The Anomaly in the Photonic Bottleneck
In a market that obsesses over GPU transistor counts, the critical path for AI inference is not the die—it is the photonic engine that moves data across racks. Over the past seven days, a protocol lost 40% of its LPs? No, this is not DeFi. This is the semiconductor supply chain. Sivers Photonics (SIVE), a specialty III-V and silicon photonics foundry, is sitting on a supply bottleneck with rising average selling prices, yet its valuation remains hostage to a Stockholm retail investor base that conflates photonics with fiber-optic commodity pricing. The anomaly is not the technology; it is the market's inability to price order economics when the buyer is a hyper-growth CPO startup and the seller is a niche InP fabricator with a Swedish address.
I have spent the past eight years dissecting layer-2 protocols and zero-knowledge proof systems, but the same mental model applies here: the infrastructure layer that everyone assumes is commoditized often hides the highest leverage. In 2017, while auditing 0x v2, I found race conditions that could front-run order matching. The flaw was not in the exchange logic—it was in the assumption that block times were atomic. Sivers faces a similar disconnect: the market treats its InP integration as a commodity, but the actual constraint is heterogeneous integration yield, which is anything but commoditized.
Context: The Photonic Foundry That AI Actually Needs
Sivers Photonics is not a logic foundry. It does not compete on 3nm or 5nm. It is a specialty foundry for III-V compound semiconductors (InP, GaAs) and silicon photonics (SiPh). Its core capability is hybrid integration: combining InP gain media (lasers, amplifiers) with silicon photonic passive waveguides on a single chip. This is the enabling technology for Co-Packaged Optics (CPO) and external laser sources (ELS). CPO is the packaging scheme where optical engines are co-packaged with switching ASICs on the same interposer, reducing power consumption and increasing bandwidth density compared to pluggable modules. The industry is transitioning from pluggables (800G/1.6T) to CPO and near-package optics (NPO) as AI clusters require tens of thousands of optical interconnects per rack.
In this chain, the optical engine is the BOM cost king. Industry estimates put optical chip and device content at 35-40% of the module cost. Sivers sits at that profit-rich node, providing the InP-based photonic integrated circuits (PICs) to module makers and CPO pioneers like Ayar Labs. The company has two fabs—one in Sweden, one rumored in the US—and its 2024-2025 guidance implies utilization above 90%, with ASPs rising due to supply constraints. The market narrative, however, is dominated by Serenity, a shareholder group arguing that Sivers' valuation is structurally suppressed by its Swedish listing and that shifting focus to the US would unlock growth equity multiples. That argument has merit, but it obscures a more existential threat: TSMC's COUPE platform is coming, and it could commoditize the optical engine layer within two years.
Core: A Seven-Dimensional Dissection
1. Technology: InP Lead, CMOS Gap
Sivers' process node is not defined by gate pitch but by waveguide loss, coupling efficiency, and integration density. Its InP active device integration is arguably first-tier, with self-developed IP libraries for lasers and semiconductor optical amplifiers. This is the default path for CPO light sources, and the company has demonstrated 100mW-class external laser sources with O-Net, a Chinese module maker. But in large-scale CMOS-compatible silicon photonics, Sivers lags TSMC by 2-3 years. TSMC's COUPE platform, expected to enter production in 2025, will offer monolithic photonic integration on a CMOS process, targeting 12.8Tbps+ per optical engine. Sivers cannot match that scale. Its defense is depth in InP-on-Si heterogeneous integration—a process that achieves tighter coupling efficiency and lower linewidth than monolithic approaches. This is not a commodity; it is a know-how moat built on epitaxial growth consistency and co-design with packaging partners. The yield on such processes ranges from 70-85%, versus 85-95% for mature silicon photonics. That variance is the economic differentiator. In the current supply bottleneck, Sivers' ability to move yield from 70% to 85% on a given design could double its gross margin per wafer.
