Here is the detail that keeps getting buried: Sivers Photonics is running hot. Supply is tight. ASPs are climbing. Six new pluggable customers walked in the door. And the stock still prices like a regional Swedish semiconductor curiosity.
From the noise of 2017 to the signal of today, I have watched markets misprice exactly this kind of bottleneck. When capacity is scarce and demand is structural, the order book is the signal. The ticker is just the lagging indicator.
Sivers is not a glamour name. It is an indium phosphide and silicon photonics specialty foundry. In the co-packaged optics era, that is the bottleneck at the center of every AI data center build-out. Serenity keeps hammering the point: the market is valuing the wrong metric. Not revenue. Not headline bookings. The economic value of the order book itself. Let me ground this in what I actually audit.
In 2026, I spent months dissecting decentralized AI compute markets. I analyzed Render Network's integration with large language models. The real constraint was never compute supply. It was the connective tissue: bandwidth between accelerators, power envelopes, and the optics carrying data at speed. In AI data centers, optical interconnect is the nervous system. A single GB200-class rack now carries tens of thousands of dollars in optical module value, up from thousands in the pre-AI era.
Sivers sits upstream of that value chain. It manufactures InP active components and silicon photonic wafers for the companies building the engines that make light carry data. Customers include CPO leaders like Ayar Labs, plus external laser source work with O-Net. When the market talks CPO adoption curves, Sivers is the quiet supplier at the top of the stack.
The company runs a dual fab structure: an existing Swedish base and a strategic orientation toward the United States. The supply bottleneck is real. The industry is restocking. Utilization is high. ASPs are rising. The market narrative still lags the physics. Photonics analysts project CPO market value reaching tens of billions of dollars by 2028. The foundry that owns the scarce InP capacity owns the pricing power. That is the trade.
This is the same infrastructure logic that will underpin decentralized AI networks, not just centralized hyperscalers. Whether the workload runs on OpenAI's clusters or on decentralized compute markets like Render and Akash, the light still has to move. Photonics is the shared railway. Sivers owns a scarce stretch of track.
Start where the money sits. In the optical communications value chain, the chip and component layer captures roughly 35-40 percent of the profit pool. Modules take 30-35 percent. System integration takes 20-25 percent. Materials capture the rest. Sivers lives in the richest segment of that stack. The chip layer is where pricing power lives, and pricing power is the scarcest asset in this cycle.
Technology requires calibration. Sivers is not competing with TSMC on dense CMOS-compatible silicon photonics. That fight is lost. TSMC's COUPE platform reaches volume production in 2025 with integration density Sivers cannot match today. GlobalFoundries' 45nm silicon photonics platform is equally ahead on CMOS-style scale. But the sell-side keeps flattening a distinction that matters: InP active integration is a first-tier capability, and it is precisely what CPO needs most. Hybrid integration - bonding InP gain media onto silicon photonic wafers - is the critical path for co-packaged optics. InP-on-silicon heterogenous integration is becoming the substrate strategy for the CPO era. Intel and Broadcom run their own InP teams. They also compete with their own customers. Sivers is a neutral foundry. In a market where every IDM is vertically integrated, neutrality is a feature. The market is still looking at wafer counts. The alpha is in the InP bottleneck.
Yield is the quiet variable. Silicon photonics platforms typically run 85-95 percent yield. InP active integration runs lower, roughly 70-85 percent. That gap decides unit economics. A fab that solves the yield problem compounds wafer value. Sivers' process know-how lives in precisely this hard middle: InP epitaxial growth and the consistency of lasers and optical amplifiers. Financial statements do not show it yet. The order book does. When supply bottlenecks meet rising ASPs, every percentage point of yield improvement is a direct margin transfer. Customer qualification cycles run twelve to eighteen months for CPO-grade components. Once qualified, switching costs are brutal. This is not a spot market.
