below where the stock trades today, so the arithmetic that follows is the conservative case. That price is a claim on profits the company has not earned yet, and the later they arrive, the less they count. Running it backward, at an ordinary cost of capital of about ten percent and with a generous 32% operating margin held the whole way, the price needs Lumentum at roughly $19.9 billion of revenue by FY2031 and $37.7 billion by FY2036. Lumentum booked $1.36 billion of revenue in FY2024 and lost half a billion dollars doing it, a GAAP figure swollen by acquisition charges, and will clear about $3 billion this fiscal year. So the price is not really asking whether CPO follows through. It is asking a $3 billion company to grow into $20 billion of revenue by FY2031 and $38 billion by FY2036.

That growth sounds impossible on its own, but the market underneath it is growing almost as fast. The fairer assessment is market share. LightCounting sizes the market for optical interconnects inside AI clusters at $26 billion in 2026, up 60% in a year, and thinks there is a reasonable chance it reaches $100 billion by 2030 if enough goes right. Lumentum's roughly $3 billion of revenue, most of it now going to cloud and AI customers, puts its share at 12% at most, and the true figure is lower because telecom and industrial are still in that total. The $67 billion mark needs $20 billion of revenue by FY2031, about 18% of where the market will have grown to by then. That is two bets, one on the market and one on Lumentum's share of it, and both have to land.

For the price to work, three things have to be true at once. The market has to get enormous, Lumentum has to keep a large share of it, and the margins have to hold the whole way. The first is the safest of the three, though a safe bet can still pay late. The other two are the same question, whether Lumentum will have pricing power over its one dominant customer, and Nvidia has already answered it. Two billion dollars went into Lumentum and two billion into Coherent on the same morning, both relationships explicitly non-exclusive, and Nvidia has named many other laser partners beyond the pair. The purchase commitments that came with the money point to real volume, and that visibility is worth something in the early years that count most. What is not guarateed are the margins. A buyer who arranges his supply that carefully does not intend to hand over the surplus. The Chiefs later paid Mahomes $450 million because he was irreplaceable. Nvidia has spent the past year making sure Lumentum is not.

The strongest counter is that capacity itself is the moat. Fabs take years to build, scarcity is what supports today's margins, and the hyperscalers designing their own silicon will need lasers from somebody too. The trouble is what the capacity costs. Lumentum is converting its profits into fabs rather than cash, which is the right decision and also the reason those earnings keep getting spent. Growth adds value only when the money sunk into it earns more than it costs, and the history of this industry is a long argument that it usually does not. The same fabs that produced a 32% operating margin last quarter produced a $546 million loss two years ago.

Then there is the calendar. Mahomes cost Kansas City a rookie contract, and the first ring arrived inside it. Optics is being paid at the top of the market and asked to deliver afterward. SemiAnalysis puts the real scale-up handoff past Rubin Ultra, out toward Feynman. A GPU generation is roughly two years, and at a ten percent discount rate a dollar that arrives two years late is worth 83 cents. Every dollar in the plan takes that haircut, and none of the physics has to be wrong for it to happen.

In 2017, nobody knew whether Mahomes's talent would translate to the NFL. The sticker price looked steep on draft night because the Chiefs were paying for uncertainty, and the bet worked out. Kansas City ended a fifty-year title drought and built a dynasty, winning three Super Bowls in five seasons. That kind of uncertainty is no longer for sale here. Everyone already knows optics takes over once copper runs out of physics, and that certainty is exactly what makes the stock expensive. The uncertainty that remains has shifted from whether optics matter to who captures the economics, when, and at what margin. Nvidia can win regardless, so long as the lasers ship on time. Its $2 billion bought supply security, reserved capacity, and a domestic laser ecosystem for its own roadmap, wrapped in a preferred security that stays invisible to anyone tracking 13F filings until it converts. Speculating as a retail investor comes with no such downside protection. The common shares carry only the equity math, and the equity math is the one part of the deal Nvidia does not need to work. Lumentum can be a very good company, the technology can arrive on schedule, and the shares can still be a poor investment. The dynasty is already in the price, and the only thing left to argue about is what year it starts.

Appendix A: Glossary

Silicon photonics is the underlying technology for building optical components, lasers aside, out of silicon on standard semiconductor manufacturing lines rather than assembling them from discrete parts.

Scale-up is the tight, high-bandwidth mesh that binds a group of GPUs into what behaves like a single machine, so that every GPU can talk to every other one at full speed. Nvidia's version is NVLink. Because the bandwidth is enormous and the latency budget is tiny, every rack-scale NVLink system Nvidia has shipped runs these links on copper, which is why their two-meter reach caps how large that single machine can get.

Scale-out is the ordinary datacenter network, InfiniBand or Ethernet, that connects those machines to each other to form a cluster of thousands. It carries roughly one-ninth of the per-GPU bandwidth of scale-up and has to span racks and rows, so it went optical years ago.

NVLink is Nvidia's proprietary scale-up interconnect. The fifth generation, used in Blackwell, moves 900 GB/s per GPU in one direction.

A pluggable transceiver is the small module that plugs into the front of a switch or server and converts electrical signals into light and back. It is the standard way optical links are built today.

A digital signal processor, or DSP, is the chip that rebuilds the electrical signal after it crosses the circuit board. The majority of pluggable optical transceivers shipping today carry one. It reshapes the waveform, recovers the timing, and decides again whether each bit was a zero or a one. This process consumes nearly half the transceiver's power.

Co-packaged optics, or CPO, moves the optical conversion off the faceplate and onto the same package as the GPU or switch chip. Because the signal no longer crosses the board, the DSP is no longer needed, which is where most of the power savings come from. The tradeoff is that the optics can no longer be unplugged.

A modulator is the component that imprints data onto a beam of light by switching it on and off, or shifting its phase, billions of times a second.

FIT, or failures in time, counts expected failures per billion device-hours. Lower is better.