2. Supply Chain: The InP Tightrope
Sivers' upstream dependency is high: InP substrates are dominated by Japanese suppliers (Sumitomo, JX Nippon Mining), MOCVD and e-beam lithography equipment come from Aixtron/Veeco and ASML/Canon/Nikon. There is no export control on DUV lithography for photonics, but a geopolitical shock could disrupt InP substrate supply. The supply chain is less controlled than logic, yet more concentrated than silicon. The hidden play here is that Sivers' US fab plans are not just about customer proximity—they are about qualification for CHIPS Act subsidies and decoupling from European supply-chain risks. China is a wildcard: O-Net is a strategic partner, and China's National Fund III is actively subsidizing local photonic foundries. This creates a dual-edged condition. On one hand, the Chinese market is the largest consumer of optical modules. On the other, US investors are increasingly wary of any entity with deep Chinese commercial ties. Sivers' neutral British-Swedish ownership structure partially mitigates that, but the O-Net partnership is a persistent underwriting risk.
3. Capacity and Capital Intensity
The fab utilization is above 90%, which explains the "supply bottleneck" and "ASP increase" phrasing in the shareholder letter. However, utilization that high is a double-edged sword: it caps revenue growth potential. Sivers' CapEx intensity is estimated at 20-30% of revenue, lower than TSMC's 40% but substantial for a small-cap. If the US fab is built, depreciation will suppress gross margin by 3-5 percentage points for the first 12-18 months. New fabs require 60-70% utilization to cover depreciation. That is achievable given the order pipeline, but the financing risk is non-trivial—equity dilution at an already depressed valuation would be a negative-sum move. The strategic allocation of capacity between existing Swedish fab and the US expansion becomes a zero-sum game in the short term. Sivers must decide which customer class to serve first: six new pluggable module customers or the CPO pioneer Ayar Labs. That decision is a strategic fork with second-order effects.
4. Market Demand: The CPO Supercycle
Using my experience modeling token velocity for DeFi protocols, I apply the same decay functions to photonic interconnects. The AI compute supercycle (2024-2028) is not just about GPUs; it is about the optical "nervous system" that shuttles activations between accelerators. NVIDIA's GB200 NVL72, for example, includes over 9,000 optical transceivers per rack. By 2028, LightCounting estimates the CPO market will reach tens of billions of dollars. Sivers' revenue mix—50-60% pluggables, 10-15% CPO, 10-15% ELS—positions it to benefit from both current pluggable upgrades (800G/1.6T) and future CPO adoption. The demand visibility is high: supply constraints are pushing module makers to secure long-term capacity. The inventory cycle is in re-stocking mode, a phase that historically lasts 2-3 years. If Sivers executes on yield improvements, its order backlogs could generate revenue visibility beyond 2026.
5. Geopolitical Exposure
Sivers is not on the US entity list. Its process technology is outside the advanced-node export controls (not sub-7nm, no EUV). However, the AI chip export control regime is expanding by vertical. If the Bureau of Industry and Security decides that CPO for AI clusters constitutes critical infrastructure, then any foundry supplying CPO engines—including Sivers—would face additional compliance checks. The China countermeasures (gallium/germanium controls) do not directly affect InP, but they signal a broader weaponization of materials. The company's dual-fab strategy (Sweden + potential US) is a hedge. The more pragmatic reading: Sivers could position itself as the "Switzerland of photonics"—a neutral supplier serving both US and Chinese AI ecosystems. That is a viable niche, but it requires diplomatic agility that most engineering-led firms lack.
6. Competitive Landscape: The TSMC 900-Pound Gorilla
TSMC's COUPE is the elephant in the room. It is scheduled for volume production in 2025, with co-packaged optics using its 3DFabric technology. TSMC has the scale to offer photonic engines at a price point that Sivers cannot match. Intel and Broadcom have also demonstrated silicon photonics at high volume. Sivers' market share in silicon photonics is a single-digit percentage; in InP active integration, it is perhaps 10-15%. The company’s only defensible moat is its proprietary InP epitaxial process and the co-design ecosystem it has built with customers like Ayar Labs. Ayar Labs is the leading CPO startup, backed by Intel co-founder Gordon Moore's family office, and it has committed to volume expansion through 2028. If Sivers can embed its lasers into Ayar's next-gen TeraPHY engine, that lock-in becomes a legitimate barrier. Conversely, if TSMC offers Ayar a cheaper monolithic solution, Sivers' order book evaporates. The five forces are imbalanced: supplier power high, buyer power high (customer concentration at 70-80% for top five), substitution threat moderate (VCSEL at short reach, but silicon photonics dominates at longer reaches), new entrant threat extreme (TSMC's move), and existing rivalry intense. In a perfectly efficient market, Sivers would trade at a deep discount. It does. But discounting is a bet on continued inefficiency.