Capacity math comes next. Utilization sits north of 90 percent by my read of the supply tightness. That is a seller's market. In a seller's market, the foundry chooses customers. That is the order book alpha nobody is graphing. The flip side: capacity limits revenue ceilings. Two fabs help. If the US capacity plan materializes, expect a 12-to-18-month equipment-to-volume timeline. Capital intensity at a specialty foundry runs 20-30 percent of revenue - lighter than TSMC's 35-45 percent, heavy for a small-cap. Depreciation will pressure gross margins by three to five points initially. At 60-70 percent utilization - which this demand cycle easily supports - the new fab covers its own cost. The strategic choice is allocation: which customers get the scarce wafers? High-value CPO clients or pluggable module buyers? Every wafer allocated to one segment is denied to another. In a seller's market, that decision is the CEO's most important job. The balance sheet will not show it for two more quarters.
Demand is not up for debate. CPO penetration moves from under 5 percent in 2024 to 20-30 percent by 2028. Ayar Labs' expansion roadmap to 2028 is the clearest signal that CPO is crossing from pilot to scale. Every one of those engines needs an external laser source. Sivers' ELS product with O-Net sits inside that attach. NVIDIA's GB200-class cluster architecture multiplies the optical module value per rack by an order of magnitude. The question is not demand. It is whether Sivers converts capacity into margin before the market re-rates it.
Geopolitics deserves a line. Sivers is UK-based with European production and a US market orientation. Not on any entity list. Its critical input - InP substrates - is dominated by Japanese suppliers. The relevant tooling is DUV, not EUV. That keeps Sivers outside the most aggressive export-control blast radius. China's gallium and germanium countermeasures do not touch its core materials. The company's neutrality has strategic value. It can serve US-facing CPO customers and maintain Chinese supply-chain ties through partners like O-Net without becoming a political target. US CHIPS Act incentives pull fab investment westward. That alignment is deliberate.
The competitive field: TSMC, Intel, Broadcom, IQE - plus a wave of Chinese challengers like Yuanjie Semiconductor and Accelink, backed by the state's third-phase semiconductor fund. Each giant has a different weakness. TSMC has scale but not InP leadership. Intel and Broadcom have technology but vertical-integration conflicts. IQE competes on volume, not on CPO-grade precision. The real risk is not one competitor. It is the market compressing photonics foundries into one undifferentiated bucket.
The consensus on the Serenity critique: Sivers suffers from a Swedish listing and a Swedish investor base. True, but that is the surface. The deeper problem is semantic. The market treats Sivers like a commodity foundry. It is not. A commodity foundry competes on price. Sivers competes on process know-how and captive customer relationships. That is a different species.
Speed runs require foresight, not just reaction. The obvious signals - revenue beats, margin prints - are backward-looking. The forward-looking signal is customer mix. Six new pluggable customers is not a headline. It is the beginning of a concentration-ratio decline. Watch how fast the top-five customer share drops from the 70-80 percent range.
Here is the angle nobody files: TSMC entering CPO is not a death sentence. It is market validation. The complexity of co-packaging - thermal management, packaging yield, testing - will scare off roughly 90 percent of would-be participants. That is not a bug. For the foundry that survives the gauntlet, it is the moat. I watched the same pattern in DeFi's composability rush. Programmable protocols expanded the design space. The complexity spike filtered out most developers. The survivors became infrastructure. Complexity is a filter. Sivers is on the right side of it.
Valuation sits at five to eight times sales. That is not cheap. But it is expensive for the wrong story. The market is pricing a cyclical foundry. The order book describes a bottleneck asset. The discount is not a discount on earnings. It is a mismatch: a UK operating story, a Swedish listing, and a US growth narrative. The same company priced through three different lenses. That is an arbitrage, not a thesis.
The ledger does not lie, but it rewards patience. The next watch item is the fab allocation statement. If Sivers signals a US capacity commitment and the customer list keeps diversifying, the re-rating follows the capital. If it keeps talking about Sweden, the discount persists. Markets price allocation, not intention. Speed runs require foresight, not just reaction. Watch the capital.