The Photonics Fault Line: Sivers Photonics and the Hidden Supply Chain of the AI-Crypto Stack

The bubble isn't the story; the story is the story selling it. While the crypto market obsesses over ETF flows, validator counts, and the latest AI-agent token launch, a different kind of bottleneck is forming in the physical layer of the AI-crypto stack — and it's not in GPUs, not in memory, not in power. It's in photonics. The optical interconnect layer that shuttles data between the compute clusters running everything from decentralized training networks to validator nodes is hitting a supply wall, and the companies sitting on that wall are the ones nobody in crypto is watching. Sivers Photonics (SIVE), a London AIM-listed III-V compound semiconductor foundry with operations in Sweden and the UK, has become a case study in how the market misprices critical infrastructure. The company doesn't mine Bitcoin, doesn't run validators, doesn't issue tokens. It fabricates indium phosphide (InP) laser components and silicon photonics chips — the core building blocks for co-packaged optics (CPO), the packaging technology that AI data centers will need to scale bandwidth without exploding power consumption. And right now, it's sitting in the middle of a supply bottleneck that the market hasn't fully priced. Friction reveals the fault lines no one else sees. The friction here is visible in the numbers: supply constraints, rising average selling prices, and a customer list that reads like a who's who of the CPO revolution — Ayar Labs, O-Net, and six new pluggable optical module customers that have come on board in recent quarters. The market doesn't reward patience; it rewards positioning. And Sivers' positioning is better than its stock price suggests. To understand why a small photonics foundry matters for the crypto ecosystem, you have to understand the physical architecture of the AI-crypto convergence. The narrative of 2025-2026 is that AI agents will transact on-chain, that decentralized compute networks will challenge centralized cloud providers, that zero-knowledge proofs will verify AI outputs. All of that requires data centers. Lots of them. And data centers require optical interconnects — the fiber optic links and photonic engines that move data between servers, between racks, between clusters. The current generation of data center interconnects relies on pluggable optical modules — the transceivers that plug into switch ports. These work, but they're power-hungry and bandwidth-limited. A typical 800G pluggable module consumes 12-15 watts, and when you scale to 1.6T and beyond, the power envelope becomes prohibitive. The next generation is CPO: co-packaged optics, where the optical engine is packaged directly on the same substrate as the switch chip. CPO eliminates the pluggable interface, reduces power consumption by 30-50%, and increases bandwidth density by an order of magnitude. It's the technology that will make 1.6T and 3.2T interconnects feasible. The market trajectory is well-documented. LightCounting projects the CPO market to reach tens of billions of dollars by 2028, with penetration growing from under 5% in 2024 to 20-30% by 2028. The drivers are the AI cluster buildouts — NVIDIA's GB200 NVL72 racks, which require thousands of dollars of optical interconnect per cabinet, and the next-generation switch ASICs from Broadcom and Cisco that are being designed with CPO in mind. Sivers Photonics sits at the top of this supply chain. As a specialty foundry focused on III-V compound semiconductors and silicon photonics, it provides the InP gain media — the lasers and optical amplifiers — that make CPO light engines work. This is not a commodity business. InP epitaxial growth is a black art that requires years of process development. The yield curves are brutal — InP active integration yields typically run 70-85%, compared to 85-95% for mature silicon photonics. The customer qualification cycles are measured in quarters, not weeks. The company's strategic position is defined by two assets: its InP active integration technology, which puts it in the first tier of photonics foundries globally, and its dual-fab footprint, which gives it capacity optionality that most competitors at its scale don't have. The market, however, has been treating Sivers like a small-cap European semiconductor stock with limited upside — a perception that a recent shareholder letter from Serenity Capital has challenged head-on. Let me break down what's actually happening at Sivers, based on the technical and financial signals that are visible in the public record. The first thing to understand is that Sivers is not competing with TSMC on silicon photonics at scale. TSMC's COUPE (Compact Universal Photonic Engine) platform, expected to enter production in 2025, represents the state of the art in CMOS-compatible silicon photonics. GlobalFoundries' 45nm silicon photonics platform is similarly formidable. Sivers doesn't have the scale or the process maturity to compete head-to-head on silicon photonics integration density. But that's not where the value is. The value in CPO light engines is in the InP active components — the lasers, the semiconductor optical amplifiers, the modulators. And this is where Sivers has genuine first-tier capability. InP epitaxial growth, the process of depositing crystalline layers of indium phosphide on a substrate to create laser structures, is one of the most difficult manufacturing processes in the semiconductor industry. The yield rates