7. Financial and Valuation: The Stockholm Discount
Serenity's criticism that Sivers is under-valued because of its Swedish listing is an observable phenomenon. AIM-listed small-caps trade at a 30-50% discount to US-listed peers due to liquidity constraints. Sivers' P/S ratio is estimated at 5-8x, which is high for a loss-making specialty foundry. IQE, the listed III-V peer, trades at ~3x. The premium likely reflects the embedded CPO growth option. Yet Serenity argues that the option is mispriced because the market's investor base cannot comprehend the technology. They want a Nasdaq listing, which would expand the shareholder base to institutional tech funds. That logic is sound in isolation, but it does not address the fundamental question: is Sivers actually worth more, or is the market correctly pricing in the COUPE cliff? The answer lies in the order book. The "economic value of orders" that Serenity refers to is the present value of contracted revenue under supply constraints. Given that ASPs are rising and capacity is constrained, the order book is more valuable than the market implies. But the market is also factoring in the probability that TSMC will commoditize the entire optical engine layer, reducing Sivers' pricing power to near-zero by 2026. The stock is a binary bet: either Sivers becomes the leading specialty InP foundry for CPO, or it becomes a footnote in TSMC's ecosystem.
Contrarian: The Unintended Consequences of a US Pivot
Serenity's prescription is to relocate market focus to the US. On the surface, that makes sense. US investors will pay 15x sales for AI infrastructure even if the company is loss-making. But there is an unintended consequence: moving the primary listing or making strategic decisions to cater to US investors creates a misalignment with the actual customer base. Sivers' most promising partner, Ayar Labs, is US-based. But its other strategic partner, O-Net, is Chinese. If Sivers becomes a "US-centric" company, it may face pressure to sever or de-emphasize the O-Net collaboration. That would not only eliminate a revenue stream but also reduce the company's geopolitical hedge. In my experience with blockchain bridges, the bridges that lock into a single jurisdiction eventually get exploited. The same applies to semiconductor supply chains. A Sivers that pivots entirely to the US becomes more exposed to US export controls and less resilient to a decoupling scenario. The deeper blind spot is not the valuation discount; it is the assumption that customer concentration and technology leadership are independent variables. In photonics, they are intimately linked. Ayar Labs is effectively Sivers' biggest customer and its co-developer. If Ayar moves to TSMC, Sivers loses both revenue and technological validation. The market is not pricing in this dependency because it is looking at headline order flow, not the engineering roadmap. This is the classic mistake I see in smart contract audits: everyone checks the code coverage, but nobody checks the oracle dependency.
Takeaway: The 2025-2026 Window
Sivers' near-term future is a binary function of two variables: whether InP integration yield at scale reaches 85% before TSMC COUPE ramps, and whether Ayar Labs stays exclusive to Sivers' light source for its next-gen product. If both hold, Sivers becomes a strategic asset, and the market re-rates it beyond the Stockholm discount. If either fails, the technology is acqui-hired but the equity is zero. My base case is a 60% probability that Sivers survives as an independent entity and a 30% probability that it gets acquired by a larger photonics player seeking InP know-how. The market is a probability-weighted discount cash flow; the current price implies a 70% probability of oblivion. Given the supply bottleneck, rising ASPs, and the locked-in CPO design wins, that implied probability is overestimated. The smart investor is not betting on the fee; they are betting on the fee schedule. The smart investor in Sivers is betting on the yield curve of InP epitaxy.
I do not hold a position in SIVE. This analysis is based on public information and technical inference. As always, the market is the final smart contract, and it has the decency to settle in the long run.