Supply bottlenecks whisper the truth before balance sheets ever do.
Over the past twelve months, something shifted in the optical chip market. Not with fanfare, but with the quiet pressure of extended lead times and climbing average selling prices. For Sivers Photonics (SIVE), a III-V compound semiconductor foundry operating out of Sweden with roots in the UK, this shift is existential. The market is finally waking up to what a handful of CPO builders have known for years: the real chokepoint in AI infrastructure isn't just the GPU. It's the light engine.
Code over hype. But in photonics, the code is written in indium phosphide.
I spent the early part of my career as an economic analyst in Shenzhen, watching supply chains bend under the weight of demand spikes. I learned that when a supplier tells you they're at capacity, they're actually telling you they've been at capacity for six months. Sivers has been saying this quietly for a year. The market just hasn't been listening.
The InP Advantage and Its Limits
The narrative that matters here is not about CMOS logic nodes. It's about something more arcane: the hybrid integration of indium phosphide gain media with silicon photonics passive structures. Sivers sits in a narrow but critical lane. They are one of the few specialty foundries globally that can credibly manufacture InP active devices—lasers and optical amplifiers—with enough consistency to feed the co-packaged optics (CPO) pipeline.
Here's what the mainstream coverage misses: the industry's flagship players are all racing toward the same destination, but they're taking different roads. TSMC is building its COUPE platform, a massive-scale silicon photonics engine slated for mass production in 2025. GlobalFoundries has its 45nm silicon photonics platform. Intel and Broadcom are vertically integrating. These giants bring capital, scale, and customer relationships that Sivers cannot match in absolute terms.
But the comparison is flawed if you treat it as a single-axis race. Sivers' differentiation is not in CMOS-compatible silicon photonics at scale—they trail TSMC there by perhaps two to three years. Their edge is in InP active integration, where they sit in the industry's first tier. This is the difference between owning the real estate and owning the mineral rights beneath it. The real estate is competitive. The mineral rights are not.
I've audited enough supply chains to know that when a foundry says its core IP is in epitaxial growth know-how, what they're really saying is that the process is fragile, the tolerances are unforgiving, and the learning curve is steep. That's a moat, but it's also a constraint. It means every single percentage point of yield improvement is worth more than a percentage point of market share. There are six new pluggable customers mentioned in the pipeline, and an external laser source (ELS) product developed with O-Net. The customer concentration is still high—probably 70-80% from the top five—but the diversification is real.
Truth decays slowly. So does yield. The question is which one you're willing to wait for.
The Dual-Fab Tension
The most underappreciated strategic signal here is the mention of capacity allocation across two wafer fabs. This is not a footnote; it's the entire story. Sivers appears to be operating a dual-site strategy—Sweden and the United States—and that split is both a hedge and a drag.
The US site, if it exists as suspected, is closer to the customer base that matters: Ayar Labs, the Intel-backed CPO pioneer, and the broader ecosystem of American AI infrastructure builders. Geographic proximity in photonics matters more than in digital logic because co-design with packaging partners is iterative. You can't debug a hybrid integration yield issue over a Zoom call. You need to be in the same cleanroom.
But the Sweden base carries the legacy investor structure. Serenity's critique—that the more time Sivers spends in the Swedish market, the more its valuation gets constrained by that market's investment logic—is sharper than it appears. This isn't just about listing venue. It's about the difference between retail investors who want quarterly narratives and institutional investors who can price a two-year CPO growth curve. The same company, with the same order book, is valued differently depending on which audience is doing the pricing.
The strategic tension is obvious: allocate capacity to the high-value CPO customers to win the future, or to the pluggable module customers who pay the bills today. In a supply-constrained environment, capacity allocation is strategy. Every wafer assigned is a bet on a thesis.
Market Structure and the AI Demand Curve
The demand picture is the cleanest part of this analysis, and it's genuinely strong. AI clusters are bandwidth-hungry in ways that traditional data center architectures never were. A single NVIDIA GB200 NVL72 rack can carry tens of thousands of dollars in optical interconnect value. The migration from 800G to 1.6T modules is happening faster than the supply chain can absorb. CPO penetration is expected to move from under 5% in 2024 to 20-30% by 2028, according to industry trackers like LightCounting.