for InP active integration typically run 70-85%, compared to 85-95% for mature silicon photonics platforms. The know-how required to achieve acceptable yields is accumulated over years of process iteration. Based on my experience auditing semiconductor supply chains, I can tell you that the InP integration capability is the moat here. TSMC can throw billions at silicon photonics, but InP epitaxy is a different game — it requires specialized MOCVD equipment, proprietary growth recipes, and a tolerance for low yields that a high-volume foundry like TSMC isn't structured to absorb. This is why TSMC's COUPE platform relies on external InP partners for the active components. Sivers' InP capability is the reason Ayar Labs — the Intel-spun CPO pioneer — has Sivers in its supply chain. The gap in silicon photonics integration is real, but it's not the right comparison. Sivers is a specialty foundry, not a general-purpose one. The right comparison is against IQE in III-V epitaxy, and against the InP teams at Intel and Broadcom. In that comparison, Sivers holds its own — it's in the first tier for InP active integration, with a technology roadmap that aligns with the CPO industry's 2025-2028 expansion window. The technology roadmap is worth examining in detail. Sivers' core process involves hybrid integration — bonding InP gain media onto silicon photonic wafers to create active photonic integrated circuits. This is the technical path that CPO manufacturers are adopting, and it's fundamentally different from the monolithic integration approach that Intel has pursued. Hybrid integration offers better performance per dollar, but it requires the foundry to have deep expertise in both InP epitaxy and silicon photonics assembly. Sivers has both, which is rare at its scale. The most underappreciated data point in the Sivers story is the supply bottleneck. The company has been operating at high capacity utilization — the signals point to north of 90% — and average selling prices are rising. This is the classic setup for a pricing power inflection: demand exceeding supply, with no near-term capacity relief in sight. The bottleneck is structural. The photonics supply chain is not like the logic chip supply chain, where multiple foundries can add capacity in parallel. InP substrate supply is concentrated in a handful of Japanese suppliers — Sumitomo Electric, JX Nippon Mining — and the epitaxial growth capacity is even more concentrated. When the CPO market inflects, as it's projected to do from less than 5% penetration in 2024 to 20-30% by 2028, the InP supply chain will be the binding constraint. This is where Sivers' dual-fab footprint becomes strategically significant. The company operates fabs in Sweden and has been signaling a potential US expansion. The US fab, if it materializes, would serve two purposes: it would bring capacity closer to the American CPO customers (Ayar Labs, and the broader US data center ecosystem), and it would position Sivers to capture CHIPS Act funding. The strategic logic is sound — the US is where the demand is, and the US is where the policy support is. But there's a tension here. The capacity allocation between the two fabs is a strategic choice that will define the company's near-term revenue mix. Allocate capacity to the pluggable optical module customers (the current revenue base, growing at 15-20%) or to the CPO customers (the future revenue base, growing at 50%+)? This is the kind of decision that separates companies that capture a market inflection from those that miss it. The equipment picture adds another layer of complexity. Photonics manufacturing requires DUV lithography (248nm/193nm) rather than EUV, which means the equipment isn't subject to the same export controls as advanced logic manufacturing. But the specialized equipment — MOCVD systems for epitaxial growth, high-precision e-beam lithography for sub-micron waveguide patterning — has delivery lead times of 6-12 months. Any capacity expansion decision made today won't contribute to revenue until 2025-2026 at the earliest. The customer concentration risk is real. The top five customers likely account for 70-80% of revenue, with Ayar Labs and O-Net as the anchor customers. This is a double-edged sword. On one hand, deep customer relationships in CPO are a strategic asset — they represent qualification cycles that competitors can't easily replicate. On the other hand, the loss of a single anchor customer would be a material revenue event. The six new pluggable customers that have come on board are a positive signal — they suggest the company is diversifying its customer base beyond the CPO pioneers. But the diversification is happening at the pluggable end of the market, not the CPO end. The CPO customer concentration remains high, and it will remain high until the CPO market reaches true scale. The O-Net relationship is particularly interesting from a geopolitical perspective. O-Net is a Chinese optical module manufacturer, and its partnership with Sivers on external laser source (ELS) products creates a China exposure that some US investors might view as a risk. But it also creates optionality — if the Chinese data center market continues to grow, Sivers has a channel into it. The company's position as a UK/Sweden entity gives it a degree of neutrality that pure US or pure Chinese suppliers don't have. The financial picture is the weakest part of the Sivers story. The company is marginally profitable at best, with estimated gross margins in the 30-40% range — respectable for a specialty foundry, but not transformative. Research and development expenses are likely fully expensed, which is conservative but suppresses reported earnings. Free cash flow is likely negative, reflecting the capital intensity of the business and the working capital drag from long customer payment cycles. The valuation is where the debate gets interesting. At an estimated price-to-sales multiple of 5-8x, Sivers is not cheap. The market is pricing in CPO growth expectations, but the question is whether those expectations are sufficient. Serenity Capital's argument is that the market is underpricing the economic value of the order book — that the combination of supply constraints, rising ASPs, and CPO customer commitments represents more revenue certainty than the current valuation reflects. The return on invested capital picture is also challenging. With estimated ROIC in the 3-5% range against a weighted average cost of capital of 10-12%, Sivers is currently destroying value. The bull case is that the CPO inflection will drive ROIC above WACC within 2-3 years, as capacity utilization stays high and ASPs continue to rise. The bear case is that the company will need to raise capital for the US fab expansion, diluting existing shareholders before the ROIC inflection materializes. Here's the angle that nobody's talking about: the "Swedish discount" is a real phenomenon, and it's distorting the valuation of a company that should be trading like a US-listed AI infrastructure play. Sivers is listed on London's AIM market, with a shareholder base that skews toward European retail investors and local institutions. The Swedish market, where the company has deep roots, has a particular investor psychology — one that values dividend yields, stable cash flows, and conservative balance sheets. That's not the profile of a company sitting on the edge of a CPO market inflection. The result is a structural valuation discount: Sivers trades at a multiple that reflects the Swedish market's investment logic, not the growth trajectory of the CPO market. This is the same dynamic that plays out across the crypto ecosystem — projects listed on the wrong exchange, with the wrong investor base, trade at a discount to their fundamental value. The fix is the same in both cases: move to where the growth investors are. Serenity's push for a US market focus isn't just about customer proximity — it's about investor proximity. A US listing, or even a US investor relations push, would expose Sivers to a shareholder base that understands CPO, understands AI infrastructure, and is willing to pay growth multiples for it. The second contrarian angle is the TSMC threat. The conventional wisdom is that TSMC's COUPE platform will crush specialty foundries like Sivers when it enters production in 2025. I think that's wrong, for a specific technical reason: TSMC's COUPE is a silicon photonics platform, and silicon photonics needs InP gain media. TSMC doesn't have first-tier InP epitaxial capability, and it's not clear that it wants to build it — the yield economics of InP are fundamentally different from the yield economics of CMOS logic. TSMC's likely path is to partner with InP specialists, which means companies like Sivers become suppliers to the TSMC ecosystem rather than competitors to it. The third angle is geopolitical. Sivers is a UK/Sweden company — a neutral player in the US-China technology competition. This neutrality has strategic value. As the US tightens export controls on AI-related technology, and as China responds with its own restrictions, the companies that can serve both ecosystems without triggering compliance issues will be at a premium. Sivers' position as a non-Chinese, non-US foundry with InP capability makes it a potential bridge between the two blocs — a role that has real option value. The fourth angle, and the one that connects most directly to the crypto ecosystem, is the AI-crypto convergence. The decentralized compute narrative — networks of distributed GPUs and specialized hardware that can be rented for AI training and inference — depends on the same optical interconnect infrastructure that centralized data centers use. If decentralized compute networks scale, they'll need photonics. Sivers is positioned to supply that infrastructure regardless of whether the compute is centralized or decentralized. The next 18 months will determine whether Sivers Photonics captures the CPO inflection or gets caught in the crossfire. The watch items are specific: TSMC's COUPE platform production timeline, Ayar Labs' 2028 expansion plans, the US fab decision, and the customer diversification progress. Each of these is a binary event that could move the stock 30-50% in either direction. The deeper lesson for the crypto ecosystem is about infrastructure blindness. We obsess over the application layer — the tokens, the protocols, the narratives — while ignoring the physical layer that makes everything work. The photonics supply chain is as critical to the AI-crypto convergence as GPUs are, and it's dramatically less understood. The companies that sit on this supply chain, like Sivers, are the picks-and-shovels plays of the next cycle. The question isn't whether they'll matter — it's whether the market will recognize it before the bottleneck becomes visible to everyone. The bubble isn't the story; the story is the story selling it. And right now, the story selling the AI-crypto convergence is missing its most important chapter: the photonics layer.