This is not a cyclical story. This is a structural shift. The optical interconnect is the nervous system of AI. Every time someone talks about scaling inference or training larger models, they're implicitly talking about needing more optical bandwidth. The question is whether the supply side can keep up.
Here's where the hidden signal matters. The analysis mentions Ayar Labs' expansion plans through 2028. That's not just a customer update. That's an indicator that CPO is moving from pilot to production. And when a marquee player starts planning multi-year capacity expansion, the entire upstream becomes strategic. Sivers is positioned as a core supplier to that pipeline. The foundry is not a bet on a technology. It's a bet on a timeline.
The Contrarian Blind Spot
Now let me complicate the narrative, because if I don't, I'm doing my readers a disservice.
The bear case here is not about technology. Sivers has the tech. The bear case is about the economics of being a small foundry in a market where the biggest player is about to enter your lane. TSMC's COUPE platform, if it hits its 2025 timeline, has the potential to compress the entire CPO supply chain. When TSMC enters a market, they don't just compete. They redefine the cost curve. A specialty foundry with sub-scale production can survive that if it owns an irreplaceable capability. InP integration might be that capability. But it might also be a capability TSMC simply decides to partner for, rather than build.
The second blind spot is financial fragility. Small specialty foundries run tight. Capital expenditures for new fab capacity are not optional if you want to grow, but the cash flow to fund them is thin. If Sivers launches a US expansion, the depreciation hit could pressure gross margins by three to five percentage points for several quarters. In a bear market, where survival matters more than gains, investors don't always have the patience for a story that gets worse before it gets better.
And there's the geopolitical layer, which I keep coming back to because it's the part most mainstream coverage gets wrong. Sivers is a UK/Sweden company. It's not on any US entity list. It's not exposed to the advanced-node export controls that dominate the headlines. But the optics of doing business with a Chinese partner like O-Net, while courting American institutional investors, creates a narrative risk that needs to be managed. The neutrality that makes Sivers a bridge between East and West is also, ironically, the thing that makes it vulnerable to suspicion from both sides. In this environment, neutrality is not a position. It's a negotiation.
The Valuation Mismatch
The core of Serenity's argument is that the actual economic value of Sivers' order book is not reflected in its valuation. I think there's something to this. When a company operates in a supply-constrained market, the quality of orders matters more than the quantity. An order that comes with ASP expansion, multi-year visibility, and strategic customer partnerships is worth more than a marginal order booked at flat pricing. The market, constrained by liquidity and limited coverage, doesn't always discriminate between these.
My read on the valuation is that it's pricing in the CPO narrative but not the execution path. The PS multiple of 5-8x assumes growth, but it doesn't price the risk that growth gets delayed by yield challenges or competition. That's the gap—the market is either too optimistic about the speed of the transition, or too pessimistic about Sivers' positioning. The truth is probably somewhere in between, which means the stock is likely range-bound until the next major data point emerges.
The Signal in the Noise
So where does this leave us? Sivers Photonics is a company with real technical capability, real market positioning, and real structural tailwinds. It's also a company with concentrated customer risk, competitive threats from the largest semiconductor manufacturer in the world, and a balance sheet that will be tested if the expansion goes ahead.
This is not a story about whether the technology works. It's a story about whether the market can correctly price a niche player in a rapidly consolidating value chain. The Swedish market constraints are real, but they're not the only constraint. The real constraint is whether Sivers can convert its technical lead into production-scale economics before the giants arrive.
The next twelve months will tell us whether the supply bottleneck was a moment or a movement. If the bottleneck persists, Sivers has pricing power and the valuation re-rates. If it eases—if TSMC's COUPE arrives on schedule and compresses margins across the supply chain—Sivers' positioning gets tested in ways that no amount of narrative can protect.
Build anyway. That's the only option. The market might not understand the InP advantage today, but it will eventually. The question is whether Sivers can hold the line long enough to be the one telling the story.
Hold the line.